YOLO-LS: A Novel Deep Learning Framework for Brain Tumor Segmentation in Magnetic Resonance Imaging

YOLO-LS: A Novel Deep Learning Framework for Brain Tumor Segmentation in Magnetic Resonance Imaging | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article YOLO-LS: A Novel Deep Learning Framework for Brain Tumor Segmentation in Magnetic Resonance Imaging Jinghui Chen, Yan Hu, Tao Yang, Zhipeng Sun, Lianxin Xie, Hongjia Zhao This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8027109/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 12 You are reading this latest preprint version Abstract Brain tumors, as highly heterogeneous intracranial diseases, require accurate recognition and segmentation for clinical diagnosis, surgical planning, and prognosis evaluation. However, manual annotation of traditional MRI images suffers from subjective bias and low efficiency. This study proposes an improved algorithm, YOLO-LS, based on the YOLO11-Seg model to enhance real-time recognition, detection, and high-precision segmentation of brain tumor MRI images. Specific improvements include the adoption of ShuffleNet V1 as a lightweight backbone network to reduce parameters and computational complexity; the introduction of the DySample dynamic upsampling mechanism to enhance detail recovery; and the optimization of the C3k2 module into a C3k2-PoolingFormer block for efficient cross-scale feature fusion. Experiments were conducted using the Figshare dataset (3064 images) for training and internal testing, with the Kaggle dataset (300 images) serving as external validation. Results indicate that YOLO-LS achieved a bounding box mAP50 of 0.953 and a Dice coefficient of 0.910 on the internal test set, with GFLOPs reduced to 8.1, representing a 2.9% precision improvement and a 15.6% reduction in computational load compared to the baseline YOLO11. Ablation experiments and comparisons with models such as U-Net and SegNet confirmed the effectiveness of the improvements; heatmaps further validated the model's precise focus on tumor boundaries. On the external test set, the model demonstrated strong generalization, with an overall Dice coefficient of 0.895. This method achieves an excellent balance among precision, efficiency, and interpretability, showing significant potential for clinical applications and future extensions to multimodal fusion and federated learning. Biological sciences/Cancer Biological sciences/Computational biology and bioinformatics Physical sciences/Engineering Physical sciences/Mathematics and computing Health sciences/Medical research brain tumor artificial intelligence image segmentation YOLO11 magnetic resonance YOLO-LS Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 Figure 17 Figure 18 1 Introduction Brain tumors represent a class of heterogeneous diseases occurring within the cranial cavity, often accompanied by rapid progression, neurological impairment, and significant treatment challenges( 1 , 2 ). Brain tumors can be classified into primary and metastatic types, with primary tumors such as gliomas, meningiomas, and pituitary adenomas exhibiting high heterogeneity not only in biological behavior but also in growth location, invasiveness, and treatment response( 3 , 4 ). For instance, gliomas have a five-year survival rate of only 5–10%, while meningiomas generally have a better prognosis; however, their invasive and blurred boundaries often lead to incomplete surgical resection, increasing the risk of recurrence( 5 ). Early and accurate identification of tumor boundaries and infiltration extent is critical for developing surgical resection, radiotherapy, and chemotherapy plans, as well as for prognosis assessment. Nevertheless, traditional image-based manual annotation is time-consuming and subject to inter-observer variability, failing to meet the demands of large-scale, real-time clinical applications. Magnetic resonance imaging (MRI), with its multi-sequence, multi-parameter, and high soft-tissue contrast characteristics, has become a common imaging modality for brain tumor diagnosis, staging, and preoperative planning( 6 , 7 ). Therefore, high-precision automated brain tumor recognition and segmentation hold significant value in clinical applications. In recent years, methods based on deep learning networks have made substantial progress in MRI brain tumor segmentation tasks. Convolutional neural network (CNN) architectures, such as U-Net and its variants, have become benchmark models widely used in brain tumor challenges. For example, Pourmahboubi et al. proposed a U-Net architecture incorporating pretrained VGG19 weights, which improved segmentation precision of brain tumors in MRI images through transfer learning. This approach achieved superior performance on the TCGA low-grade glioma dataset, outperforming traditional U-Net variants, particularly in handling class imbalance( 8 ). Lu et al. developed the SF-SSM UNet model, which serialized MRI slices and utilized the Mamba state space model to capture spatiotemporal correlations, further optimizing feature learning via Fourier transform-based frequency domain analysis. This model demonstrated robustness in multimodal images on the BraTS-2019 dataset( 9 ). These CNN-based methods enhanced local feature extraction through attention mechanisms or residual connections but remain limited by fixed receptive fields, hindering effective capture of long-range dependencies. Subsequently, the introduction of Transformer architectures advanced the field of brain tumor segmentation, with their global attention mechanisms aiding in handling tumor heterogeneity and boundary ambiguity. Jia et al. proposed the BiTr-Unet model, combining CNN with Vision Transformer in the U-Net encoder via dual Transformer layers to capture local and long-range dependencies in 3D images, enhanced by CBAM modules for feature extraction( 10 ). Zhu et al. designed the multiPI-TransBTS framework, integrating multi-path multi-physics information fusion and adaptive feature modules, which improved Dice coefficients and sensitivity on the BraTS-2020 dataset, particularly in multimodal MRI sequence fusion( 11 ). Huang et al. introduced the TMA-TransBTS model with 3D multi-scale self-attention and cross-attention modules, achieving state-of-the-art performance on the BraTS2019 dataset and improving training stability via deep supervision strategies( 12 ). Huang et al. proposed a Transformer-based generative adversarial network (GAN), combining ResNet and Transformer blocks in the bottleneck layer with multi-scale L1 loss for discriminator optimization, surpassing traditional methods on the BraTS-2015 dataset and validating generalization on subsequent datasets( 13 ). These Transformer-based methods achieved breakthroughs in multimodal fusion and global context modeling but suffer from high computational complexity, limiting real-time applications. The YOLO (You Only Look Once) series, as a representative single-stage object detection framework( 14 ), has garnered widespread attention since its introduction in 2015 for its efficient real-time performance and end-to-end design. This series models object detection as a single regression problem, directly predicting bounding boxes and classes on image grids to balance speed and precision. Its evolution spans multiple generations: YOLOv1 established the foundational architecture( 15 ), YOLOv2 introduced anchor boxes and batch normalization( 16 ), and YOLOv3 enhanced multi-scale target detection via multi-scale predictions and the Darknet-53 backbone( 17 ). Subsequent versions like YOLOv4 and v5 integrated CSPNet, PANet, and Mosaic data augmentation into the backbone and neck networks( 18 ), continuously optimizing performance and training efficiency. Recent versions from YOLOv6 to 11 have explored reparameterization, anchor-free detection heads, and decoupling of classification and regression tasks, extending to instance segmentation tasks with models such as YOLOv8-Seg and YOLO11-Seg( 19 ). These models incorporate mask prediction branches within the detection framework, enabling pixel-level segmentation while maintaining real-time capabilities, providing efficient and robust solutions for fine contour segmentation tasks, including medical imaging. The YOLO series, as a real-time object detection framework, has been extended to brain tumor segmentation. Recent studies have explored YOLO variants in MRI brain tumor image processing to improve real-time diagnostic efficiency. Montalbo proposed a transfer learning model based on YOLOv4-Tiny for computer-aided diagnosis of brain tumors, fine-tuned on T1-weighted CE-MRI images to achieve accurate detection of gliomas, meningiomas, and pituitary adenomas( 20 ). Abdusalomov et al. integrated CBAM and feature pyramid networks into YOLOv7 for brain tumor detection, emphasizing real-time performance and generalization on multimodal MRI data for gliomas, meningiomas, and pituitary adenomas( 21 ). These early works laid the foundation but still face challenges in small tumor detection and boundary ambiguity. Yang et al. developed variants based on YOLOv5s, incorporating ASPP modules for multi-scale context capture and CBAM and CA for enhanced feature attention, achieving precise segmentation of tumor regions on the Kaggle brain tumor dataset while optimizing computational load( 22 ). Abraham et al. proposed a feature extraction network based on YOLOv8, improved with dilated convolutions and dual feature pyramid networks (DFPN), combined with dynamic non-monotonic focusing mechanisms to enhance detection of complex tumor morphologies, particularly for small tumors and unclear boundaries in MRI images( 23 ). Priyadharshini et al. compared YOLOv9, YOLOv10, and YOLO11 variants, evaluating their detection and segmentation performance on Figshare and BraTS-2020 datasets, highlighting the advantages of multi-scale attention mechanisms and Transformer-based backbones in handling tumor heterogeneity( 24 ). These YOLO variants balanced precision and efficiency in edge computing scenarios through attention fusion and lightweight improvements but require further optimization for small lesions and boundary ambiguity. Despite various improvement strategies, several issues remain inadequately addressed: conventional convolutional stacking dilutes fine-grained features of boundary-ambiguous, highly heterogeneous small lesions through repeated pooling; achieving true lightweight deployment while ensuring high precision remains challenging; and differences in imaging protocols and quality across data sources necessitate specific validation of model generalization on external data. Based on these practical needs, this study proposes targeted improvements on YOLO11, naming the improved model YOLO-LS. The strategies include replacing the backbone with ShuffleNetV1 for lightweighting, introducing content-based dynamic upsampling DySample to enhance detail recovery, and replacing the C3k2 module with C3k2-PoolingFormer to optimize cross-scale feature fusion. These improvements aim to maintain or enhance recognition and segmentation precision while significantly reducing computational load and improving identification of small-volume and boundary-ambiguous tumors. In summary, the YOLO-LS model proposed in this study, based on targeted improvements to the YOLO11-Seg model, is expected to achieve a better balance between real-time recognition, detection, and high-precision segmentation of brain tumor MRI images, thereby enhancing the model's usability and interpretability in clinical environments. The following sections detail the proposed model architecture, datasets, training details, quantitative and qualitative evaluation results, and comparisons with existing methods along with ablation analyses. 2 Baseline Model The YOLO series, as a classic algorithm in object detection, is renowned for its real-time capabilities, accuracy, and ease of use. It has been widely extended to instance segmentation tasks, particularly suitable for recognition and precise boundary segmentation of brain tumor MRI images. In clinical scenarios, the end-to-end design of this series enables efficient processing of various MRI modalities, achieving tumor localization, type classification, and pixel-level segmentation to assist in real-time diagnosis and preoperative planning. YOLO11, as a current mainstream algorithm for object recognition, detection, and segmentation, offers five scale variants: YOLO11n, YOLO11s, YOLO11m, YOLO11l, and YOLO11x, accommodating diverse deployment needs from lightweight edge devices to high-performance servers. Its core architecture consists of a backbone network (for multi-scale feature extraction), a neck network (for feature fusion and enhancement), and a head network (for bounding box regression and mask prediction), supporting the YOLO11-Seg variant for direct output of instance segmentation results for tumor regions. Compared to previous versions, YOLO11's key innovations include the introduction of the C3k2 module as an optimized CSP variant, which optimizes information flow through feature map splitting and small-kernel convolutions to improve computational efficiency; this module supports parameterized configuration for dual-module serialization to increase feature depth or degradation to YOLOv8's C2f module, indirectly enhancing capture precision of heterogeneous textures in brain tumors; the addition of the C2PSA module fuses CSP structures with partial self-attention (PSA) mechanisms, where features are split via 1×1 convolutions—one path passes directly, the other undergoes PSA multi-scale convolution extraction, SE module channel weighting, and Softmax for point-wise weighting—to strengthen focus on key tumor regions, significantly improving detection robustness for multi-scale tumors in complex MRI images; the detection head borrows from YOLOv10 by incorporating depthwise separable convolutions to reduce redundant computations and accelerate inference speed, suitable for resource-constrained clinical workstations. However, the baseline YOLO11 model faces challenges in brain tumor MRI segmentation tasks: high computational costs in the backbone network hinder deployment on edge devices; insufficient feature extraction capabilities lead to loss of tumor details; and low-resolution feature maps struggle to capture small-volume, boundary-ambiguous tumor regions, limiting detection accuracy and segmentation precision. In light of these issues, YOLO11-Seg is selected as the baseline model for improvements in this paper, with its network structure shown in Fig. 1 , laying the foundation for subsequent lightweighting and detail enhancement optimizations tailored to brain tumor characteristics. 3 Materials and Methods 3.1 Improved Model Although YOLO11 has achieved significant improvements in precision and lightweighting for brain tumor MRI image recognition and segmentation tasks, repeated convolution and pooling operations in its architecture may lead to dilution or loss of fine-grained features in small-volume tumors during feature map compression. Additionally, the high heterogeneity of brain tumors in structure, morphology, edges, and location can induce model bias toward negative predictions, increasing the risk of false negatives. To address these challenges, targeted improvement strategies are proposed in this study to further enhance recognition, detection accuracy, and segmentation precision while maintaining lightweight characteristics. However, directly selecting larger-scale variants would substantially increase parameters and computational complexity, constraining deployment on clinical edge devices. Therefore, optimizations are performed based on the smallest-scale variant "n" of the YOLO11-Seg model, with main improvements including: ( 1 ) replacing the backbone network with ShuffleNet V1 for lightweight feature extraction; ( 2 ) introducing the DySample dynamic upsampling mechanism to enhance detail recovery; ( 3 ) improving the C3k2 module to C3k2-PoolingFormer for optimized cross-scale feature fusion. The improved network structure is shown in Fig. 2 . 3.2 ShuffleNet V1 Lightweight Network Extending YOLO series models to instance segmentation has achieved excellent balance, but the default backbone still suffers from large parameter counts and high computational complexity, limiting deployment on embedded devices. To overcome this bottleneck, an efficient lightweight architecture is adopted: ShuffleNet v1 replaces the original YOLO11 backbone to construct an instance segmentation network that balances high speed and high precision. ShuffleNet v1, proposed by the Megvii Technology team, is an advanced CNN architecture with extremely high computational efficiency( 25 ). Its core concepts involve introducing two novel operations—pointwise grouped convolution and channel shuffle—to significantly reduce computational complexity while maintaining model precision. ( 1 ) Pointwise grouped convolution reduces the computational complexity of 1×1 convolutions by applying grouped convolutions at the 1×1 layer, ensuring each convolution operates only on corresponding input channel groups, thereby substantially lowering costs; ( 2 ) Channel shuffle addresses the lack of inter-group information interaction in group convolutions by uniformly shuffling channel information in input feature maps. Specifically, for a convolutional layer with g groups and g×n output channels, the output channel dimension is reshaped to (g, n), transposed, and then flattened as input for the next layer. This operation ensures input and output channels are fully correlated. Assume an input feature map with width Wi, height Hi, channel count Ci, output channels Co, and kernel size K×K. Parameter counts for each convolution type are calculated as follows: ( 1 ) standard convolution (sconv) parameters: P sc =C i ×C o ×K×K; ( 2 ) group convolution (gc) parameters: P gc =×C o ×K×K; ( 3 ) depthwise separable convolution (dw) parameters: P dw = C i ×K×K. Thus, standard convolution has the highest parameter count, group convolution is 1/g of that, and depthwise separable convolution has the lowest, at 1/C o of standard. Combining group and depthwise separable convolutions greatly reduces network parameters and computations. The ShuffleNet unit is designed based on bottleneck unit principles. In its residual branch, a computationally economical 3×3 depthwise separable convolution is applied to the 3×3 layer. The first 1×1 layer is replaced with pointwise grouped convolution followed by channel shuffle, and the second pointwise grouped convolution restores channel dimensions to match the shortcut path. When stride is applied, a 3×3 average pooling is added to the shortcut path, and element-wise addition is replaced with channel concatenation. ShuffleNet V1 is composed of stacked ShuffleNet units divided into three stages, with the first building block in each stage applying stride = 2, maintaining other hyperparameters within the stage, and doubling output channels for the next stage. Network connectivity sparsity and complexity are controlled by adjusting group number g and channel scaling factor s to achieve lightweight feature extraction. The basic module network structure of ShuffleNet V1 is shown in Fig. 3 . 3.3 Introduction of DySample Upsampling YOLO11 employs traditional nearest-neighbor interpolation for upsampling, which may fail to effectively capture feature details and semantic information, impacting model performance in prediction tasks. DySample is a lightweight and efficient upsampling algorithm that uses content-aware dynamic upsampling to adjust sampling positions based on local image features, more accurately restoring feature point positions and details( 26 ). The DySample upsampling process consists primarily of generating sampling point set S and dynamic feature resampling, as shown in Fig. 4 . Its point sampling concept efficiently conserves computational resources. The sampling point generation process is as follows: ( 1 ) Based on input feature map size and upsampling factor s, a regular grid coordinate G representing initial sampling positions is generated. ( 2 ) Dynamic offsets O are generated. Input feature map X first passes through a linear convolutional layer to produce initial offsets, constrained by a static range factor (0.25), then undergoes pixel rearrangement to redistribute channel information to high-resolution spatial dimensions, yielding offset set O. ( 3 ) Dynamic offsets O are added to regular grid sampling coordinates G to obtain the final dynamic sampling point set S. The dynamic feature resampling process uses PyTorch's built-in grid_sample function with bilinear interpolation on input feature map X based on dynamic sampling point set S to smoothly generate high-resolution feature values, resulting in high-resolution feature map X′. To increase offset flexibility, dynamic per-point range factors are generated via linear projection of input features, multiplied point-wise by a sigmoid function and 0.5 static range factor to ensure equivalence with the original static constraint, as shown in Fig. 5 . 