Comparison of 2D and 2.5D deep learning features based on multi-parametric magnetic resonance imaging for brain metastases classification

Comparison of 2D and 2.5D deep learning features based on multi-parametric magnetic resonance imaging for brain metastases classification | 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 Research Article Comparison of 2D and 2.5D deep learning features based on multi-parametric magnetic resonance imaging for brain metastases classification Jinling Zhu, Jixuan Deng, Ruizhe Xu, Li Zou, Xin Xie, Ye Tian, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7210303/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Objective This study aims to compare deep learning (DL) features extracted from 2D and 2.5D data via multiparametric magnetic resonance imaging (MRI) to classify brain metastases (BMs) originating from lung cancer (LC), breast cancer (BC), and gastrointestinal cancer (GIC). Methods This retrospective study analyzed MR images from 328 patients with brain metastases (BMs), which were randomly divided into training (N = 229) and test (N = 99) sets at a 7:3 ratio. From the primary lesion slice, we obtained adjacent slices in both the superior-inferior and anterior-posterior directions, constructing a series of two-dimensional (2D) images. DL features were extracted from these slices via pretrained convolutional neural networks (CNNs), including DenseNet121, ResNet50, and ResNet101. A multi-instance learning (MIL) framework was then applied to integrate features into a comprehensive representation. The 2D model, which uses the tumor’s maximal cross-sectional slice as input, followed an identical processing pipeline to that of the 2.5D model. All feature sets were evaluated via machine learning algorithms. Diagnostic performance was assessed via fivefold cross-validation, with accuracy and area under the curve (AUC) metrics quantified for analysis. Results The best classification results were obtained from multiparametric MR images combined with ResNet50. In the test set, the accuracy and AUC of the optimal 2.5D model were 0.906 and 0.961 (95% confidence interval [CI], 0.926–0.978), respectively. The accuracy and AUC of the optimal 2D model were 0.660 and 0.686 (95% CI, 0.622–0.751), respectively. Conclusions On the basis of multiparametric MRI data, the 2.5D-based DL model is feasible for distinguishing the origins of BMs. Brain Metastases 2.5D Deep Learning Multi-Instance Learning Magnetic Resonance Imaging Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Brain metastases (BMs) may be responsible for neurological symptoms in patients with undiagnosed malignancies 1 . Approximately 10–40% of cancer patients develop BMs during the course of the illness 2 . The common sources of their primary tumors, which usually originate from BMs, are lung cancer (LC), breast cancer (BC), melanoma, and gastrointestinal cancer (GIC) 3 . Approximately 15% of patients with BMs have an unknown primary tumor 4 , 5 . Patients with BMs typically present with severe neurological symptoms and have a poor prognosis 6 . Early detection of BMs can help reduce mortality and treatment-related toxicity 7 . However, invasive biopsies impose a significant burden on patients, particularly those in poor physical condition. Performing extensive whole-body imaging or using other definitive identification methods may unnecessarily delay individualized treatment of patients with BMs. This underscores the importance and urgency of developing a noninvasive diagnostic method for BMs. In recent years, deep learning (DL), especially the convolutional neural networks (CNNs) architecture, has achieved remarkable success in the field of medical image processing. DL can automatically extract deeper information from medical images without relying on predefined definitions from human experts. In some medical tasks, DL features show greater potential than hand-crafted radiomic features do 8 – 10 . However, most existing studies focus on extracting features from two-dimensional (2D) or three-dimensional (3D) regions of interest (ROIs). The significant advantage of 3D segmentation over 2D segmentation lies in its ability to utilize 3D spatial information comprehensively. However, 3D segmentation also faces several limitations, such as high computational costs for 3D networks, substantial GPU memory consumption, and numerous parameters that can potentially lead to overfitting 11 . In contrast, the 2.5D method employs the stacking of adjacent slices of the lesion as input rather than the entire lesion volume, making it an effective alternative 12 . In this study, a 2.5D DL classification model was developed utilizing multi-instance learning (MIL) methodology 13 . This approach focuses on processing all 2D magnetic resonance imaging (MRI) slices from the same subject to generate an overall classification score. Unlike traditional supervised learning, MIL operates on training data organized as labeled "bags", where each bag contains multiple unlabeled instances 14 . The goal of MIL is to learn a classifier from labeled positive and negative bags and predict the category of unknown bags, which is suitable for image classification tasks. In our previous