3.4 Improved C3k2 Module In the YOLO11 segmentation framework adopted in this study, the C3k2 module serves as a core component in the neck network, providing multi-scale feature extraction and fusion. It is an enhanced version of the C3 module from YOLOv7 and YOLOv8, designed to optimize feature fusion and computational efficiency through Cross Stage Partial (CSP) structures. The C3k2 module typically includes a backbone path and multiple bottleneck branches, where the backbone path uses convolutional layers for downsampling and channel adjustment, branches extract multi-scale features via repeated bottleneck blocks, and features are fused via concatenation. This design replaces the previous C2f block in YOLO11, further improving speed and accuracy, particularly in real-time object detection tasks. However, in medical imaging tasks like brain tumor segmentation, the C3k2 module has potential limitations: reliance on traditional convolutions may lead to insufficient capture of complex tumor boundaries and fine-grained textures, while computational overhead may be high for high-resolution MRI images, posing challenges for deployment in resource-limited scenarios. Thus, efficient lightweight improvements to the C3k2 module in YOLO11 are urgently needed. PoolFormer is an efficient visual Transformer variant proposed by Sea AI Lab in 2022( 27 ), as shown in Fig. 6 . Based on the MetaFormer concept, it replaces the computationally intensive self-attention mechanism in Transformer architectures with simple spatial pooling as the Token Mixer module, significantly reducing computational complexity and parameters while maintaining competitive performance in tasks such as image classification, object detection, and semantic segmentation. The overall PoolFormer architecture uses a hierarchical design, typically divided into 4 stages, each containing multiple PoolFormer blocks, as shown in Fig. 7 . Each PoolFormer block's core structure includes: ( 1 ) Patch Embedding: In the first stage, input images are converted to patch tokens via convolution; ( 2 ) Token Mixer: Average pooling is used for local feature aggregation instead of global attention, focusing more on local spatial relationships and reducing FLOPs; ( 3 ) Channel Mixer: Multi-layer perceptron (MLP) layers perform channel-dimensional feature transformations, typically including Layer Normalization, GELU activation, and DropPath regularization; ( 4 ) Residual Connections: Each block uses residual connections for training stability. The main computational processes are as follows: Where L is input data; Pool denotes average pooling; Norm denotes group normalization; \(\:{\text{W}}_{1}\) and \(\:{\text{W}}_{2}\) are learnable parameters in MLP; σ is the GELU activation function. Based on this, the C3k2-PoolingFormer module is proposed to improve the original C3k2 module in YOLO11. Specifically, this module retains the CSP structural framework of C3k2 but replaces bottleneck blocks in branch paths with PoolFormer blocks. The C3k2-PoolingFormer module dynamically configures its processing paths via built-in parameters to address varying task demands. When the parameter is "True," the C3k-PoolingFormer path is activated, with its core advantage lying in utilizing PoolFormer for efficient token interaction, significantly enhancing local feature representation capabilities, excelling in handling brain tumor features in complex MRI backgrounds. When the parameter is "False," the module switches to the PoolFormer path, which enhances global context awareness on a lightweight basis, capturing long-range dependencies crucial for understanding image spatial relationships. The network architecture is shown in Fig. 8 . 3.5 Data Acquisition Medical imaging data used in this study were sourced from two distinct open public datasets. The first dataset is the Figshare brain tumor dataset provided by Guangzhou Campus of Southern Medical University( 28 , 29 ), containing 3064 MRI images from 233 patients, covering three main brain tumor types: meningioma (n = 708), glioma (n = 1426), and pituitary adenoma (n = 930). This dataset, named Dataset 1, was used for model training and internal testing to evaluate performance. The second dataset is from the Kaggle platform( 30 ), with 300 randomly selected MRI images (100 each of meningioma, glioma, and pituitary adenoma) serving as an external test set for model performance evaluation. All images were pre-annotated for segmentation targets using LabelImg by two radiologists with intermediate titles or above, with results reviewed on-site by two radiologists with associate senior titles or above to ensure experimental accuracy. Five-fold cross-validation was employed for model training in this study. From Dataset 1, 2800 images were randomly divided into 5 equal subsets of 560 images each. These were split in an 8:2 ratio into training and validation sets, with 4 subsets as training and the remaining as validation, cycled five times to create five different data distribution combinations. The remaining 264 images served as an internal independent test set to evaluate the optimal model's segmentation performance. The training set was used for parameter learning; the validation set independently monitored the training process and triggered early stopping; the test set was strictly isolated for final performance evaluation only. This strategy detects overfitting tendencies in real time during training, ensuring optimal model performance. Annotation distribution characteristics and inter-class correlations in the training dataset were visualized through distribution statistics in Figs. 9 and 10 . 3.6 Experimental Environment Configuration Experiments were conducted on Windows 11 operating system, with CPU model Intel(R) Core(TM) Ultra 9 275 HX, GPU model NVIDIA GeForce GTX 5070 Laptop, 16G RAM, deep learning framework PyTorch-2.8.0, CUDA-12.9, and Python-3.9.23. Model hyperparameters are shown in Table 1 . Table 1 Model hyperparameter settings Optimizer Initial Learning Rate Batch Epoch loU Momentum Weight Decay AdamW 0.01 16 200 0.7 0.937 0.0005 3.7 Experimental Evaluation Metrics Multiple evaluation metrics were adopted in this study to comprehensively assess model performance in brain tumor MRI image recognition and segmentation tasks. These metrics include precision (Precision), recall (Recall), mean average precision (mAP), intersection over union (IoU), and Dice coefficient (Dice Coefficient), along with model computational complexity indicators such as GFLOPs (Giga Floating Point Operations). Here, mAP50 denotes average precision at IoU = 0.5, while mAP50-95 represents average precision across IoU from 0.5 to 0.95, providing a stricter and more comprehensive evaluation. These metrics are particularly suitable for medical image processing, quantifying model accuracy and robustness in detecting brain tumor positions, classifying tumor types, and precisely segmenting tumor boundaries, while assessing computational efficiency and lightweighting for clinical deployment. Calculation formulas for evaluation metrics are shown in Table 2 . Table 2 Experimental evaluation metrics Indicator Calculation formula Precision \(\:\mathbf{P}=\frac{\mathbf{T}\mathbf{P}}{\mathbf{T}\mathbf{P}+\mathbf{F}\mathbf{P}}\) Recall \(\:\mathbf{R}=\frac{\mathbf{T}\mathbf{P}}{\mathbf{T}\mathbf{P}+\mathbf{F}\mathbf{N}}\) mean Average Precision(mAP) \(\:\mathbf{m}\mathbf{A}\mathbf{P}=\frac{1}{\mathbf{N}}\sum\:_{\mathbf{i}=1}^{\mathbf{N}}\mathbf{A}{\mathbf{P}}_{\mathbf{i}}\) Intersection over Union (IoU) \(\:\mathbf{I}\mathbf{o}\mathbf{U}=\frac{\mathbf{T}\mathbf{P}}{\mathbf{T}\mathbf{P}+\mathbf{F}\mathbf{P}+\mathbf{F}\mathbf{N}}\) Mean Intersection over Union (mIoU) \(\:\mathbf{m}\mathbf{I}\mathbf{o}\mathbf{U}=\frac{1}{\mathbf{N}}\sum\:_{\mathbf{i}=1}^{\mathbf{N}}\mathbf{I}\mathbf{o}{\mathbf{U}}_{\mathbf{i}}\) Dice Coefficient \(\:\mathbf{D}\mathbf{i}\mathbf{c}\mathbf{e}=\frac{2\times\:\mathbf{T}\mathbf{P}}{2\times\:\mathbf{T}\mathbf{P}+\mathbf{F}\mathbf{P}+\mathbf{F}\mathbf{N}}\) In the formulas, TP represents true positives (correctly detected or segmented brain tumor pixels/bounding boxes by the model), FP represents false positives (mis-predicted non-tumor regions), FN represents false negatives (missed brain tumor regions), and N is the number of classes. mAP50 was selected as the primary evaluation metric in this experiment to verify the proposed model's effectiveness in brain tumor MRI image recognition and segmentation tasks, thereby comprehensively assessing the method's practicality and computational efficiency. Additionally, to further analyze model reliability in clinical applications, mean IoU (mIoU) and Dice coefficient (Dice) were calculated as supplementary segmentation performance metrics, with GFLOPs compared to evaluate lightweighting efficiency improvements before and after enhancements, ensuring practicality in resource-limited environments. 4 Results and Analysis 4.1 Ablation Experiments To verify the contributions of the proposed improvement modules to model performance, ablation experiments were conducted based on the YOLO11n model. Experiments used the internal independent test set from Dataset 1 (264 images) for evaluation, with metrics including mAP50, precision (Precision), recall (Recall), mean IoU (mIoU), Dice coefficient (Dice), and GFLOPs. All models were trained under identical experimental environments, hyperparameter settings, and five-fold cross-validation to ensure result robustness. Ablation experiments progressively introduced three main improvements: ( 1 ) replacing the backbone with ShuffleNet V1; ( 2 ) introducing DySample upsampling; ( 3 ) improving the C3k2 module to C3k2-PoolingFormer. Training curves for the YOLO11n and YOLO-LS models are shown in Figs. 11 and 12 . Table 3 Ablation experiment results ShuffleNetV1 Dysample C3k2-PoolingFormer P(B) R(B) mAP50(B) P(M) R(M) mAP50(M) mIoU Dice GFLOPs 0.914 0.873 0.924 0.92 0.877 0.929 0.815 0.898 9.6 √ 0.906 0.893 0.934 0.914 0.892 0.937 0.823 0.903 8.4 √ 0.916 0.883 0.936 0.922 0.885 0.939 0.823 0.903 9.5 √ 0.913 0.88 0.931 0.92 0.881 0.933 0.818 0.9 9.2 √ √ 0.908 0.903 0.946 0.916 0.9 0.947 0.831 0.908 8.3 √ √ 0.905 0.9 0.941 0.914 0.896 0.941 0.826 0.905 8.2 √ √ 0.915 0.89 0.943 0.922 0.889 0.943 0.827 0.905 8.9 √ √ √ 0.907 0.91 0.953 0.916 0.904 0.951 0.835 0.91 8.1 Based on the ablation experiment results in Table 3 and Fig. 13 , the progressive introduction of the three improvement modules significantly enhanced YOLO11n model performance in brain tumor MRI image segmentation tasks. The baseline YOLO11n model performed well, with mAP50(B) of 0.924, precision(B) of 0.914, recall(B) of 0.873, mIoU of 0.815, Dice of 0.898, and GFLOPs of 9.6, but showed room for improvement in recall and segmentation precision for small or boundary-ambiguous tumors. After replacing the backbone with ShuffleNet V1, precision(B) slightly decreased to 0.906, but recall(B) increased to 0.893, mAP50(B) to 0.934, mIoU to 0.823, Dice to 0.903, and GFLOPs decreased by 12.5% to 8.4, demonstrating effective multi-scale feature extraction under lightweight design. Introducing DySample upsampling further improved performance, with mAP50(B) reaching 0.936, recall(B) 0.883, mIoU and Dice both 0.903, and GFLOPs 9.5, indicating that its dynamic sampling mechanism significantly enhanced detail retention for small-target tumors. When using the C3k2-PoolingFormer module alone, mAP50(B) increased to 0.931, mIoU to 0.818, Dice to 0.900, and GFLOPs to 9.2, optimizing feature fusion for complex textures via pooling operations. Integrating all improvements yielded the best performance: mAP50(B) of 0.953, recall(B) of 0.910, mIoU of 0.835, Dice of 0.910, and GFLOPs reduced to 8.1, a 15.6% decrease from baseline. Overall, improvements elevated mAP50(B), mIoU, and Dice by 2.9%, 2.0%, and 1.2%, respectively. Training curves before and after improvements showed faster convergence and higher robustness, indicating an excellent balance between precision and computational efficiency, suitable for lightweight deployment in clinical environments. 4.2 YOLO Series Comparison Experiments To further validate the superiority of the YOLO-LS model in brain tumor MRI image recognition and segmentation tasks, comparison experiments were conducted with other mainstream YOLO series versions, including YOLOv5, YOLOv8, YOLO12, and the baseline YOLO11 model. All comparison models were evaluated on the same internal independent test set from Dataset 1 (264 images), under identical experimental environments, hyperparameter settings, and five-fold cross-validation. Evaluation metrics aligned with those in ablation experiments. The experiments aimed to examine performance across different YOLO versions in precision, segmentation accuracy, and computational efficiency, particularly for complex tumor morphologies and boundary ambiguity. Comparison results show that the YOLO-LS model outperformed baseline models across multiple metrics. Specifically, the improved model significantly reduced computational complexity while enhancing detection and segmentation precision. This indicates that the improvement strategies effectively addressed feature extraction deficiencies and small-target detection challenges in original YOLO series for medical imaging tasks. Detailed comparison results are shown in Table 4 and Fig. 14 . Table 4 YOLO series comparison experiment results Model P(B) R(B) mAP50(B) P(M) R(M) mAP50(M) mIoU Dice GFLOPs YOLOv5 0.892 0.851 0.908 0.898 0.855 0.913 0.795 0.876 7.3 YOLOv8 0.905 0.885 0.931 0.911 0.889 0.935 0.822 0.9 11.3 YOLO11 0.914 0.873 0.924 0.92 0.877 0.929 0.815 0.898 9.6 YOLO12 0.898 0.868 0.919 0.904 0.872 0.924 0.809 0.888 9.8 YOLO-LS 0.907 0.91 0.953 0.916 0.904 0.951 0.835 0.91 8.1 4.3 Comparison Experiments with Different Networks To comprehensively evaluate the overall performance of the YOLO-LS model in brain tumor MRI image recognition and segmentation tasks, comparisons were made with various non-YOLO series networks widely used in medical image segmentation, including U-Net, HRNet, SegNet, and Deeplab V3+. All models were evaluated on the same internal independent test set from Dataset 1, with consistent experimental settings. Comparison results are shown in Table 5 and Fig. 15 . From the comparison results in Table 5 and Fig. 15 , YOLO-LS exhibited outstanding performance across multiple key metrics. Specifically, YOLO-LS surpassed U-Net, SegNet, and HRNet in precision and recall, second only to Deeplab V3+. In segmentation quality, YOLO-LS achieved mIoU of 0.835 and Dice of 0.91, comparable to Deeplab V3 + but significantly higher than U-Net and SegNet. Notably, YOLO-LS demonstrated substantial advantages in computational efficiency, with GFLOPs of only 8.1, far below U-Net (38.5), SegNet (25.3), HRNetV2 (49.7), and Deeplab V3+ (43.8). This indicates that YOLO-LS maintained high segmentation precision while significantly reducing computational resource demands, showing potential for deployment in resource-limited clinical environments. Overall, YOLO-LS achieved a good balance among precision, recall, segmentation quality, and computational efficiency. Although slightly inferior to Deeplab V3 + in some metrics, its extremely low computational complexity makes it more advantageous in practical applications, particularly in clinical scenarios requiring real-time or near-real-time processing. This highlights the practical potential of YOLO-LS's real-time advantages combined with lightweight improvements in medical imaging tasks. Table 5 Different network comparison experiment results Model Precision Recall mIoU Dice GFLOPs U-Net 0.895 0.865 0.818 0.888 38.5 HR 0.91 0.88 0.83 0.9 49.7 SegNet 0.875 0.842 0.785 0.865 25.3 Deeplab V3+ 0.922 0.892 0.842 0.915 43.8 YOLO-LS 0.916 0.904 0.835 0.91 8.1 To visually demonstrate the performance of different network models in brain tumor MRI image segmentation tasks, segmentation effect visualizations were used for qualitative analysis. As shown in Fig. 16 , these visualizations clearly present each model's capabilities in handling tumor boundaries, morphologies, and heterogeneous textures, providing intuitive evidence of the proposed YOLO-LS model's superiority in precise segmentation compared to U-Net, HRNet, SegNet, and Deeplab V3+. This visualization analysis complements the quantitative metrics in Table 5 , further emphasizing the improved model's potential value in clinical diagnosis and treatment planning for precise tumor segmentation. 4.4 Generalization Capability Evaluation To evaluate the generalization capability of the proposed YOLO-LS model in brain tumor MRI image segmentation tasks, Dataset 2 was used as an external independent test set. This dataset originates from different sources, differing from training Dataset 1 in imaging equipment, patient populations, and image quality, making it suitable for testing model robustness on unseen data. The YOLO-LS model, using the best weights trained on Dataset 1, was directly applied to Dataset 2 for performance evaluation without additional fine-tuning. Evaluation metrics aligned with previous experiments. Specific metrics for each tumor type and overall performance of the YOLO-LS model on the external independent test set are shown in Table 6 . Table 6 External test set generalization capability evaluation results Category P(B) R(B) mAP50(B) P(M) R(M) mAP50(M) mIoU Dice Meningioma 0.902 0.905 0.938 0.91 0.902 0.936 0.828 0.903 Glioma 0.885 0.888 0.928 0.892 0.885 0.925 0.812 0.888 Pituitary 0.908 0.912 0.942 0.915 0.908 0.94 0.83 0.905 All 0.898 0.902 0.936 0.906 0.898 0.934 0.82 0.895 Based on the evaluation results in Table 6 , the YOLO-LS model exhibited excellent overall performance on the external independent test set, with bounding box detection precision (P(B)) of 0.898, recall (R(B)) of 0.902, and mAP50(B) of 0.936; mask segmentation precision (P(M)) of 0.906, recall (R(M)) of 0.898, and mAP50(M) of 0.934. For segmentation quality, mean IoU (mIoU) was 0.82, and Dice coefficient was 0.895. These metrics indicate that the model maintained high levels of recognition, detection, and segmentation performance on unseen data, validating its strong generalization capability. By tumor type, pituitary adenoma showed the best recognition and segmentation, with mAP50(B) of 0.942 and Dice of 0.905; meningioma followed, with mAP50(B) of 0.938 and Dice of 0.903; glioma metrics were slightly lower but still reached mAP50(B) of 0.928 and Dice of 0.888. This demonstrates the YOLO-LS model's good adaptability to different pathological types of brain tumors, effectively handling image differences from various data sources. However, glioma showed relatively lower recall and Dice on the external test set, possibly due to its often blurred boundaries and strong heterogeneity in MRI, though overall performance remained acceptable. To more intuitively display the YOLO-LS model's recognition and segmentation effects on the external test set, visualization results are provided as shown in Fig. 17 . This figure compares segmentation outputs of models before and after improvements on typical samples, clearly showing YOLO-LS's advantages in tumor boundary localization and detail recovery, further validating its potential in practical clinical applications. 4.5 Heatmaps To deeply analyze the model's decision-making process and focus on key brain tumor regions, Grad-CAM (Gradient-weighted Class Activation Mapping) was employed to generate heatmaps. This method visualizes the model's emphasis on different input image regions during prediction by computing gradient weights, with red areas indicating high attention (typically corresponding to tumor boundaries and textures) and blue areas indicating low attention. Heatmaps were generated based on YOLO-LS model predictions on MRI images from Dataset 1's internal test set, aiming to verify correct focus on lesion areas rather than background noise or irrelevant structures. Heatmap displays for the improved model on MRI images of different brain tumor types are shown in Fig. 18 . Figure 18 heatmap examples (top: label visualization of original images; middle: YOLO11-Seg model heatmap results; bottom: YOLO-LS model heatmap results after improvements). Results show that YOLO-LS heatmaps more precisely highlight tumor core and edge regions, with more concentrated attention distribution, avoiding dispersion to non-tumor areas. This benefits from the C3k2-PoolingFormer module's pooling operations enhancing local feature aggregation and DySample upsampling's detail retention capabilities. For example, in glioma images, heatmaps clearly show model focus on heterogeneous textures and invasive boundaries; in pituitary adenoma images, attention centers on tumor core positions, reducing mis-focus on adjacent normal tissues. Notably, in small tumor cases, improved model activation coverage is higher, further explaining its recall advantages. These visualization results not only validate model interpretability but also provide references for clinicians to understand AI decision bases. 5 Discussion The YOLO-LS improvement framework based on YOLO11 is proposed in this study, significantly enhancing efficiency in brain tumor MRI image recognition and segmentation through integration of ShuffleNet V1 lightweight backbone, DySample dynamic upsampling mechanism, and C3k2-PoolingFormer module. Experimental results indicate that the model achieved mAP50 of 0.953 on the internal test set with a 15.6% reduction in GFLOPs, also demonstrating good generalization on the external test set. These improvements address clinical challenges such as tumor heterogeneity, boundary ambiguity, and small-volume lesions, achieving an effective balance between precision and lightweighting, particularly suitable for resource-limited clinical environments. The outstanding performance of the YOLO-LS model stems from the targeted improvement strategies proposed in this study. The introduction of the ShuffleNet V1 backbone, through pointwise grouped convolutions and channel shuffle operations, substantially reduced parameters and computations with minimal loss in feature representation. This aligns with the growing demand for lightweight models in medical image analysis( 31 ). The DySample dynamic upsampling mechanism replaces traditional interpolation methods, with its content-aware sampling point generation strategy effectively mitigating detail loss due to position insensitivity during upsampling, crucial for precisely delineating brain tumors, especially small lesions with blurred, infiltrative boundaries. The innovation of the C3k2-PoolingFormer module lies in introducing PoolFormer's simple and efficient pooling-based Token Mixer mechanism into YOLO's neck network, enhancing capture of long-distance contextual information without the high computational costs of Transformers, better addressing brain tumor heterogeneity. Ablation experiments clearly demonstrate the independent contributions of each module and their synergistic effects. Furthermore, a prominent contribution of this study is the model's exceptional computational efficiency. Compared to mature segmentation architectures like U-Net, HRNet, SegNet, and Deeplab V3+, YOLO-LS achieves comparable or higher precision with only a fraction of their computational overhead. For example, YOLO-LS reaches a Dice score of 0.91 with only 8.1 GFLOPs, while the widely used U-Net requires 38.5 GFLOPs for a slightly lower Dice of 0.888. This nearly fivefold reduction in computational demand highlights the model's suitability for real-time clinical applications and deployment on resource-constrained hardware, crucial for expanding access to sophisticated diagnostic technologies. Beyond performance on known datasets, model generalization is essential for clinical utility( 32 , 33 ). YOLO-LS was evaluated on an independent external dataset without retraining or fine-tuning, with results confirming its robustness. The model maintained high performance across different tumor types from new data sources, with an overall Dice score of 0.895. This indicates that YOLO-LS has learned fundamental, transferable imaging biomarkers for brain tumors rather than overfitting specific training data features, enhancing its potential for real-world deployment( 34 , 35 ). Additionally, to ensure the model is not a "black box," Grad-CAM was used for visualization explanations. Generated heatmaps provide strong evidence that architectural improvements directly translate to more clinically relevant decision processes. Compared to the baseline, YOLO-LS exhibits more precise and focused attention on tumor cores and boundaries while ignoring irrelevant background anatomy. This enhanced focus is attributed to C3k2-PoolingFormer's improved feature aggregation and DySample's detail retention, explaining observed segmentation precision improvements and offering clinicians a visualization tool to build trust in model predictions. Although the YOLO-LS model achieved significant progress, limitations persist. First, the training dataset primarily relies on T1-weighted CE-MRI sequences, which enhance tumor regions well, but complete brain tumor assessment often requires complementary information from multi-modal MRI sequences like T2 and FLAIR; future extensions to multimodal fusion are needed for enhanced robustness. Second, the external test set is relatively small; while validating generalization, further evaluation on larger-scale, multi-center clinical datasets is required. Generalization robustness of deep learning models across institutions, scanners, and imaging parameters remains a major challenge in the field( 36 ). To prove true clinical applicability, validation on larger-scale, multi-center, prospectively collected real-world datasets is essential. Looking ahead, the following directions warrant deeper exploration: ( 1 ) Multimodal and temporal information fusion: Explore effective fusion strategies integrating T1, T2, FLAIR, DWI, and other sequences, and further investigate utilization of dynamic contrast-enhanced MRI (DCE-MRI) or diffusion tensor imaging (DTI) temporal data to provide richer imaging biomarkers for treatment response assessment and prognosis prediction. ( 2 ) Personalized adaptive learning: Develop online or meta-learning strategies enabling rapid fine-tuning based on specific patient follow-up imaging data, adapting to tumor morphological evolution during treatment for personalized precise segmentation. ( 3 ) Deep integration with clinical workflows: Interface segmentation results with radiation therapy planning systems (RTPS) or neuronavigation systems for automatic tumor volume delineation and dose calculation, or assist in surgical path planning, truly leveraging AI in clinical decision support. ( 4 ) Prospective clinical validation: Design rigorous clinical trials to quantitatively assess impacts of introducing YOLO-LS on radiologist diagnostic efficiency, delineation consistency, and ultimate patient treatment outcomes. ( 5 ) Collaborative training under privacy protection: To address privacy and compliance barriers in multi-center data sharing, introduce federated learning and other privacy-preserving techniques for cross-institutional collaborative training without moving raw data, fully utilizing distributed medical data resources to further enhance model generalization and robustness. Overall, the YOLO-LS model proposed in this study provides an efficient, lightweight, and interpretable solution for AI-assisted brain tumor diagnosis. It not only demonstrates feasibility of high-precision segmentation under lightweight constraints but also lays a solid technical foundation for subsequent research and clinical deployment. Through continued multimodal expansions, large-scale clinical validations, and seamless integration with clinical workflows, this framework holds promise as a reliable tool for neuroradiologists, ultimately contributing to improved brain tumor patient care processes and outcomes. 6 Conclusion A YOLO-LS model based on improved YOLO11 is proposed in this study for brain tumor MRI image recognition and segmentation. Through innovative integration of ShuffleNet V1 lightweight backbone, DySample dynamic upsampling, and C3k2-PoolingFormer module, the model achieved excellent performance with mAP50 = 0.953 and Dice = 0.91 on Figshare and Kaggle datasets, with GFLOPs reduced to 8.1, representing a 2.9% precision improvement and 15.6% computational load reduction compared to the baseline. Ablation experiments and comparative analyses confirmed contributions of each improvement module, with heatmap visualizations further validating model interpretability and precise focus on tumor boundaries. These results indicate that YOLO-LS achieves breakthroughs in balancing real-time capabilities and high precision, showing significant clinical application potential, especially for auxiliary diagnosis on edge devices. Future work will focus on multimodal data fusion, larger-scale clinical validation, and integration with radiation therapy planning to further enhance model generalization and practicality. This framework provides a novel tool for precision medicine in brain tumors, poised to improve patient prognosis and reduce healthcare burdens. Declarations Funding No Funding Author Contribution J.C. performed the experiments, analyzed the data, prepared the figures, and wrote the original draft. Y.H., T.Y., Z.S., and L.X. contributed to conceptualization, methodology, data curation, software implementation, and validation. H.Z. supervised the project, provided resources, and reviewed and edited the manuscript. All authors reviewed the manuscript. Data Availability The datasets analyzed during the current study are available in the following public repositories: The Figshare brain tumor dataset is available at https://www.kaggle.com/datasets/ashkhagan/figshare-brain-tumor-dataset (as referenced in citations 28 and 29). The Kaggle brain tumor MRI dataset is available at https://www.kaggle.com/datasets/masoudnickparvar/brain-tumor-mri-dataset (as referenced in citation 30). No new datasets were generated in this study. References Vankdothu, R., Hameed, M. A. & Fatima, H. A Brain Tumor Identification and Classification Using Deep Learning Based on Cnn-Lstm Method. Comput. Electr. Eng. 101 , 107960. https://doi.org/10.1016/j.compeleceng.2022.107960 (2022). Nazir, M., Shakil, S. & Khurshid, K. Role of Deep Learning in Brain Tumor Detection and Classification (2015 to 2020): A Review. Comput. Med. Imaging Graph. 91 , 101940. https://doi.org/10.1016/j.compmedimag.2021.101940 (2021). Kumar, S. & Mankame, D. P. Optimization Driven Deep Convolution Neural Network for Brain Tumor Classification. Biocybernetics Biomedical Eng. 40 (3), 1190–1204. https://doi.org/10.1016/j.bbe.2020.05.009 (2020). Karayegen, G. & Aksahin, M. F. Brain Tumor Prediction on Mr Images with Semantic Segmentation by Using Deep Learning Network and 3d Imaging of Tumor Region. Biomed. Signal Process. Control . 66 , 102458. https://doi.org/10.1016/j.bspc.2021.102458 (2021). Ostrom, Q. T. et al. Cbtrus Statistical Report: Primary Brain and Other Central Nervous System Tumors Diagnosed in the United States in 2015–2019. Neuro-Oncology 24(Supplement_5):v1-v95. (2022). 10.1093/neuonc/noac202 Vankdothu, R. & Hameed, M. A. Brain Tumor Mri Images Identification and Classification Based on the Recurrent Convolutional Neural Network. Measurement: Sens. 24 , 100412. https://doi.org/10.1016/j.measen.2022.100412 (2022). Rajinikanth, V., Kadry, S. & Nam, Y. Convolutional-Neural-Network Assisted Segmentation and Svm Classification of Brain Tumor in Clinical Mri Slices. Inform. Technol. Control . 50 (2), 342–356. 10.5755/j01.itc.50.2.28087 (2021). Pourmahboubi, A., Arsalani Saeed, N. & Tabrizchi, H. A Brain Tumor Segmentation Enhancement in Mri Images Using U-Net and Transfer Learning. BMC Med. Imaging . 25 (1), 307. 10.1186/s12880-025-01837-4 (2025). Lu, J. et al. A Sequential Flow Unet for Mri Brain Tumor Segmentation Based on State-Space-Model. Appl. Soft Comput. 186 , 114069. 10.1016/j.asoc.2025.114069 (2025). Jia, Q. & Shu, H. (eds) Bitr-Unet: A Cnn-Transformer Combined Network For mri Brain Tumor Segmentation. Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries; (2022). 2022//; Cham: Springer International Publishing. Zhu, H. et al. A Multi-Path Learning Framework for Brain Tumor Image Segmentation Based on Multi-Physical Information. Comput. Biol. Med. 191 , 110148. https://doi.org/10.1016/j.compbiomed.2025.110148 (2025). Huang, Y., Chen, L. & Zhou, C. Multi-Modal Brain Tumor Segmentation Via 3d Multi-Scale Self-Attention and Cross-Attention. arXiv preprint arXiv:250409088 (2025). Huang, L. et al. A Transformer-Based Generative Adversarial Network for Brain Tumor Segmentation. Front. NeuroSci. 16. 10.3389/fnins.2022.1054948 (2022). Kotthapalli, M., Ravipati, D. & Bhatia, R. Yolov1 to Yolov11: A Comprehensive Survey of Real-Time Object Detection Innovations and Challenges. arXiv preprint arXiv:250802067 (2025). Zhao, Z. Q., Zheng, P., Xu, S. T. & Wu, X. Object Detection with Deep Learning: A Review. IEEE Trans. Neural Networks Learn. Syst. 30 (11), 3212–3232. 10.1109/TNNLS.2018.2876865 (2019). Luo, Y. The Evolution of Yolo: From Yolov1 to Yolov11 with a Focus on Yolov7's Innovations in Object Detection, Theoretical and Natural Science, 87, 1, 82–90 ,. (2025). (2025). (2025). Ali, M. L. & Zhang, Z. The Yolo Framework: A Comprehensive Review of Evolution, Applications, and Benchmarks in Object Detection (Preprints, 2024). 10.20944/preprints202410.1785.v1 Murat, A. A. & Kiran, M. S. A Comprehensive Review on Yolo Versions for Object Detection. Eng. Sci. Technol. Int. J. 70 , 102161 (2025). Ultralytics Yolov11: Ultralytics Yolo Documentation – Instance Segmentation: Ultralytics [2025-11-02]. (2024). Available from: https://docs.ultralytics.com/tasks/segment/ Montalbo, F. J. A Computer-Aided Diagnosis of Brain Tumors Using a Fine-Tuned Yolo-Based Model with Transfer Learning. KSII Trans. Internet Inf. Syst. 14 , 4816–4834. 10.3837/tiis.2020.12.011 (2021). Abdusalomov, A. B., Mukhiddinov, M. & Whangbo, T. K. Brain Tumor Detection Based on Deep Learning Approaches and Magnetic Resonance Imaging. Cancers (Basel) 15(16). Epub 20230818. (2023). 10.3390/cancers15164172 Yang, T. et al. Application of Mri Image Segmentation Algorithm for Brain Tumors Based on Improved Yolo. Front. Neurosci. 18 , 1510175. 10.3389/fnins.2024.1510175 (2024). Epub 20250107. Abraham, L. A., Palanisamy, G. & Goutham, V. Dilated Convolution and Yolov8 Feature Extraction Network: An Improved Method for Mri-Based Brain Tumor Detection. IEEE Access. 13 , 27238–27256. 10.1109/ACCESS.2025.3539924 (2025). Priyadharshini, S., Bhoopalan, R., Manikandan, D. & Ramaswamy, K. A Successive Framework for Brain Tumor Interpretation Using Yolo Variants. Sci. Rep. 15 (1), 27973. 10.1038/s41598-025-13155-4 (2025). Zhang, X., Zhou, X., Lin, M. & Sun, J. (eds) Shufflenet: An Extremely Efficient Convolutional Neural Network for Mobile Devices. Proceedings of the IEEE conference on computer vision and pattern recognition ; (2018). Liu, W., Lu, H., Fu, H. & Cao, Z. (eds) Learning to Upsample by Learning to Sample. Proceedings of the IEEE/CVF international conference on computer vision ; (2023). Yu, W. et al. (eds) Metaformer Is Actually What You Need for Vision. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition ; (2022). Ashkhagan Figshare Brain Tumor Dataset: Kaggle [2025-11-02]. (2025). Available from: https://www.kaggle.com/datasets/ashkhagan/figshare-brain-tumor-dataset Cheng, J. et al. Retrieval of Brain Tumors by Adaptive Spatial Pooling and Fisher Vector Representation. PLOS ONE . 11 (6), e0157112. 10.1371/journal.pone.0157112 (2016). Nickparvar, M. Brain Tumor Mri Dataset: Kaggle [2025-11-02]. (2021). Available from: https://www.kaggle.com/datasets/masoudnickparvar/brain-tumor-mri-dataset Al-Milaji, Z. & Yousif, H. Lightweight Deep Learning Model Optimization for Medical Image Analysis. Int. J. Imaging Syst. Technol. 34 (5), e23173 (2024). Tran, A. T., Zeevi, T. & Payabvash, S. Strategies to Improve the Robustness and Generalizability of Deep Learning Segmentation and Classification in Neuroimaging. BioMedInformatics 5 (2), 20 (2025). Dorfner, F. J., Patel, J. B., Kalpathy-Cramer, J., Gerstner, E. R. & Bridge, C. P. A Review of Deep Learning for Brain Tumor Analysis in Mri. NPJ Precision Oncol. 9 (1), 2 (2025). Aiya, A. J. et al. Optimized Deep Learning for Brain Tumor Detection: A Hybrid Approach with Attention Mechanisms and Clinical Explainability. Sci. Rep. 15 (1), 31386. 10.1038/s41598-025-04591-3 (2025). Aziz, N. et al. Precision Meets Generalization: Enhancing Brain Tumor Classification Via Pretrained Densenet with Global Average Pooling and Hyperparameter Tuning. PLOS ONE . 19 (9), e0307825. 10.1371/journal.pone.0307825 (2024). Matta, S. et al. A Systematic Review of Generalization Research in Medical Image Classification. Comput. Biol. Med. 183 , 109256 (2024). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 14 Jan, 2026 Reviews received at journal 05 Jan, 2026 Reviews received at journal 22 Dec, 2025 Reviewers agreed at journal 20 Dec, 2025 Reviews received at journal 20 Dec, 2025 Reviewers agreed at journal 08 Dec, 2025 Reviewers agreed at journal 08 Dec, 2025 Reviewers invited by journal 11 Nov, 2025 Editor invited by journal 10 Nov, 2025 Editor assigned by journal 08 Nov, 2025 Submission checks completed at journal 08 Nov, 2025 First submitted to journal 04 Nov, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8027109","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":543250877,"identity":"97ecebe5-87d6-4891-8ca8-febb1cce7c35","order_by":0,"name":"Jinghui Chen","email":"","orcid":"","institution":"The Affiliated People’s Hospital of Fujian University of Traditional Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Jinghui","middleName":"","lastName":"Chen","suffix":""},{"id":543250878,"identity":"41dc7553-5ee5-40a7-b251-9cf8d7401e08","order_by":1,"name":"Yan Hu","email":"","orcid":"","institution":"The Affiliated People’s Hospital of Fujian University of Traditional Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yan","middleName":"","lastName":"Hu","suffix":""},{"id":543250879,"identity":"5e7e928b-66d3-4c55-95a8-92ea9725b03f","order_by":2,"name":"Tao Yang","email":"","orcid":"","institution":"The Affiliated People’s Hospital of Fujian University of Traditional Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Tao","middleName":"","lastName":"Yang","suffix":""},{"id":543250880,"identity":"20031b3d-7bc0-42d6-be34-c6f474cee326","order_by":3,"name":"Zhipeng Sun","email":"","orcid":"","institution":"The Affiliated People’s Hospital of Fujian University of Traditional Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Zhipeng","middleName":"","lastName":"Sun","suffix":""},{"id":543250881,"identity":"f36c15dc-a0fa-4fa2-aefe-dc1227aa1ad7","order_by":4,"name":"Lianxin Xie","email":"","orcid":"","institution":"The Affiliated People’s Hospital of Fujian University of Traditional Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Lianxin","middleName":"","lastName":"Xie","suffix":""},{"id":543250882,"identity":"b9dae019-2abd-4930-8b2b-e5300a0f81ff","order_by":5,"name":"Hongjia Zhao","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyUlEQVRIiWNgGAWjYBACAxAhYcDAz8De2PjwAylaJBt4DjcbSxCtBQgkGyTS2wR4iNFizt57+IVFgZ2E/MyHbQwSDHZyug0EtFj2nEuzkDBIlmCcndj2oIAh2djsACGH3cgxM5AwYK5jlk5sN5BgOJC4jUgt9RJskgfbJHiI1GL8QMLgsASPBCOxWs6cMQMG8nEJCZ5EYCAbEOOX4z3GnyX+VEvItx9/+PBDhZ0cQS1AwCaNiEEDPOqQAPNHotLJKBgFo2AUjFwAAGbPPS2RO3GDAAAAAElFTkSuQmCC","orcid":"","institution":"The Affiliated People’s Hospital of Fujian University of Traditional Chinese Medicine","correspondingAuthor":true,"prefix":"","firstName":"Hongjia","middleName":"","lastName":"Zhao","suffix":""}],"badges":[],"createdAt":"2025-11-04 09:23:47","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8027109/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8027109/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":96422924,"identity":"597a9134-f620-44ac-913a-696051e6d511","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3825199,"visible":true,"origin":"","legend":"","description":"","filename":"2SR.docx","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/1e5fe4aaee36dc1f2ceb47b4.docx"},{"id":96455044,"identity":"547a3697-0e1d-43ad-90ee-4c35f6bb9d50","added_by":"auto","created_at":"2025-11-21 10:03:27","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7961,"visible":true,"origin":"","legend":"","description":"","filename":"c825c0385b034c5da761059fe03c4afb.json","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/47177a355c2db584c7654429.json"},{"id":96422907,"identity":"4b7d403e-7285-4b0d-bccd-ea6c758628bf","added_by":"auto","created_at":"2025-11-21 01:13:28","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":838232,"visible":true,"origin":"","legend":"","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/cc8b6d2002fbdd1807bee1d4.png"},{"id":96422919,"identity":"58ea5ee4-5b13-4eeb-9b50-aacda970d6f6","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"jpg","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":213312,"visible":true,"origin":"","legend":"","description":"","filename":"10.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/c420e8b70a9cf3e93b27a25d.jpg"},{"id":96422915,"identity":"5416e951-055c-4462-98be-572b9f0708ad","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":347249,"visible":true,"origin":"","legend":"","description":"","filename":"11.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/0854c45b874a4e31bfb16f19.png"},{"id":96454785,"identity":"42a1bf22-b8fc-41c4-9a4e-c5c1eb9c73e2","added_by":"auto","created_at":"2025-11-21 10:03:08","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":367671,"visible":true,"origin":"","legend":"","description":"","filename":"12.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/9bbd587c9709b4b75bcc7f4e.png"},{"id":96455428,"identity":"540e04ee-1e68-4783-8c9c-8fcd7a2efd49","added_by":"auto","created_at":"2025-11-21 10:04:08","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":320817,"visible":true,"origin":"","legend":"","description":"","filename":"13.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/0071dc211f1b3507684aeaac.png"},{"id":96422953,"identity":"84739261-1752-4a4d-a567-7354f37f33da","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":207594,"visible":true,"origin":"","legend":"","description":"","filename":"14.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/a7ad8e7dc3050adc2657059f.png"},{"id":96422933,"identity":"81035bfe-7855-47a6-ae6d-ecda5f2d7c97","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":70587,"visible":true,"origin":"","legend":"","description":"","filename":"15.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/87a68c80460283d0e170f445.png"},{"id":96422941,"identity":"a92f1798-9770-419d-9508-7df4598388b9","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1667451,"visible":true,"origin":"","legend":"","description":"","filename":"16.