research, we validated the practicality of the 2.5D method in differentiating the pathological types of BMs from those of LC 15 . Building upon these findings, the current study aims to further explore the applicability of the 2.5D method in distinguishing among the LC, GIC, and BC of BMs. Materials and methods Patients This single-center retrospective study was approved by the ethics committee of our hospital, and written informed consent was not needed. The MRI and clinical data of patients with BMs were collected in our hospital between December 2010 and April 2023. The inclusion criteria were as follows: (1) BMs confirmed by histopathology or clinical and imaging follow-up; (2) no previous treatment or surgery for BMs; and (3) the pathological type of the primary malignant tumor was determined by pathological examination, and (4) there was only one primary tumor. The exclusion criteria were as follows: (1) metastatic lesion diameter less than 5 mm; (2) images affected by artifacts; and (3) lack of any required sequence of CE-T1WI and FLAIR images. The detailed inclusion and exclusion criteria as well as the participant enrollment process are shown in Fig. 1 . Thus, a total of 328 patients were included in the study. Among all patients, 32 had breast cancer, 37 had gastrointestinal cancer, and 259 had lung cancer. The study randomly divided the samples into a training cohort, comprising 70% (N = 229) of the data, and a testing cohort, containing the remaining 30% (N = 99). The clinical features of the patients with BMs we collected included age, sex, number of tumors, maximum tumor diameter, and maximum diameter of edema. Image Acquisition and Preprocessing All images were acquired from an image archiving and communication system (PACS) and saved in digital imaging and communication in medicine (DICOM) file format for subsequent analysis. RIs were obtained via 1.5 T MRI and 3.0 T MRI. The MRI sequences included FLAIR and CE-T1WI. The acquisition parameters for the different machines are given in Supplementary Table 1. The dataset comprises two imaging sequences: FLAIR and CE-T1WI. The region of interest (ROI) was meticulously delineated via ITK-SNAP (version 3.8.0; http://www.itksnap.org ) by a junior radiologist blinded to the diagnostic and clinical information, followed by review and revision by an experienced senior radiologist. To enhance precision and consistency in medical image analysis, the resolution was uniformly adjusted to 1 mm×1 mm×1 mm, optimizing image quality for subsequent analytical processes 16. Automated rigid alignment was performed to achieve spatial positional alignment of corresponding anatomical structures between the CE-T1WI and FLAIR sequences. All raw images underwent N4 bias field correction to address intensity nonuniformity. Data generation A method was developed to assemble a series of 2D images by extracting adjacent slices along the superior-inferior and anterior-posterior axes relative to a central slice. Hyperparameter optimization revealed that the use of seven-layer 2D images optimally balances lesion information richness with computational efficiency. Adjacent slices at positions ± 1, ±2, and ± 4 relative to the central slice were selected, resulting in seven 2D images per patient. These images, centered on the maximal cross-sectional slice of the ROI, encompass partial three-dimensional structural data, hence termed 2.5D data. Conversely, the single maximum cross-section of the ROI is defined as 2D data. The cropping was performed via the OKT-crop_max_roi tool from the OnekeyAI Platform, with parameters configured to capture extended cross-sectional contexts of the ROI by including slices at + 1, +2, + 4, -1, -2, and − 4. The study design and pipeline are illustrated in Fig. 2. Development of the DL Model This study integrated the generated 2.5D data within a transfer learning framework. The efficacy of several established deep learning architectures—DenseNet121, ResNet50, and ResNet101—was evaluated. These models, pretrained on the ImageNet Large Scale Visual Recognition Challenge 2012 (ILSVRC-2012) dataset, were employed for analysis. To ensure intensity uniformity across the dataset, grayscale values of selected slices underwent min–max normalization, scaled to the range [-1, 1]. Each cropped subregion image was subsequently resized to 224 × 224 pixels via nearest neighbor interpolation to match the input requirements of the chosen models. Given dataset constraints, the learning rate was systematically optimized to enhance model generalizability via a cosine decay strategy. The implemented learning rate parameters were as follows: $$\:{\eta\:}_{t}={\eta\:}_{min}^{i}+\frac{1}{2}\left({\eta\:}_{max}^{i}-{\eta\:}_{min}^{i}\right)\left(1+cos\left(\frac{{T}_{cur}}{{T}_{i}}\pi\:\right)\right)$$ The minimum learning rate, \(\:{\eta\:}_{min}^{i}\) , is set to 0, whereas the maximum learning rate, \(\:{\eta\:}_{max}^{i}\) , is set to 0.01. The parameter \(\:{T}_{i}\) denotes the number of iteration epochs. Other hyperparameters are configured as follows: the optimizer is stochastic gradient descent (SGD), and the loss function used is softmax cross entropy. Details can be found in the Supplementary material. Multi-Instance