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/acca01070e32a839234bb8c8.png"},{"id":96422934,"identity":"b183f0a5-d47f-46b5-94c6-99522ca189f9","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":551541,"visible":true,"origin":"","legend":"","description":"","filename":"17.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/f9ecf25c8f50b715c0531b5f.png"},{"id":96422914,"identity":"3019cbd9-907e-4606-b878-932a4832a5ab","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1227972,"visible":true,"origin":"","legend":"","description":"","filename":"18.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/fa23b1e74743cb685dc2c166.png"},{"id":96422930,"identity":"f2d8becb-0aef-48f0-a4c6-c0960f771934","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1028780,"visible":true,"origin":"","legend":"","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/dcfd711434d39ce2cc586638.png"},{"id":96454888,"identity":"bdd0c940-7f81-4014-964c-b0e6c0423ca2","added_by":"auto","created_at":"2025-11-21 10:03:16","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":269493,"visible":true,"origin":"","legend":"","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/438abab94c15e7f46058c2d2.png"},{"id":96454395,"identity":"9e9a306b-f67b-430a-b6de-3ae03a07b6e6","added_by":"auto","created_at":"2025-11-21 10:02:41","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":223285,"visible":true,"origin":"","legend":"","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/aecea41691d2449c2e92bb6b.png"},{"id":96422917,"identity":"1c0a7e09-22dd-4289-9c15-9d53e1d4a7c0","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":156724,"visible":true,"origin":"","legend":"","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/ac7eecafa061ce79d1f4e302.png"},{"id":96422938,"identity":"d464363e-c05d-4580-a7e1-47519782805d","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":211921,"visible":true,"origin":"","legend":"","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/354351472f62f2662aaae78c.png"},{"id":96455020,"identity":"959426b6-63fb-414b-a712-9d12692358db","added_by":"auto","created_at":"2025-11-21 10:03:26","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":311090,"visible":true,"origin":"","legend":"","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/d4533def9ce3ffba9b99a2e6.png"},{"id":96422928,"identity":"ed10ad7e-82aa-4e36-8502-8b534dd0cd09","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":260447,"visible":true,"origin":"","legend":"","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/ed7b942f1045d8cba4e45b9a.png"},{"id":96422960,"identity":"ccdb2a37-afa5-4d8d-90ba-22ac8f61e935","added_by":"auto","created_at":"2025-11-21 01:13:30","extension":"jpg","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":134826,"visible":true,"origin":"","legend":"","description":"","filename":"9.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/88a73cf064978ca28c6beba5.jpg"},{"id":96422978,"identity":"b95eff24-bb16-4106-ae3f-4b97dd4e41a8","added_by":"auto","created_at":"2025-11-21 01:13:30","extension":"xml","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":146397,"visible":true,"origin":"","legend":"","description":"","filename":"c825c0385b034c5da761059fe03c4afb1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/2e32c9dfd73f7da6c1ac1b33.xml"},{"id":96454949,"identity":"a5d204c9-ff40-443b-98d5-4662167a757a","added_by":"auto","created_at":"2025-11-21 10:03:20","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":297643,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/774278b4b4abb68d0e89d600.png"},{"id":96455160,"identity":"596ae5ce-7136-4804-8955-dc6dacbd4bda","added_by":"auto","created_at":"2025-11-21 10:03:43","extension":"jpeg","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":100640,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/52f1256787ca56d2ca20bc89.jpeg"},{"id":96422932,"identity":"f78b7128-5b65-4912-9e3d-921233b0a777","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":164604,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/19101e79f61e14ebc221c1ce.png"},{"id":96422937,"identity":"f1bf21ce-b07d-4fd2-be2f-9ced063f0a02","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":169266,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/944fe8395eb0fd19e6b09854.png"},{"id":96454775,"identity":"16e250a3-a4da-4791-9179-e3b7231cbec8","added_by":"auto","created_at":"2025-11-21 10:03:07","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":228107,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/dc14fcc4cd97a766ad53d911.png"},{"id":96422957,"identity":"ac25d227-9314-45c5-a454-4112a543d70a","added_by":"auto","created_at":"2025-11-21 01:13:30","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":207594,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/328d55e99427ff50c380563e.png"},{"id":96454695,"identity":"d9562cdb-d3e6-4daa-8e6f-85be3c418ecf","added_by":"auto","created_at":"2025-11-21 10:03:03","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":70587,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/8b804f8e03a39ab00b8e271d.png"},{"id":96422939,"identity":"04c8a874-9f05-4111-9fb1-65d090f4c181","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":452462,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/cbe1c8147a5e3e813a49eff3.png"},{"id":96422980,"identity":"03b8a9e3-fe66-480b-bb72-c5dfc89e74b1","added_by":"auto","created_at":"2025-11-21 01:13:31","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":514090,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage17.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/a37d580fac4587add4c07217.png"},{"id":96454985,"identity":"b64b27cd-6347-4bdf-a9cf-906517edd1c5","added_by":"auto","created_at":"2025-11-21 10:03:24","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":486799,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage18.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/2d5f792928f84727ed1067a6.png"},{"id":96454164,"identity":"047ea8b1-8e92-4a0d-962a-c34e0b425a00","added_by":"auto","created_at":"2025-11-21 10:02:25","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":212128,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/e35223696d4e362f3a9b3ddd.png"},{"id":96454828,"identity":"409fbcc1-a287-4fd2-80c5-ddcf03dc8ac5","added_by":"auto","created_at":"2025-11-21 10:03:10","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":160193,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/da0fff4d8708f3fe78f86239.png"},{"id":96422959,"identity":"94333fcf-ad6e-4f79-86dd-198aeb14d8a4","added_by":"auto","created_at":"2025-11-21 01:13:30","extension":"png","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":61339,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/5dfceac0b6aff9f8786ff244.png"},{"id":96422949,"identity":"99bee5e0-9760-45f3-98b4-ac884e65cfe0","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":55604,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/18e7d3d5fa41664ba9de8113.png"},{"id":96422942,"identity":"41518537-0de1-431b-89fb-650346ec5973","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":35,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":154108,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/e73c9229b79c596976363190.png"},{"id":96422962,"identity":"4e815e4d-cf88-4a60-924c-99ce4de87cc2","added_by":"auto","created_at":"2025-11-21 01:13:30","extension":"png","order_by":36,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":152193,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/5160cead9bbb6ba937424128.png"},{"id":96422936,"identity":"f446abea-02a7-4678-b621-549c91065511","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":37,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":158201,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/90969a6b1113c599ce9bc20e.png"},{"id":96454538,"identity":"24eea545-a330-4387-806f-967ac4227026","added_by":"auto","created_at":"2025-11-21 10:02:51","extension":"jpeg","order_by":38,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":91687,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/0d78d61bc375fcb34ff50015.jpeg"},{"id":96453988,"identity":"ffd524a2-2576-4c93-9798-810f99c6f9a3","added_by":"auto","created_at":"2025-11-21 10:02:14","extension":"png","order_by":39,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":62518,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/03704fcaf75be31ab794dce8.png"},{"id":96455006,"identity":"16f28f26-cbcc-48a4-a8a4-f6607750260b","added_by":"auto","created_at":"2025-11-21 10:03:25","extension":"png","order_by":40,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":71980,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/9546700c34aa062f3073d01d.png"},{"id":96422966,"identity":"aad3a77e-6329-4237-8ac4-66aaeb0d8844","added_by":"auto","created_at":"2025-11-21 01:13:30","extension":"png","order_by":41,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":43408,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/2c7947af531c312d65be0e67.png"},{"id":96454802,"identity":"6790a1d9-bc64-44e6-82de-76fc1e057b50","added_by":"auto","created_at":"2025-11-21 10:03:09","extension":"png","order_by":42,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":44166,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/fb7a84129eba01229b65d134.png"},{"id":96454933,"identity":"311e27e0-7aa1-428a-871d-79492f2fdc07","added_by":"auto","created_at":"2025-11-21 10:03:18","extension":"png","order_by":43,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":49738,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/a9d45133c17eb47abd481403.png"},{"id":96422973,"identity":"8424c78a-6a16-4340-b019-10615b4347c4","added_by":"auto","created_at":"2025-11-21 01:13:30","extension":"png","order_by":44,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":200095,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/d1d599d05618984d391bd4ce.png"},{"id":96422954,"identity":"c62a01c1-edcf-4e97-a2bb-8a01c3a04ffe","added_by":"auto","created_at":"2025-11-21 01:13:30","extension":"png","order_by":45,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":65735,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/044c7de767953d88b48e0895.png"},{"id":96422956,"identity":"5be8405d-2d2b-4d50-880d-d07d2f3664c2","added_by":"auto","created_at":"2025-11-21 01:13:30","extension":"png","order_by":46,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":66324,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/54b09bdf5c328e4d90f6c509.png"},{"id":96422961,"identity":"2ca32d0b-6558-4fd8-91bc-eba673bc3373","added_by":"auto","created_at":"2025-11-21 01:13:30","extension":"png","order_by":47,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":52047,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage17.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/abf267a5ba843003a14d556b.png"},{"id":96455448,"identity":"6b1b47a0-9638-4328-aa36-b5e5d34ec1cc","added_by":"auto","created_at":"2025-11-21 10:04:09","extension":"png","order_by":48,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":69458,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage18.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/37fd959a15c7c196bc6a7882.png"},{"id":96454120,"identity":"f9a5ad55-21c4-485f-8b93-b8acbc949cd6","added_by":"auto","created_at":"2025-11-21 10:02:21","extension":"png","order_by":49,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":39853,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/715260ff9973fffc640bcdab.png"},{"id":96454738,"identity":"0b2d757e-2286-45be-9977-4c9219dd07e1","added_by":"auto","created_at":"2025-11-21 10:03:05","extension":"png","order_by":50,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":42976,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/c46b0d103e6abde7aed3eb74.png"},{"id":96422981,"identity":"d81f27bb-537d-419c-ba6d-ca7336b945b9","added_by":"auto","created_at":"2025-11-21 01:13:31","extension":"png","order_by":51,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16374,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/2ef8f4f42f5efbe3226be85a.png"},{"id":96422972,"identity":"31cd4320-d36a-4e86-84fb-edb4b19120a5","added_by":"auto","created_at":"2025-11-21 01:13:30","extension":"png","order_by":52,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":13443,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/7a9aaa28e2f1b16dabbcb104.png"},{"id":96455080,"identity":"93b674d7-2008-47fc-96b8-cacb0b4a4d6c","added_by":"auto","created_at":"2025-11-21 10:03:31","extension":"png","order_by":53,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":43813,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/343cb07559ffcc89a98eb827.png"},{"id":96422947,"identity":"9fcbc6bb-d23a-437a-8925-ae9d07db1cb1","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":54,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":37539,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/50e6c0bf8c2340f6c0646dbc.png"},{"id":96454038,"identity":"38f2ce1f-74b7-4150-a586-341710d2c04e","added_by":"auto","created_at":"2025-11-21 10:02:16","extension":"png","order_by":55,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":37366,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/f9259f79b5a73639c635e235.png"},{"id":96422967,"identity":"08cd4056-b4f0-4ab0-8e4b-67f28308d817","added_by":"auto","created_at":"2025-11-21 01:13:30","extension":"png","order_by":56,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":61488,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/fa560a86f9ca07f02df070bd.png"},{"id":96455032,"identity":"83ad18c1-c5ec-4f54-97b6-6f0de29ca30b","added_by":"auto","created_at":"2025-11-21 10:03:26","extension":"xml","order_by":57,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":146630,"visible":true,"origin":"","legend":"","description":"","filename":"c825c0385b034c5da761059fe03c4afb1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/6f4703fc66094ee7903bffa3.xml"},{"id":96454919,"identity":"b0f58fe6-e5cd-440d-98ad-005bdef76c02","added_by":"auto","created_at":"2025-11-21 10:03:17","extension":"html","order_by":58,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":162124,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/0003941f69e004ee7de0cbe3.html"},{"id":96422911,"identity":"150e0c53-9f13-40dd-a3b2-23155059f966","added_by":"auto","created_at":"2025-11-21 01:13:28","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":838232,"visible":true,"origin":"","legend":"\u003cp\u003eBasic network structure of YOLO11\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/c5ec2ef2ae4732751d539da3.png"},{"id":96422910,"identity":"fc5b2ddd-6010-418b-b704-fc45ad1b018a","added_by":"auto","created_at":"2025-11-21 01:13:28","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1028780,"visible":true,"origin":"","legend":"\u003cp\u003eImproved YOLO11 network structure\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/16a56fc30c277ed3e4ac68c9.png"},{"id":96422905,"identity":"9f8b9f1c-3316-46f3-9637-a624739b92bd","added_by":"auto","created_at":"2025-11-21 01:13:28","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":269493,"visible":true,"origin":"","legend":"\u003cp\u003eBasic module of ShuffleNet v1\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/efaf5ef3ded16d179fbc8833.png"},{"id":96454435,"identity":"296fc915-d663-465c-8dc2-cbaa0b52bf0b","added_by":"auto","created_at":"2025-11-21 10:02:44","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":223285,"visible":true,"origin":"","legend":"\u003cp\u003eDySample network structure\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/4863893447ec406d63c60ba8.png"},{"id":96454796,"identity":"79f19e6d-0b0f-422f-b3a6-e2ddad047b0b","added_by":"auto","created_at":"2025-11-21 10:03:09","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":156724,"visible":true,"origin":"","legend":"\u003cp\u003eDynamic range factor\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/d4b7652c4acd8629cc16914c.png"},{"id":96422920,"identity":"1674f2a6-acd6-4628-bca8-b86a2c8de0c7","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":211921,"visible":true,"origin":"","legend":"\u003cp\u003ePoolFormer network architecture comparison\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/e7324df13bcdc536b4fc6515.png"},{"id":96422925,"identity":"adbf21f9-fb09-4de4-9285-833bf92b4fbb","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":311090,"visible":true,"origin":"","legend":"\u003cp\u003ePoolFormer hierarchical architecture\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/2ce349a63a1f360ef98747f2.png"},{"id":96454462,"identity":"5b92cb2b-7321-4ba3-99a3-620e9b987d1d","added_by":"auto","created_at":"2025-11-21 10:02:47","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":260447,"visible":true,"origin":"","legend":"\u003cp\u003eC3k2-PoolingFormer network architecture\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/b47823e735852a260fdac550.png"},{"id":96454959,"identity":"e1619ed9-1268-4352-88fb-a5563a5307d5","added_by":"auto","created_at":"2025-11-21 10:03:22","extension":"jpg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":134826,"visible":true,"origin":"","legend":"\u003cp\u003eAnnotation status of the training set\u003c/p\u003e","description":"","filename":"9.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/f62bbe117e171b8a8ad9246f.jpg"},{"id":96454570,"identity":"1a615962-1c5b-4d1a-83eb-864e60f23541","added_by":"auto","created_at":"2025-11-21 10:02:54","extension":"jpg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":213312,"visible":true,"origin":"","legend":"\u003cp\u003eAnnotation correlation map of the training set\u003c/p\u003e","description":"","filename":"10.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/b13e7aeebc5a91275215b2e8.jpg"},{"id":96422943,"identity":"2ef05c96-7f5c-45bb-95d3-aabc771f5cba","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":347249,"visible":true,"origin":"","legend":"\u003cp\u003eYOLO11n training curve\u003c/p\u003e","description":"","filename":"11.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/698f26f6f12a453c07d6d3f7.png"},{"id":96422922,"identity":"af69db7a-7c7c-41b7-9056-85f76294ad46","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":367671,"visible":true,"origin":"","legend":"\u003cp\u003eYOLO-LS training curve\u003c/p\u003e","description":"","filename":"12.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/24785099d1c0f53d40b84287.png"},{"id":96454990,"identity":"74bd8618-290f-4271-ae17-d3b88e343204","added_by":"auto","created_at":"2025-11-21 10:03:24","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":320817,"visible":true,"origin":"","legend":"\u003cp\u003eAblation experiment comparison\u003c/p\u003e","description":"","filename":"13.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/4757b477f8db317ded205646.png"},{"id":96422963,"identity":"7f42ffee-806b-43e2-9d50-80b6e5450c7f","added_by":"auto","created_at":"2025-11-21 01:13:30","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":207594,"visible":true,"origin":"","legend":"\u003cp\u003eYOLO series experiment comparison\u003c/p\u003e","description":"","filename":"14.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/6b0a531053d022ce39216c48.png"},{"id":96422912,"identity":"1cbac4f3-422b-489f-a6ec-020873baba88","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":70587,"visible":true,"origin":"","legend":"\u003cp\u003eDifferent network experiment comparison\u003c/p\u003e","description":"","filename":"15.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/4d5da50a45c1fd14ee2f7246.png"},{"id":96454473,"identity":"d4b7bfb6-d16c-40a0-b9ee-d9073f55b38e","added_by":"auto","created_at":"2025-11-21 10:02:48","extension":"png","order_by":16,"title":"Figure 16","display":"","copyAsset":false,"role":"figure","size":1667451,"visible":true,"origin":"","legend":"\u003cp\u003eVisual Comparison of Different Network Experiments\u003c/p\u003e","description":"","filename":"16.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/2cfa4561b997c5b5da55f413.png"},{"id":96422946,"identity":"4a5d3faf-6234-47da-9277-68d8fd57b416","added_by":"auto","created_at":"2025-11-21 01:13:29","extension":"png","order_by":17,"title":"Figure 17","display":"","copyAsset":false,"role":"figure","size":551541,"visible":true,"origin":"","legend":"\u003cp\u003eRecognition and segmentation result comparison of models before and after improvements on the external test set\u003c/p\u003e","description":"","filename":"17.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/7a9a87e26e454ae9c52bd42c.png"},{"id":96455075,"identity":"cb6ccfc0-0e0d-41b8-b1b6-24574619003a","added_by":"auto","created_at":"2025-11-21 10:03:31","extension":"png","order_by":18,"title":"Figure 18","display":"","copyAsset":false,"role":"figure","size":258407,"visible":true,"origin":"","legend":"\u003cp\u003eAttention heatmaps for different tumor types\u003c/p\u003e","description":"","filename":"18.png","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/f18b31515d9da22d66e567b0.png"},{"id":96708894,"identity":"90d4fd88-d7d7-46c5-9e83-df44ccb7a190","added_by":"auto","created_at":"2025-11-25 10:06:01","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":7756349,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8027109/v1/796f0cd9-203b-4027-8d43-b5e1b9b556a6.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"YOLO-LS: A Novel Deep Learning Framework for Brain Tumor Segmentation in Magnetic Resonance Imaging","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eBrain tumors represent a class of heterogeneous diseases occurring within the cranial cavity, often accompanied by rapid progression, neurological impairment, and significant treatment challenges(\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). Brain tumors can be classified into primary and metastatic types, with primary tumors such as gliomas, meningiomas, and pituitary adenomas exhibiting high heterogeneity not only in biological behavior but also in growth location, invasiveness, and treatment response(\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e). For instance, gliomas have a five-year survival rate of only 5\u0026ndash;10%, while meningiomas generally have a better prognosis; however, their invasive and blurred boundaries often lead to incomplete surgical resection, increasing the risk of recurrence(\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e). Early and accurate identification of tumor boundaries and infiltration extent is critical for developing surgical resection, radiotherapy, and chemotherapy plans, as well as for prognosis assessment. Nevertheless, traditional image-based manual annotation is time-consuming and subject to inter-observer variability, failing to meet the demands of large-scale, real-time clinical applications. Magnetic resonance imaging (MRI), with its multi-sequence, multi-parameter, and high soft-tissue contrast characteristics, has become a common imaging modality for brain tumor diagnosis, staging, and preoperative planning(\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e). Therefore, high-precision automated brain tumor recognition and segmentation hold significant value in clinical applications.\u003c/p\u003e\u003cp\u003eIn recent years, methods based on deep learning networks have made substantial progress in MRI brain tumor segmentation tasks. Convolutional neural network (CNN) architectures, such as U-Net and its variants, have become benchmark models widely used in brain tumor challenges. For example, Pourmahboubi et al. proposed a U-Net architecture incorporating pretrained VGG19 weights, which improved segmentation precision of brain tumors in MRI images through transfer learning. This approach achieved superior performance on the TCGA low-grade glioma dataset, outperforming traditional U-Net variants, particularly in handling class imbalance(\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e). Lu et al. developed the SF-SSM UNet model, which serialized MRI slices and utilized the Mamba state space model to capture spatiotemporal correlations, further optimizing feature learning via Fourier transform-based frequency domain analysis. This model demonstrated robustness in multimodal images on the BraTS-2019 dataset(\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e). These CNN-based methods enhanced local feature extraction through attention mechanisms or residual connections but remain limited by fixed receptive fields, hindering effective capture of long-range dependencies. Subsequently, the introduction of Transformer architectures advanced the field of brain tumor segmentation, with their global attention mechanisms aiding in handling tumor heterogeneity and boundary ambiguity. Jia et al. proposed the BiTr-Unet model, combining CNN with Vision Transformer in the U-Net encoder via dual Transformer layers to capture local and long-range dependencies in 3D images, enhanced by CBAM modules for feature extraction(\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e). Zhu et al. designed the multiPI-TransBTS framework, integrating multi-path multi-physics information fusion and adaptive feature modules, which improved Dice coefficients and sensitivity on the BraTS-2020 dataset, particularly in multimodal MRI sequence fusion(\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e). Huang et al. introduced the TMA-TransBTS model with 3D multi-scale self-attention and cross-attention modules, achieving state-of-the-art performance on the BraTS2019 dataset and improving training stability via deep supervision strategies(\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e). Huang et al. proposed a Transformer-based generative adversarial network (GAN), combining ResNet and Transformer blocks in the bottleneck layer with multi-scale L1 loss for discriminator optimization, surpassing traditional methods on the BraTS-2015 dataset and validating generalization on subsequent datasets(\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). These Transformer-based methods achieved breakthroughs in multimodal fusion and global context modeling but suffer from high computational complexity, limiting real-time applications.\u003c/p\u003e\u003cp\u003eThe YOLO (You Only Look Once) series, as a representative single-stage object detection framework(\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e), has garnered widespread attention since its introduction in 2015 for its efficient real-time performance and end-to-end design. This series models object detection as a single regression problem, directly predicting bounding boxes and classes on image grids to balance speed and precision. Its evolution spans multiple generations: YOLOv1 established the foundational architecture(\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e), YOLOv2 introduced anchor boxes and batch normalization(\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e), and YOLOv3 enhanced multi-scale target detection via multi-scale predictions and the Darknet-53 backbone(\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e). Subsequent versions like YOLOv4 and v5 integrated CSPNet, PANet, and Mosaic data augmentation into the backbone and neck networks(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e), continuously optimizing performance and training efficiency. Recent versions from YOLOv6 to 11 have explored reparameterization, anchor-free detection heads, and decoupling of classification and regression tasks, extending to instance segmentation tasks with models such as YOLOv8-Seg and YOLO11-Seg(\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e). These models incorporate mask prediction branches within the detection framework, enabling pixel-level segmentation while maintaining real-time capabilities, providing efficient and robust solutions for fine contour segmentation tasks, including medical imaging.\u003c/p\u003e\u003cp\u003eThe YOLO series, as a real-time object detection framework, has been extended to brain tumor segmentation. Recent studies have explored YOLO variants in MRI brain tumor image processing to improve real-time diagnostic efficiency. Montalbo proposed a transfer learning model based on YOLOv4-Tiny for computer-aided diagnosis of brain tumors, fine-tuned on T1-weighted CE-MRI images to achieve accurate detection of gliomas, meningiomas, and pituitary adenomas(\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e). Abdusalomov et al. integrated CBAM and feature pyramid networks into YOLOv7 for brain tumor detection, emphasizing real-time performance and generalization on multimodal MRI data for gliomas, meningiomas, and pituitary adenomas(\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e). These early works laid the foundation but still face challenges in small tumor detection and boundary ambiguity. Yang et al. developed variants based on YOLOv5s, incorporating ASPP modules for multi-scale context capture and CBAM and CA for enhanced feature attention, achieving precise segmentation of tumor regions on the Kaggle brain tumor dataset while optimizing computational load(\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e). Abraham et al. proposed a feature extraction network based on YOLOv8, improved with dilated convolutions and dual feature pyramid networks (DFPN), combined with dynamic non-monotonic focusing mechanisms to enhance detection of complex tumor morphologies, particularly for small tumors and unclear boundaries in MRI images(\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e). Priyadharshini et al. compared YOLOv9, YOLOv10, and YOLO11 variants, evaluating their detection and segmentation performance on Figshare and BraTS-2020 datasets, highlighting the advantages of multi-scale attention mechanisms and Transformer-based backbones in handling tumor heterogeneity(\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e). These YOLO variants balanced precision and efficiency in edge computing scenarios through attention fusion and lightweight improvements but require further optimization for small lesions and boundary ambiguity. Despite various improvement strategies, several issues remain inadequately addressed: conventional convolutional stacking dilutes fine-grained features of boundary-ambiguous, highly heterogeneous small lesions through repeated pooling; achieving true lightweight deployment while ensuring high precision remains challenging; and differences in imaging protocols and quality across data sources necessitate specific validation of model generalization on external data. Based on these practical needs, this study proposes targeted improvements on YOLO11, naming the improved model YOLO-LS. The strategies include replacing the backbone with ShuffleNetV1 for lightweighting, introducing content-based dynamic upsampling DySample to enhance detail recovery, and replacing the C3k2 module with C3k2-PoolingFormer to optimize cross-scale feature fusion. These improvements aim to maintain or enhance recognition and segmentation precision while significantly reducing computational load and improving identification of small-volume and boundary-ambiguous tumors.\u003c/p\u003e\u003cp\u003eIn summary, the YOLO-LS model proposed in this study, based on targeted improvements to the YOLO11-Seg model, is expected to achieve a better balance between real-time recognition, detection, and high-precision segmentation of brain tumor MRI images, thereby enhancing the model's usability and interpretability in clinical environments. The following sections detail the proposed model architecture, datasets, training details, quantitative and qualitative evaluation results, and comparisons with existing methods along with ablation analyses.\u003c/p\u003e"},{"header":"2 Baseline Model","content":"\u003cp\u003eThe YOLO series, as a classic algorithm in object detection, is renowned for its real-time capabilities, accuracy, and ease of use. It has been widely extended to instance segmentation tasks, particularly suitable for recognition and precise boundary segmentation of brain tumor MRI images. In clinical scenarios, the end-to-end design of this series enables efficient processing of various MRI modalities, achieving tumor localization, type classification, and pixel-level segmentation to assist in real-time diagnosis and preoperative planning. YOLO11, as a current mainstream algorithm for object recognition, detection, and segmentation, offers five scale variants: YOLO11n, YOLO11s, YOLO11m, YOLO11l, and YOLO11x, accommodating diverse deployment needs from lightweight edge devices to high-performance servers. Its core architecture consists of a backbone network (for multi-scale feature extraction), a neck network (for feature fusion and enhancement), and a head network (for bounding box regression and mask prediction), supporting the YOLO11-Seg variant for direct output of instance segmentation results for tumor regions. Compared to previous versions, YOLO11's key innovations include the introduction of the C3k2 module as an optimized CSP variant, which optimizes information flow through feature map splitting and small-kernel convolutions to improve computational efficiency; this module supports parameterized configuration for dual-module serialization to increase feature depth or degradation to YOLOv8's C2f module, indirectly enhancing capture precision of heterogeneous textures in brain tumors; the addition of the C2PSA module fuses CSP structures with partial self-attention (PSA) mechanisms, where features are split via 1\u0026times;1 convolutions\u0026mdash;one path passes directly, the other undergoes PSA multi-scale convolution extraction, SE module channel weighting, and Softmax for point-wise weighting\u0026mdash;to strengthen focus on key tumor regions, significantly improving detection robustness for multi-scale tumors in complex MRI images; the detection head borrows from YOLOv10 by incorporating depthwise separable convolutions to reduce redundant computations and accelerate inference speed, suitable for resource-constrained clinical workstations. However, the baseline YOLO11 model faces challenges in brain tumor MRI segmentation tasks: high computational costs in the backbone network hinder deployment on edge devices; insufficient feature extraction capabilities lead to loss of tumor details; and low-resolution feature maps struggle to capture small-volume, boundary-ambiguous tumor regions, limiting detection accuracy and segmentation precision. In light of these issues, YOLO11-Seg is selected as the baseline model for improvements in this paper, with its network structure shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, laying the foundation for subsequent lightweighting and detail enhancement optimizations tailored to brain tumor characteristics.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"3 Materials and Methods","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e3.1 Improved Model\u003c/h2\u003e\u003cp\u003eAlthough YOLO11 has achieved significant improvements in precision and lightweighting for brain tumor MRI image recognition and segmentation tasks, repeated convolution and pooling operations in its architecture may lead to dilution or loss of fine-grained features in small-volume tumors during feature map compression. Additionally, the high heterogeneity of brain tumors in structure, morphology, edges, and location can induce model bias toward negative predictions, increasing the risk of false negatives. To address these challenges, targeted improvement strategies are proposed in this study to further enhance recognition, detection accuracy, and segmentation precision while maintaining lightweight characteristics. However, directly selecting larger-scale variants would substantially increase parameters and computational complexity, constraining deployment on clinical edge devices. Therefore, optimizations are performed based on the smallest-scale variant \"n\" of the YOLO11-Seg model, with main improvements including: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) replacing the backbone network with ShuffleNet V1 for lightweight feature extraction; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) introducing the DySample dynamic upsampling mechanism to enhance detail recovery; (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) improving the C3k2 module to C3k2-PoolingFormer for optimized cross-scale feature fusion. The improved network structure is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e3.2 ShuffleNet V1 Lightweight Network\u003c/h2\u003e\u003cp\u003eExtending YOLO series models to instance segmentation has achieved excellent balance, but the default backbone still suffers from large parameter counts and high computational complexity, limiting deployment on embedded devices. To overcome this bottleneck, an efficient lightweight architecture is adopted: ShuffleNet v1 replaces the original YOLO11 backbone to construct an instance segmentation network that balances high speed and high precision.\u003c/p\u003e\u003cp\u003eShuffleNet v1, proposed by the Megvii Technology team, is an advanced CNN architecture with extremely high computational efficiency(\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e). Its core concepts involve introducing two novel operations\u0026mdash;pointwise grouped convolution and channel shuffle\u0026mdash;to significantly reduce computational complexity while maintaining model precision. (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) Pointwise grouped convolution reduces the computational complexity of 1\u0026times;1 convolutions by applying grouped convolutions at the 1\u0026times;1 layer, ensuring each convolution operates only on corresponding input channel groups, thereby substantially lowering costs; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) Channel shuffle addresses the lack of inter-group information interaction in group convolutions by uniformly shuffling channel information in input feature maps. Specifically, for a convolutional layer with g groups and g\u0026times;n output channels, the output channel dimension is reshaped to (g, n), transposed, and then flattened as input for the next layer. This operation ensures input and output channels are fully correlated.\u003c/p\u003e\u003cp\u003eAssume an input feature map with width Wi, height Hi, channel count Ci, output channels Co, and kernel size K\u0026times;K. Parameter counts for each convolution type are calculated as follows: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) standard convolution (sconv) parameters: P\u003csub\u003esc\u003c/sub\u003e=C\u003csub\u003ei\u003c/sub\u003e\u0026times;C\u003csub\u003eo\u003c/sub\u003e\u0026times;K\u0026times;K; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) group convolution (gc) parameters: P\u003csub\u003egc\u003c/sub\u003e=\u0026times;C\u003csub\u003eo\u003c/sub\u003e\u0026times;K\u0026times;K; (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) depthwise separable convolution (dw) parameters: P\u003csub\u003edw\u003c/sub\u003e = C\u003csub\u003ei\u003c/sub\u003e\u0026times;K\u0026times;K. Thus, standard convolution has the highest parameter count, group convolution is 1/g of that, and depthwise separable convolution has the lowest, at 1/C\u003csub\u003eo\u003c/sub\u003e of standard. Combining group and depthwise separable convolutions greatly reduces network parameters and computations.\u003c/p\u003e\u003cp\u003eThe ShuffleNet unit is designed based on bottleneck unit principles. In its residual branch, a computationally economical 3\u0026times;3 depthwise separable convolution is applied to the 3\u0026times;3 layer. The first 1\u0026times;1 layer is replaced with pointwise grouped convolution followed by channel shuffle, and the second pointwise grouped convolution restores channel dimensions to match the shortcut path. When stride is applied, a 3\u0026times;3 average pooling is added to the shortcut path, and element-wise addition is replaced with channel concatenation. ShuffleNet V1 is composed of stacked ShuffleNet units divided into three stages, with the first building block in each stage applying stride\u0026thinsp;=\u0026thinsp;2, maintaining other hyperparameters within the stage, and doubling output channels for the next stage. Network connectivity sparsity and complexity are controlled by adjusting group number g and channel scaling factor s to achieve lightweight feature extraction. The basic module network structure of ShuffleNet V1 is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e3.3 Introduction of DySample Upsampling\u003c/h2\u003e\u003cp\u003eYOLO11 employs traditional nearest-neighbor interpolation for upsampling, which may fail to effectively capture feature details and semantic information, impacting model performance in prediction tasks. DySample is a lightweight and efficient upsampling algorithm that uses content-aware dynamic upsampling to adjust sampling positions based on local image features, more accurately restoring feature point positions and details(\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e). The DySample upsampling process consists primarily of generating sampling point set S and dynamic feature resampling, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIts point sampling concept efficiently conserves computational resources. The sampling point generation process is as follows: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) Based on input feature map size and upsampling factor s, a regular grid coordinate G representing initial sampling positions is generated. (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) Dynamic offsets O are generated. Input feature map X first passes through a linear convolutional layer to produce initial offsets, constrained by a static range factor (0.25), then undergoes pixel rearrangement to redistribute channel information to high-resolution spatial dimensions, yielding offset set O. (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) Dynamic offsets O are added to regular grid sampling coordinates G to obtain the final dynamic sampling point set S.\u003c/p\u003e\u003cp\u003eThe dynamic feature resampling process uses PyTorch's built-in grid_sample function with bilinear interpolation on input feature map X based on dynamic sampling point set S to smoothly generate high-resolution feature values, resulting in high-resolution feature map X\u0026prime;.\u003c/p\u003e\u003cp\u003eTo increase offset flexibility, dynamic per-point range factors are generated via linear projection of input features, multiplied point-wise by a sigmoid function and 0.5 static range factor to ensure equivalence with the original static constraint, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e3.4 Improved C3k2 Module\u003c/h2\u003e\u003cp\u003eIn the YOLO11 segmentation framework adopted in this study, the C3k2 module serves as a core component in the neck network, providing multi-scale feature extraction and fusion. It is an enhanced version of the C3 module from YOLOv7 and YOLOv8, designed to optimize feature fusion and computational efficiency through Cross Stage Partial (CSP) structures. The C3k2 module typically includes a backbone path and multiple bottleneck branches, where the backbone path uses convolutional layers for downsampling and channel adjustment, branches extract multi-scale features via repeated bottleneck blocks, and features are fused via concatenation. This design replaces the previous C2f block in YOLO11, further improving speed and accuracy, particularly in real-time object detection tasks. However, in medical imaging tasks like brain tumor segmentation, the C3k2 module has potential limitations: reliance on traditional convolutions may lead to insufficient capture of complex tumor boundaries and fine-grained textures, while computational overhead may be high for high-resolution MRI images, posing challenges for deployment in resource-limited scenarios. Thus, efficient lightweight improvements to the C3k2 module in YOLO11 are urgently needed.\u003c/p\u003e\u003cp\u003ePoolFormer is an efficient visual Transformer variant proposed by Sea AI Lab in 2022(\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e), as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. Based on the MetaFormer concept, it replaces the computationally intensive self-attention mechanism in Transformer architectures with simple spatial pooling as the Token Mixer module, significantly reducing computational complexity and parameters while maintaining competitive performance in tasks such as image classification, object detection, and semantic segmentation. The overall PoolFormer architecture uses a hierarchical design, typically divided into 4 stages, each containing multiple PoolFormer blocks, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e. Each PoolFormer block's core structure includes: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) Patch Embedding: In the first stage, input images are converted to patch tokens via convolution; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) Token Mixer: Average pooling is used for local feature aggregation instead of global attention, focusing more on local spatial relationships and reducing FLOPs; (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) Channel Mixer: Multi-layer perceptron (MLP) layers perform channel-dimensional feature transformations, typically including Layer Normalization, GELU activation, and DropPath regularization; (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) Residual Connections: Each block uses residual connections for training stability. The main computational processes are as follows:\u003c/p\u003e\u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/132203_cef980177e9a226b/132203_custom_files/img1763656867.png\" width=\"695\" height=\"108\"\u003e\u003c/p\u003e\u003cp\u003eWhere L is input data; Pool denotes average pooling; Norm denotes group normalization; \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{W}}_{1}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{W}}_{2}\\)\u003c/span\u003e\u003c/span\u003e are learnable parameters in MLP; σ is the GELU activation function.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eBased on this, the C3k2-PoolingFormer module is proposed to improve the original C3k2 module in YOLO11. Specifically, this module retains the CSP structural framework of C3k2 but replaces bottleneck blocks in branch paths with PoolFormer blocks. The C3k2-PoolingFormer module dynamically configures its processing paths via built-in parameters to address varying task demands. When the parameter is \"True,\" the C3k-PoolingFormer path is activated, with its core advantage lying in utilizing PoolFormer for efficient token interaction, significantly enhancing local feature representation capabilities, excelling in handling brain tumor features in complex MRI backgrounds. When the parameter is \"False,\" the module switches to the PoolFormer path, which enhances global context awareness on a lightweight basis, capturing long-range dependencies crucial for understanding image spatial relationships. The network architecture is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e3.5 Data Acquisition\u003c/h2\u003e\u003cp\u003eMedical imaging data used in this study were sourced from two distinct open public datasets. The first dataset is the Figshare brain tumor dataset provided by Guangzhou Campus of Southern Medical University(\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e), containing 3064 MRI images from 233 patients, covering three main brain tumor types: meningioma (n\u0026thinsp;=\u0026thinsp;708), glioma (n\u0026thinsp;=\u0026thinsp;1426), and pituitary adenoma (n\u0026thinsp;=\u0026thinsp;930). This dataset, named Dataset 1, was used for model training and internal testing to evaluate performance. The second dataset is from the Kaggle platform(\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e), with 300 randomly selected MRI images (100 each of meningioma, glioma, and pituitary adenoma) serving as an external test set for model performance evaluation.\u003c/p\u003e\u003cp\u003eAll images were pre-annotated for segmentation targets using LabelImg by two radiologists with intermediate titles or above, with results reviewed on-site by two radiologists with associate senior titles or above to ensure experimental accuracy. Five-fold cross-validation was employed for model training in this study. From Dataset 1, 2800 images were randomly divided into 5 equal subsets of 560 images each. These were split in an 8:2 ratio into training and validation sets, with 4 subsets as training and the remaining as validation, cycled five times to create five different data distribution combinations. The remaining 264 images served as an internal independent test set to evaluate the optimal model's segmentation performance. The training set was used for parameter learning; the validation set independently monitored the training process and triggered early stopping; the test set was strictly isolated for final performance evaluation only. This strategy detects overfitting tendencies in real time during training, ensuring optimal model performance. Annotation distribution characteristics and inter-class correlations in the training dataset were visualized through distribution statistics in Figs.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e and \u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e3.6 Experimental Environment Configuration\u003c/h2\u003e\u003cp\u003eExperiments were conducted on Windows 11 operating system, with CPU model Intel(R) Core(TM) Ultra 9 275 HX, GPU model NVIDIA GeForce GTX 5070 Laptop, 16G RAM, deep learning framework PyTorch-2.8.0, CUDA-12.9, and Python-3.9.23. Model hyperparameters are shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eModel hyperparameter settings\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOptimizer\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eInitial Learning Rate\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eBatch\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eEpoch\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eloU\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eMomentum\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eWeight Decay\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAdamW\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e0.01\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e16\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e200\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e0.7\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.937\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.0005\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e3.7 Experimental Evaluation Metrics\u003c/h2\u003e\u003cp\u003eMultiple evaluation metrics were adopted in this study to comprehensively assess model performance in brain tumor MRI image recognition and segmentation tasks. These metrics include precision (Precision), recall (Recall), mean average precision (mAP), intersection over union (IoU), and Dice coefficient (Dice Coefficient), along with model computational complexity indicators such as GFLOPs (Giga Floating Point Operations). Here, mAP50 denotes average precision at IoU\u0026thinsp;=\u0026thinsp;0.5, while mAP50-95 represents average precision across IoU from 0.5 to 0.95, providing a stricter and more comprehensive evaluation. These metrics are particularly suitable for medical image processing, quantifying model accuracy and robustness in detecting brain tumor positions, classifying tumor types, and precisely segmenting tumor boundaries, while assessing computational efficiency and lightweighting for clinical deployment. Calculation formulas for evaluation metrics are shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eExperimental evaluation metrics\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"2\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIndicator\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCalculation formula\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003ePrecision\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathbf{P}=\\frac{\\mathbf{T}\\mathbf{P}}{\\mathbf{T}\\mathbf{P}+\\mathbf{F}\\mathbf{P}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eRecall\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathbf{R}=\\frac{\\mathbf{T}\\mathbf{P}}{\\mathbf{T}\\mathbf{P}+\\mathbf{F}\\mathbf{N}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003emean Average Precision(mAP)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathbf{m}\\mathbf{A}\\mathbf{P}=\\frac{1}{\\mathbf{N}}\\sum\\:_{\\mathbf{i}=1}^{\\mathbf{N}}\\mathbf{A}{\\mathbf{P}}_{\\mathbf{i}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eIntersection over Union (IoU)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathbf{I}\\mathbf{o}\\mathbf{U}=\\frac{\\mathbf{T}\\mathbf{P}}{\\mathbf{T}\\mathbf{P}+\\mathbf{F}\\mathbf{P}+\\mathbf{F}\\mathbf{N}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMean Intersection over Union (mIoU)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathbf{m}\\mathbf{I}\\mathbf{o}\\mathbf{U}=\\frac{1}{\\mathbf{N}}\\sum\\:_{\\mathbf{i}=1}^{\\mathbf{N}}\\mathbf{I}\\mathbf{o}{\\mathbf{U}}_{\\mathbf{i}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eDice Coefficient\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathbf{D}\\mathbf{i}\\mathbf{c}\\mathbf{e}=\\frac{2\\times\\:\\mathbf{T}\\mathbf{P}}{2\\times\\:\\mathbf{T}\\mathbf{P}+\\mathbf{F}\\mathbf{P}+\\mathbf{F}\\mathbf{N}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIn the formulas, TP represents true positives (correctly detected or segmented brain tumor pixels/bounding boxes by the model), FP represents false positives (mis-predicted non-tumor regions), FN represents false negatives (missed brain tumor regions), and N is the number of classes. mAP50 was selected as the primary evaluation metric in this experiment to verify the proposed model's effectiveness in brain tumor MRI image recognition and segmentation tasks, thereby comprehensively assessing the method's practicality and computational efficiency. Additionally, to further analyze model reliability in clinical applications, mean IoU (mIoU) and Dice coefficient (Dice) were calculated as supplementary segmentation performance metrics, with GFLOPs compared to evaluate lightweighting efficiency improvements before and after enhancements, ensuring practicality in resource-limited environments.\u003c/p\u003e\u003c/div\u003e"},{"header":"4 Results and Analysis","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e4.1 Ablation Experiments\u003c/h2\u003e\u003cp\u003eTo verify the contributions of the proposed improvement modules to model performance, ablation experiments were conducted based on the YOLO11n model. Experiments used the internal independent test set from Dataset 1 (264 images) for evaluation, with metrics including mAP50, precision (Precision), recall (Recall), mean IoU (mIoU), Dice coefficient (Dice), and GFLOPs. All models were trained under identical experimental environments, hyperparameter settings, and five-fold cross-validation to ensure result robustness. Ablation experiments progressively introduced three main improvements: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) replacing the backbone with ShuffleNet V1; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) introducing DySample upsampling; (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) improving the C3k2 module to C3k2-PoolingFormer. Training curves for the YOLO11n and YOLO-LS models are shown in Figs.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e and \u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eAblation experiment results\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"12\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c12\" colnum=\"12\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eShuffleNetV1\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDysample\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eC3k2-PoolingFormer\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eP(B)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eR(B)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003emAP50(B)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eP(M)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eR(M)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003emAP50(M)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003emIoU\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c11\"\u003e\u003cp\u003eDice\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c12\"\u003e\u003cp\u003eGFLOPs\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.914\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.873\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.924\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.92\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.877\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.929\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.815\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.898\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e9.6\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026radic;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.906\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.893\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.934\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.914\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.892\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.937\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.823\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.903\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e8.4\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026radic;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.916\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.883\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.936\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.922\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.885\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.939\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.823\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.903\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e9.5\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u0026radic;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.913\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.88\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.931\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.92\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.881\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.933\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.818\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e9.2\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026radic;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026radic;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.908\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.903\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.946\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.916\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.947\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.831\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.908\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e8.3\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026radic;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u0026radic;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.905\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.941\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.914\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.896\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.941\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.826\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.905\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e8.2\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026radic;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u0026radic;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.915\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.943\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.922\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.889\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.943\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.827\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.905\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e8.9\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026radic;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026radic;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u0026radic;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.907\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.953\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.916\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.904\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.951\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.835\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e8.1\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eBased on the ablation experiment results in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e, the progressive introduction of the three improvement modules significantly enhanced YOLO11n model performance in brain tumor MRI image segmentation tasks. The baseline YOLO11n model performed well, with mAP50(B) of 0.924, precision(B) of 0.914, recall(B) of 0.873, mIoU of 0.815, Dice of 0.898, and GFLOPs of 9.6, but showed room for improvement in recall and segmentation precision for small or boundary-ambiguous tumors. After replacing the backbone with ShuffleNet V1, precision(B) slightly decreased to 0.906, but recall(B) increased to 0.893, mAP50(B) to 0.934, mIoU to 0.823, Dice to 0.903, and GFLOPs decreased by 12.5% to 8.4, demonstrating effective multi-scale feature extraction under lightweight design. Introducing DySample upsampling further improved performance, with mAP50(B) reaching 0.936, recall(B) 0.883, mIoU and Dice both 0.903, and GFLOPs 9.5, indicating that its dynamic sampling mechanism significantly enhanced detail retention for small-target tumors. When using the C3k2-PoolingFormer module alone, mAP50(B) increased to 0.931, mIoU to 0.818, Dice to 0.900, and GFLOPs to 9.2, optimizing feature fusion for complex textures via pooling operations. Integrating all improvements yielded the best performance: mAP50(B) of 0.953, recall(B) of 0.910, mIoU of 0.835, Dice of 0.910, and GFLOPs reduced to 8.1, a 15.6% decrease from baseline. Overall, improvements elevated mAP50(B), mIoU, and Dice by 2.9%, 2.0%, and 1.2%, respectively. Training curves before and after improvements showed faster convergence and higher robustness, indicating an excellent balance between precision and computational efficiency, suitable for lightweight deployment in clinical environments.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e4.2 YOLO Series Comparison Experiments\u003c/h2\u003e\u003cp\u003eTo further validate the superiority of the YOLO-LS model in brain tumor MRI image recognition and segmentation tasks, comparison experiments were conducted with other mainstream YOLO series versions, including YOLOv5, YOLOv8, YOLO12, and the baseline YOLO11 model. All comparison models were evaluated on the same internal independent test set from Dataset 1 (264 images), under identical experimental environments, hyperparameter settings, and five-fold cross-validation. Evaluation metrics aligned with those in ablation experiments. The experiments aimed to examine performance across different YOLO versions in precision, segmentation accuracy, and computational efficiency, particularly for complex tumor morphologies and boundary ambiguity.\u003c/p\u003e\u003cp\u003eComparison results show that the YOLO-LS model outperformed baseline models across multiple metrics. Specifically, the improved model significantly reduced computational complexity while enhancing detection and segmentation precision. This indicates that the improvement strategies effectively addressed feature extraction deficiencies and small-target detection challenges in original YOLO series for medical imaging tasks. Detailed comparison results are shown in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e14\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eYOLO series comparison experiment results\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"10\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eP(B)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eR(B)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003emAP50(B)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eP(M)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eR(M)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003emAP50(M)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003emIoU\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eDice\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003eGFLOPs\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eYOLOv5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.892\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.851\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.908\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.898\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.855\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.913\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.795\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.876\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e7.3\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eYOLOv8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.905\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.885\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.931\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.911\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.889\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.935\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.822\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e11.3\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eYOLO11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.914\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.873\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.924\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.92\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.877\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.929\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.815\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.898\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e9.6\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eYOLO12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.898\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.868\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.919\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.904\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.872\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.924\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.809\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.888\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e9.8\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eYOLO-LS\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.907\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.953\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.916\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.904\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.951\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.835\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e8.1\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003e4.3 Comparison Experiments with Different Networks\u003c/h2\u003e\u003cp\u003eTo comprehensively evaluate the overall performance of the YOLO-LS model in brain tumor MRI image recognition and segmentation tasks, comparisons were made with various non-YOLO series networks widely used in medical image segmentation, including U-Net, HRNet, SegNet, and Deeplab V3+. All models were evaluated on the same internal independent test set from Dataset 1, with consistent experimental settings. Comparison results are shown in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e15\u003c/span\u003e.\u003c/p\u003e\u003cp\u003eFrom the comparison results in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e15\u003c/span\u003e, YOLO-LS exhibited outstanding performance across multiple key metrics. Specifically, YOLO-LS surpassed U-Net, SegNet, and HRNet in precision and recall, second only to Deeplab V3+. In segmentation quality, YOLO-LS achieved mIoU of 0.835 and Dice of 0.91, comparable to Deeplab V3\u0026thinsp;+\u0026thinsp;but significantly higher than U-Net and SegNet. Notably, YOLO-LS demonstrated substantial advantages in computational efficiency, with GFLOPs of only 8.1, far below U-Net (38.5), SegNet (25.3), HRNetV2 (49.7), and Deeplab V3+ (43.8). This indicates that YOLO-LS maintained high segmentation precision while significantly reducing computational resource demands, showing potential for deployment in resource-limited clinical environments. Overall, YOLO-LS achieved a good balance among precision, recall, segmentation quality, and computational efficiency. Although slightly inferior to Deeplab V3\u0026thinsp;+\u0026thinsp;in some metrics, its extremely low computational complexity makes it more advantageous in practical applications, particularly in clinical scenarios requiring real-time or near-real-time processing. This highlights the practical potential of YOLO-LS's real-time advantages combined with lightweight improvements in medical imaging tasks.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eDifferent network comparison experiment results\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003emIoU\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eDice\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eGFLOPs\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eU-Net\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.895\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.865\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.818\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.888\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e38.5\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.88\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.83\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e49.7\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSegNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.875\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.842\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.785\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.865\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e25.3\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDeeplab V3+\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.922\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.892\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.842\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.915\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e43.8\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eYOLO-LS\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.916\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.904\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.835\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e8.1\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo visually demonstrate the performance of different network models in brain tumor MRI image segmentation tasks, segmentation effect visualizations were used for qualitative analysis. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e16\u003c/span\u003e, these visualizations clearly present each model's capabilities in handling tumor boundaries, morphologies, and heterogeneous textures, providing intuitive evidence of the proposed YOLO-LS model's superiority in precise segmentation compared to U-Net, HRNet, SegNet, and Deeplab V3+. This visualization analysis complements the quantitative metrics in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, further emphasizing the improved model's potential value in clinical diagnosis and treatment planning for precise tumor segmentation.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003e4.4 Generalization Capability Evaluation\u003c/h2\u003e\u003cp\u003eTo evaluate the generalization capability of the proposed YOLO-LS model in brain tumor MRI image segmentation tasks, Dataset 2 was used as an external independent test set. This dataset originates from different sources, differing from training Dataset 1 in imaging equipment, patient populations, and image quality, making it suitable for testing model robustness on unseen data. The YOLO-LS model, using the best weights trained on Dataset 1, was directly applied to Dataset 2 for performance evaluation without additional fine-tuning. Evaluation metrics aligned with previous experiments. Specific metrics for each tumor type and overall performance of the YOLO-LS model on the external independent test set are shown in Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eExternal test set generalization capability evaluation results\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"9\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCategory\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eP(B)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eR(B)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003emAP50(B)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eP(M)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eR(M)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003emAP50(M)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003emIoU\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eDice\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMeningioma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.902\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.905\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.938\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.902\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.936\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.828\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.903\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGlioma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.885\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.888\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.928\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.892\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.885\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.925\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.812\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.888\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePituitary\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.908\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.912\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.942\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.915\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.908\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.83\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.905\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAll\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.898\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.902\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.936\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.906\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.898\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.934\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.82\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.895\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eBased on the evaluation results in Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, the YOLO-LS model exhibited excellent overall performance on the external independent test set, with bounding box detection precision (P(B)) of 0.898, recall (R(B)) of 0.902, and mAP50(B) of 0.936; mask segmentation precision (P(M)) of 0.906, recall (R(M)) of 0.898, and mAP50(M) of 0.934. For segmentation quality, mean IoU (mIoU) was 0.82, and Dice coefficient was 0.895. These metrics indicate that the model maintained high levels of recognition, detection, and segmentation performance on unseen data, validating its strong generalization capability.\u003c/p\u003e\u003cp\u003eBy tumor type, pituitary adenoma showed the best recognition and segmentation, with mAP50(B) of 0.942 and Dice of 0.905; meningioma followed, with mAP50(B) of 0.938 and Dice of 0.903; glioma metrics were slightly lower but still reached mAP50(B) of 0.928 and Dice of 0.888. This demonstrates the YOLO-LS model's good adaptability to different pathological types of brain tumors, effectively handling image differences from various data sources. However, glioma showed relatively lower recall and Dice on the external test set, possibly due to its often blurred boundaries and strong heterogeneity in MRI, though overall performance remained acceptable.\u003c/p\u003e\u003cp\u003eTo more intuitively display the YOLO-LS model's recognition and segmentation effects on the external test set, visualization results are provided as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e17\u003c/span\u003e. This figure compares segmentation outputs of models before and after improvements on typical samples, clearly showing YOLO-LS's advantages in tumor boundary localization and detail recovery, further validating its potential in practical clinical applications.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e4.5 Heatmaps\u003c/h2\u003e\u003cp\u003eTo deeply analyze the model's decision-making process and focus on key brain tumor regions, Grad-CAM (Gradient-weighted Class Activation Mapping) was employed to generate heatmaps. This method visualizes the model's emphasis on different input image regions during prediction by computing gradient weights, with red areas indicating high attention (typically corresponding to tumor boundaries and textures) and blue areas indicating low attention. Heatmaps were generated based on YOLO-LS model predictions on MRI images from Dataset 1's internal test set, aiming to verify correct focus on lesion areas rather than background noise or irrelevant structures. Heatmap displays for the improved model on MRI images of different brain tumor types are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig18\" class=\"InternalRef\"\u003e18\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig18\" class=\"InternalRef\"\u003e18\u003c/span\u003e heatmap examples (top: label visualization of original images; middle: YOLO11-Seg model heatmap results; bottom: YOLO-LS model heatmap results after improvements). Results show that YOLO-LS heatmaps more precisely highlight tumor core and edge regions, with more concentrated attention distribution, avoiding dispersion to non-tumor areas. This benefits from the C3k2-PoolingFormer module's pooling operations enhancing local feature aggregation and DySample upsampling's detail retention capabilities. For example, in glioma images, heatmaps clearly show model focus on heterogeneous textures and invasive boundaries; in pituitary adenoma images, attention centers on tumor core positions, reducing mis-focus on adjacent normal tissues. Notably, in small tumor cases, improved model activation coverage is higher, further explaining its recall advantages. These visualization results not only validate model interpretability but also provide references for clinicians to understand AI decision bases.\u003c/p\u003e\u003c/div\u003e"},{"header":"5 Discussion","content":"\u003cp\u003eThe YOLO-LS improvement framework based on YOLO11 is proposed in this study, significantly enhancing efficiency in brain tumor MRI image recognition and segmentation through integration of ShuffleNet V1 lightweight backbone, DySample dynamic upsampling mechanism, and C3k2-PoolingFormer module. Experimental results indicate that the model achieved mAP50 of 0.953 on the internal test set with a 15.6% reduction in GFLOPs, also demonstrating good generalization on the external test set. These improvements address clinical challenges such as tumor heterogeneity, boundary ambiguity, and small-volume lesions, achieving an effective balance between precision and lightweighting, particularly suitable for resource-limited clinical environments.\u003c/p\u003e\u003cp\u003eThe outstanding performance of the YOLO-LS model stems from the targeted improvement strategies proposed in this study. The introduction of the ShuffleNet V1 backbone, through pointwise grouped convolutions and channel shuffle operations, substantially reduced parameters and computations with minimal loss in feature representation. This aligns with the growing demand for lightweight models in medical image analysis(\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e). The DySample dynamic upsampling mechanism replaces traditional interpolation methods, with its content-aware sampling point generation strategy effectively mitigating detail loss due to position insensitivity during upsampling, crucial for precisely delineating brain tumors, especially small lesions with blurred, infiltrative boundaries. The innovation of the C3k2-PoolingFormer module lies in introducing PoolFormer's simple and efficient pooling-based Token Mixer mechanism into YOLO's neck network, enhancing capture of long-distance contextual information without the high computational costs of Transformers, better addressing brain tumor heterogeneity. Ablation experiments clearly demonstrate the independent contributions of each module and their synergistic effects. Furthermore, a prominent contribution of this study is the model's exceptional computational efficiency. Compared to mature segmentation architectures like U-Net, HRNet, SegNet, and Deeplab V3+, YOLO-LS achieves comparable or higher precision with only a fraction of their computational overhead. For example, YOLO-LS reaches a Dice score of 0.91 with only 8.1 GFLOPs, while the widely used U-Net requires 38.5 GFLOPs for a slightly lower Dice of 0.888. This nearly fivefold reduction in computational demand highlights the model's suitability for real-time clinical applications and deployment on resource-constrained hardware, crucial for expanding access to sophisticated diagnostic technologies.\u003c/p\u003e\u003cp\u003eBeyond performance on known datasets, model generalization is essential for clinical utility(\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e). YOLO-LS was evaluated on an independent external dataset without retraining or fine-tuning, with results confirming its robustness. The model maintained high performance across different tumor types from new data sources, with an overall Dice score of 0.895. This indicates that YOLO-LS has learned fundamental, transferable imaging biomarkers for brain tumors rather than overfitting specific training data features, enhancing its potential for real-world deployment(\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e, \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e). Additionally, to ensure the model is not a \"black box,\" Grad-CAM was used for visualization explanations. Generated heatmaps provide strong evidence that architectural improvements directly translate to more clinically relevant decision processes. Compared to the baseline, YOLO-LS exhibits more precise and focused attention on tumor cores and boundaries while ignoring irrelevant background anatomy. This enhanced focus is attributed to C3k2-PoolingFormer's improved feature aggregation and DySample's detail retention, explaining observed segmentation precision improvements and offering clinicians a visualization tool to build trust in model predictions.\u003c/p\u003e\u003cp\u003eAlthough the YOLO-LS model achieved significant progress, limitations persist. First, the training dataset primarily relies on T1-weighted CE-MRI sequences, which enhance tumor regions well, but complete brain tumor assessment often requires complementary information from multi-modal MRI sequences like T2 and FLAIR; future extensions to multimodal fusion are needed for enhanced robustness. Second, the external test set is relatively small; while validating generalization, further evaluation on larger-scale, multi-center clinical datasets is required. Generalization robustness of deep learning models across institutions, scanners, and imaging parameters remains a major challenge in the field(\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e). To prove true clinical applicability, validation on larger-scale, multi-center, prospectively collected real-world datasets is essential. Looking ahead, the following directions warrant deeper exploration: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) Multimodal and temporal information fusion: Explore effective fusion strategies integrating T1, T2, FLAIR, DWI, and other sequences, and further investigate utilization of dynamic contrast-enhanced MRI (DCE-MRI) or diffusion tensor imaging (DTI) temporal data to provide richer imaging biomarkers for treatment response assessment and prognosis prediction. (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) Personalized adaptive learning: Develop online or meta-learning strategies enabling rapid fine-tuning based on specific patient follow-up imaging data, adapting to tumor morphological evolution during treatment for personalized precise segmentation. (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) Deep integration with clinical workflows: Interface segmentation results with radiation therapy planning systems (RTPS) or neuronavigation systems for automatic tumor volume delineation and dose calculation, or assist in surgical path planning, truly leveraging AI in clinical decision support. (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) Prospective clinical validation: Design rigorous clinical trials to quantitatively assess impacts of introducing YOLO-LS on radiologist diagnostic efficiency, delineation consistency, and ultimate patient treatment outcomes. (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e) Collaborative training under privacy protection: To address privacy and compliance barriers in multi-center data sharing, introduce federated learning and other privacy-preserving techniques for cross-institutional collaborative training without moving raw data, fully utilizing distributed medical data resources to further enhance model generalization and robustness.\u003c/p\u003e\u003cp\u003eOverall, the YOLO-LS model proposed in this study provides an efficient, lightweight, and interpretable solution for AI-assisted brain tumor diagnosis. It not only demonstrates feasibility of high-precision segmentation under lightweight constraints but also lays a solid technical foundation for subsequent research and clinical deployment. Through continued multimodal expansions, large-scale clinical validations, and seamless integration with clinical workflows, this framework holds promise as a reliable tool for neuroradiologists, ultimately contributing to improved brain tumor patient care processes and outcomes.\u003c/p\u003e"},{"header":"6 Conclusion","content":"\u003cp\u003eA YOLO-LS model based on improved YOLO11 is proposed in this study for brain tumor MRI image recognition and segmentation. Through innovative integration of ShuffleNet V1 lightweight backbone, DySample dynamic upsampling, and C3k2-PoolingFormer module, the model achieved excellent performance with mAP50\u0026thinsp;=\u0026thinsp;0.953 and Dice\u0026thinsp;=\u0026thinsp;0.91 on Figshare and Kaggle datasets, with GFLOPs reduced to 8.1, representing a 2.9% precision improvement and 15.6% computational load reduction compared to the baseline. Ablation experiments and comparative analyses confirmed contributions of each improvement module, with heatmap visualizations further validating model interpretability and precise focus on tumor boundaries. These results indicate that YOLO-LS achieves breakthroughs in balancing real-time capabilities and high precision, showing significant clinical application potential, especially for auxiliary diagnosis on edge devices. Future work will focus on multimodal data fusion, larger-scale clinical validation, and integration with radiation therapy planning to further enhance model generalization and practicality. This framework provides a novel tool for precision medicine in brain tumors, poised to improve patient prognosis and reduce healthcare burdens.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFunding\u003c/h2\u003e\u003cp\u003eNo Funding\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eJ.C. performed the experiments, analyzed the data, prepared the figures, and wrote the original draft. Y.H., T.Y., Z.S., and L.X. contributed to conceptualization, methodology, data curation, software implementation, and validation. H.Z. supervised the project, provided resources, and reviewed and edited the manuscript. All authors reviewed the manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets analyzed during the current study are available in the following public repositories: The Figshare brain tumor dataset is available at https://www.kaggle.com/datasets/ashkhagan/figshare-brain-tumor-dataset (as referenced in citations 28 and 29). The Kaggle brain tumor MRI dataset is available at https://www.kaggle.com/datasets/masoudnickparvar/brain-tumor-mri-dataset (as referenced in citation 30). No new datasets were generated in this study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eVankdothu, R., Hameed, M. A. \u0026amp; Fatima, H. A Brain Tumor Identification and Classification Using Deep Learning Based on Cnn-Lstm Method. \u003cem\u003eComput. Electr. Eng.\u003c/em\u003e \u003cb\u003e101\u003c/b\u003e, 107960. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.compeleceng.2022.107960\u003c/span\u003e\u003cspan address=\"10.1016/j.compeleceng.2022.107960\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNazir, M., Shakil, S. \u0026amp; Khurshid, K. Role of Deep Learning in Brain Tumor Detection and Classification (2015 to 2020): A Review. \u003cem\u003eComput. Med. Imaging Graph.\u003c/em\u003e \u003cb\u003e91\u003c/b\u003e, 101940. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.compmedimag.2021.101940\u003c/span\u003e\u003cspan address=\"10.1016/j.compmedimag.2021.101940\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKumar, S. \u0026amp; Mankame, D. P. Optimization Driven Deep Convolution Neural Network for Brain Tumor Classification. \u003cem\u003eBiocybernetics Biomedical Eng.\u003c/em\u003e \u003cb\u003e40\u003c/b\u003e (3), 1190\u0026ndash;1204. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.bbe.2020.05.009\u003c/span\u003e\u003cspan address=\"10.1016/j.bbe.2020.05.009\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKarayegen, G. \u0026amp; Aksahin, M. F. Brain Tumor Prediction on Mr Images with Semantic Segmentation by Using Deep Learning Network and 3d Imaging of Tumor Region. \u003cem\u003eBiomed. Signal Process. Control\u003c/em\u003e. \u003cb\u003e66\u003c/b\u003e, 102458. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.bspc.2021.102458\u003c/span\u003e\u003cspan address=\"10.1016/j.bspc.2021.102458\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOstrom, Q. T. et al. Cbtrus Statistical Report: Primary Brain and Other Central Nervous System Tumors Diagnosed in the United States in 2015\u0026ndash;2019. \u003cem\u003eNeuro-Oncology\u003c/em\u003e 24(Supplement_5):v1-v95. (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/neuonc/noac202\u003c/span\u003e\u003cspan address=\"10.1093/neuonc/noac202\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVankdothu, R. \u0026amp; Hameed, M. A. Brain Tumor Mri Images Identification and Classification Based on the Recurrent Convolutional Neural Network. \u003cem\u003eMeasurement: Sens.\u003c/em\u003e \u003cb\u003e24\u003c/b\u003e, 100412. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.measen.2022.100412\u003c/span\u003e\u003cspan address=\"10.1016/j.measen.2022.100412\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRajinikanth, V., Kadry, S. \u0026amp; Nam, Y. Convolutional-Neural-Network Assisted Segmentation and Svm Classification of Brain Tumor in Clinical Mri Slices. \u003cem\u003eInform. Technol. Control\u003c/em\u003e. \u003cb\u003e50\u003c/b\u003e (2), 342\u0026ndash;356. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.5755/j01.itc.50.2.28087\u003c/span\u003e\u003cspan address=\"10.5755/j01.itc.50.2.28087\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePourmahboubi, A., Arsalani Saeed, N. \u0026amp; Tabrizchi, H. A Brain Tumor Segmentation Enhancement in Mri Images Using U-Net and Transfer Learning. \u003cem\u003eBMC Med. Imaging\u003c/em\u003e. \u003cb\u003e25\u003c/b\u003e (1), 307. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12880-025-01837-4\u003c/span\u003e\u003cspan address=\"10.1186/s12880-025-01837-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLu, J. et al. A Sequential Flow Unet for Mri Brain Tumor Segmentation Based on State-Space-Model. \u003cem\u003eAppl. Soft Comput.\u003c/em\u003e \u003cb\u003e186\u003c/b\u003e, 114069. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.asoc.2025.114069\u003c/span\u003e\u003cspan address=\"10.1016/j.asoc.2025.114069\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJia, Q. \u0026amp; Shu, H. (eds) Bitr-Unet: A Cnn-Transformer Combined Network For mri Brain Tumor Segmentation. Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries; (2022). 2022//; Cham: Springer International Publishing.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhu, H. et al. A Multi-Path Learning Framework for Brain Tumor Image Segmentation Based on Multi-Physical Information. \u003cem\u003eComput. Biol. Med.\u003c/em\u003e \u003cb\u003e191\u003c/b\u003e, 110148. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.compbiomed.2025.110148\u003c/span\u003e\u003cspan address=\"10.1016/j.compbiomed.2025.110148\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHuang, Y., Chen, L. \u0026amp; Zhou, C. Multi-Modal Brain Tumor Segmentation Via 3d Multi-Scale Self-Attention and Cross-Attention. \u003cem\u003earXiv preprint arXiv:250409088\u003c/em\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHuang, L. et al. A Transformer-Based Generative Adversarial Network for Brain Tumor Segmentation. \u003cem\u003eFront. NeuroSci.\u003c/em\u003e 16. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fnins.2022.1054948\u003c/span\u003e\u003cspan address=\"10.3389/fnins.2022.1054948\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKotthapalli, M., Ravipati, D. \u0026amp; Bhatia, R. Yolov1 to Yolov11: A Comprehensive Survey of Real-Time Object Detection Innovations and Challenges. \u003cem\u003earXiv preprint arXiv:250802067\u003c/em\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhao, Z. Q., Zheng, P., Xu, S. T. \u0026amp; Wu, X. Object Detection with Deep Learning: A Review. \u003cem\u003eIEEE Trans. Neural Networks Learn. Syst.\u003c/em\u003e \u003cb\u003e30\u003c/b\u003e (11), 3212\u0026ndash;3232. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TNNLS.2018.2876865\u003c/span\u003e\u003cspan address=\"10.1109/TNNLS.2018.2876865\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLuo, Y. \u003cem\u003eThe Evolution of Yolo: From Yolov1 to Yolov11 with a Focus on Yolov7's Innovations in Object Detection, Theoretical and Natural Science, 87, 1, 82\u0026ndash;90\u003c/em\u003e,. (2025). (2025). (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAli, M. L. \u0026amp; Zhang, Z. \u003cem\u003eThe Yolo Framework: A Comprehensive Review of Evolution, Applications, and Benchmarks in Object Detection\u003c/em\u003e (Preprints, 2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.20944/preprints202410.1785.v1\u003c/span\u003e\u003cspan address=\"10.20944/preprints202410.1785.v1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMurat, A. A. \u0026amp; Kiran, M. S. A Comprehensive Review on Yolo Versions for Object Detection. \u003cem\u003eEng. Sci. Technol. Int. J.\u003c/em\u003e \u003cb\u003e70\u003c/b\u003e, 102161 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eUltralytics Yolov11: Ultralytics Yolo Documentation \u0026ndash; Instance Segmentation: Ultralytics [2025-11-02]. (2024). Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://docs.ultralytics.com/tasks/segment/\u003c/span\u003e\u003cspan address=\"https://docs.ultralytics.com/tasks/segment/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMontalbo, F. J. A Computer-Aided Diagnosis of Brain Tumors Using a Fine-Tuned Yolo-Based Model with Transfer Learning. \u003cem\u003eKSII Trans. Internet Inf. Syst.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e, 4816\u0026ndash;4834. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3837/tiis.2020.12.011\u003c/span\u003e\u003cspan address=\"10.3837/tiis.2020.12.011\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAbdusalomov, A. B., Mukhiddinov, M. \u0026amp; Whangbo, T. K. Brain Tumor Detection Based on Deep Learning Approaches and Magnetic Resonance Imaging. \u003cem\u003eCancers (Basel)\u003c/em\u003e 15(16). Epub 20230818. (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/cancers15164172\u003c/span\u003e\u003cspan address=\"10.3390/cancers15164172\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang, T. et al. Application of Mri Image Segmentation Algorithm for Brain Tumors Based on Improved Yolo. \u003cem\u003eFront. Neurosci.\u003c/em\u003e \u003cb\u003e18\u003c/b\u003e, 1510175. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fnins.2024.1510175\u003c/span\u003e\u003cspan address=\"10.3389/fnins.2024.1510175\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024). Epub 20250107.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAbraham, L. A., Palanisamy, G. \u0026amp; Goutham, V. Dilated Convolution and Yolov8 Feature Extraction Network: An Improved Method for Mri-Based Brain Tumor Detection. \u003cem\u003eIEEE Access.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 27238\u0026ndash;27256. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ACCESS.2025.3539924\u003c/span\u003e\u003cspan address=\"10.1109/ACCESS.2025.3539924\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePriyadharshini, S., Bhoopalan, R., Manikandan, D. \u0026amp; Ramaswamy, K. A Successive Framework for Brain Tumor Interpretation Using Yolo Variants. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e (1), 27973. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-025-13155-4\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-13155-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang, X., Zhou, X., Lin, M. \u0026amp; Sun, J. (eds) Shufflenet: An Extremely Efficient Convolutional Neural Network for Mobile Devices. \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition\u003c/em\u003e; (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu, W., Lu, H., Fu, H. \u0026amp; Cao, Z. (eds) Learning to Upsample by Learning to Sample. \u003cem\u003eProceedings of the IEEE/CVF international conference on computer vision\u003c/em\u003e; (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYu, W. et al. (eds) Metaformer Is Actually What You Need for Vision. \u003cem\u003eProceedings of the IEEE/CVF conference on computer vision and pattern recognition\u003c/em\u003e; (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAshkhagan Figshare Brain Tumor Dataset: Kaggle [2025-11-02]. (2025). Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.kaggle.com/datasets/ashkhagan/figshare-brain-tumor-dataset\u003c/span\u003e\u003cspan address=\"https://www.kaggle.com/datasets/ashkhagan/figshare-brain-tumor-dataset\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCheng, J. et al. Retrieval of Brain Tumors by Adaptive Spatial Pooling and Fisher Vector Representation. \u003cem\u003ePLOS ONE\u003c/em\u003e. \u003cb\u003e11\u003c/b\u003e (6), e0157112. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pone.0157112\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0157112\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2016).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNickparvar, M. Brain Tumor Mri Dataset: Kaggle [2025-11-02]. (2021). Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.kaggle.com/datasets/masoudnickparvar/brain-tumor-mri-dataset\u003c/span\u003e\u003cspan address=\"https://www.kaggle.com/datasets/masoudnickparvar/brain-tumor-mri-dataset\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAl-Milaji, Z. \u0026amp; Yousif, H. Lightweight Deep Learning Model Optimization for Medical Image Analysis. \u003cem\u003eInt. J. Imaging Syst. Technol.\u003c/em\u003e \u003cb\u003e34\u003c/b\u003e (5), e23173 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTran, A. T., Zeevi, T. \u0026amp; Payabvash, S. Strategies to Improve the Robustness and Generalizability of Deep Learning Segmentation and Classification in Neuroimaging. \u003cem\u003eBioMedInformatics\u003c/em\u003e \u003cb\u003e5\u003c/b\u003e (2), 20 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDorfner, F. J., Patel, J. B., Kalpathy-Cramer, J., Gerstner, E. R. \u0026amp; Bridge, C. P. A Review of Deep Learning for Brain Tumor Analysis in Mri. \u003cem\u003eNPJ Precision Oncol.\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e (1), 2 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAiya, A. J. et al. Optimized Deep Learning for Brain Tumor Detection: A Hybrid Approach with Attention Mechanisms and Clinical Explainability. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e (1), 31386. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-025-04591-3\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-04591-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAziz, N. et al. Precision Meets Generalization: Enhancing Brain Tumor Classification Via Pretrained Densenet with Global Average Pooling and Hyperparameter Tuning. \u003cem\u003ePLOS ONE\u003c/em\u003e. \u003cb\u003e19\u003c/b\u003e (9), e0307825. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pone.0307825\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0307825\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMatta, S. et al. A Systematic Review of Generalization Research in Medical Image Classification. \u003cem\u003eComput. Biol. Med.\u003c/em\u003e \u003cb\u003e183\u003c/b\u003e, 109256 (2024).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"brain tumor, artificial intelligence, image segmentation, YOLO11, magnetic resonance, YOLO-LS","lastPublishedDoi":"10.21203/rs.3.rs-8027109/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8027109/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eBrain tumors, as highly heterogeneous intracranial diseases, require accurate recognition and segmentation for clinical diagnosis, surgical planning, and prognosis evaluation. However, manual annotation of traditional MRI images suffers from subjective bias and low efficiency. This study proposes an improved algorithm, YOLO-LS, based on the YOLO11-Seg model to enhance real-time recognition, detection, and high-precision segmentation of brain tumor MRI images. Specific improvements include the adoption of ShuffleNet V1 as a lightweight backbone network to reduce parameters and computational complexity; the introduction of the DySample dynamic upsampling mechanism to enhance detail recovery; and the optimization of the C3k2 module into a C3k2-PoolingFormer block for efficient cross-scale feature fusion. Experiments were conducted using the Figshare dataset (3064 images) for training and internal testing, with the Kaggle dataset (300 images) serving as external validation. Results indicate that YOLO-LS achieved a bounding box mAP50 of 0.953 and a Dice coefficient of 0.910 on the internal test set, with GFLOPs reduced to 8.1, representing a 2.9% precision improvement and a 15.6% reduction in computational load compared to the baseline YOLO11. Ablation experiments and comparisons with models such as U-Net and SegNet confirmed the effectiveness of the improvements; heatmaps further validated the model's precise focus on tumor boundaries. On the external test set, the model demonstrated strong generalization, with an overall Dice coefficient of 0.895. This method achieves an excellent balance among precision, efficiency, and interpretability, showing significant potential for clinical applications and future extensions to multimodal fusion and federated learning.\u003c/p\u003e","manuscriptTitle":"YOLO-LS: A Novel Deep Learning Framework for Brain Tumor Segmentation in Magnetic Resonance Imaging","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-21 01:13:23","doi":"10.21203/rs.3.rs-8027109/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-01-14T06:21:59+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-05T12:55:13+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-12-22T06:00:49+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"232309210707116023637890615009825973046","date":"2025-12-20T07:55:00+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-12-20T06:58:50+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"244688694468544683665538938171846328228","date":"2025-12-09T04:33:44+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"146799412909647911860695728120510078070","date":"2025-12-09T04:01:36+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-11-11T10:21:35+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-11-10T13:51:25+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-11-08T09:19:00+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-11-08T09:17:31+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-11-04T09:19:39+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a31260f3-3574-4383-8bcb-ec367f32bcad","owner":[],"postedDate":"November 21st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":57793086,"name":"Biological sciences/Cancer"},{"id":57793087,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":57793088,"name":"Physical sciences/Engineering"},{"id":57793089,"name":"Physical sciences/Mathematics and computing"},{"id":57793090,"name":"Health sciences/Medical research"}],"tags":[],"updatedAt":"2026-04-10T02:23:13+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-21 01:13:23","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8027109","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8027109","identity":"rs-8027109","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}