Learning Fusion This study implemented two fusion techniques embedded within the MIL framework. Under the MIL paradigm, a bag is labeled positive if it contains ≥ 1 positive instance, whereas it is assigned a negative label only when all instances are negative 14 , 17 . Under this assumption, the MIL paradigm is well suited for image classification because images (bags) are generally classified according to some of their subregions (instances) 18 . Details can be found in the Supplementary materials: Predictive likelihood histogram (PLH) : Using 2.5D DL models, we generated histograms representing the distribution of predictive probabilities and labels across each slice in the 2.5D images, encapsulating image features. Bag of Words (BoW) : The complete image was segmented into slices, with probabilities and predictions extracted from each. This process yielded 2 * 7 predictive results per sample (derived from both 2.5D and multimodel analyses). These results were treated analogously to word frequencies within a document, and TF-IDF weighting was applied to characterize the features. Feature Fusion : We combined features from PLH and BoW with radiomic features (based on CE-T1WI and FLAIR) to create a comprehensive feature set. This integrated approach fuses diverse data sources to represent image characteristics efficiently. These features were then used in machine learning algorithms to build models that enhance classification performance. Model Construction and Validation Clinical Signature Univariate analyses were performed on clinical features via the same models applied to the 2.5D DL data to identify features associated with BM classification. A clinical model was subsequently constructed utilizing features that demonstrated statistical significance (p < 0.05). 2.5D and 2D deep learning signatures : The feature selection process is described in the supplementary material. The combined features were input into machine learning algorithms similar to those used in radiomic feature modeling to develop the 2.5DMIL_Rad signature and 2D signature. Within the training set, 5-fold cross-validation was employed alongside grid search for hyperparameter optimization. To assess the effectiveness of multisequence fusion, an identical modeling approach was applied to evaluate performance via exclusively CE-T1WI data and exclusively FLAIR data. Receiver operating characteristic (ROC) curves were plotted, and the area under the curve (AUC), accuracy, sensitivity and specificity were calculated to evaluate the performance of various classification models. Statistical analysis The normality of the clinical features was assessed via the Shapiro‒Wilk test. Continuous variables were evaluated for significance via analysis of variance (ANOVA), whereas categorical variables were analyzed with chi-square (χ²) tests. All the data analyses were conducted via Python 3.7.12 on the OnekeyAI platform v3.1.8. For statistical analyses, we used statsmodels v0.13.2, and for radiomic feature extraction, we employed PyRadiomics v3.0.1. The machine learning algorithms were implemented via scikit-learn 1.0.2. All the DL models were developed via PyTorch 1.11.0, with CUDA 11.3.1 and cuDNN 8.2.1 for hardware acceleration. Results Clinical baseline characteristics Table 1 presents the clinical characteristics of the patients in the two data cohorts. According to the inclusion and exclusion criteria, 328 patients were ultimately included in this study, including 208 males and 120 females, with an age range of 30–85 years and an average age of 63.17 ± 10.54 years. The training group included 229 patients, and the test group included 99 patients. Age, sex, number of tumors, maximum tumor diameter, and maximum edema diameter were not significantly different between the two cohorts (P > 0.05). Table 1 Baseline characteristics of the training and test cohorts. Feature_name ALL train Test Pvalue Age 63.17 ± 10.54 62.88 ± 10.07 63.85 ± 11.57 0.247 Maximun_diameter_of_edema 40.95 ± 28.29 39.02 ± 27.53 45.43 ± 29.63 0.077 Number_of_tumors 4.14 ± 5.29 4.05 ± 5.48 4.33 ± 4.83 0.296 Maximun_tumor_diameter 23.25 ± 11.64 22.92 ± 12.14 24.02 ± 10.40 0.145 Sex 0.353 male 208(63.41) 141(61.57) 67(67.68) Female 120(36.59) 88(38.43) 32(32.32) Performance of the 2.5D and 2D Deep Learning Model The performance of each model on the basis of a single sequence can be seen in Supplementary Table 2 and Table 3 . Among the three DL models, ResNet50 achieved the best overall classification performance. Therefore, we used the feature set obtained from the ResNet50 model to construct RF, SVM, and LR classification models via the scikitlearn machine learning library. Table 2 and Table 3 indicate the performance of ResNet50 on the 2.5D and 2D DL signatures. Figure 3 and Fig. 4 show the AUCs for the 2.5D and 2D models in the cohorts. For the 2.5D DL model, the LR model had a high AUC of 0.994 (95% CI: 0.991–0.998) in the training cohort, indicating excellent discrimination. The AUC was slightly lower at 0.961 (95% CI: 0.938–0.984) in the test cohort but still indicated good generalizability. The performance of the 2.5D ResNet50 model on the basis of a single sequence (CE-T1WI or FLAIR) is presented in Supplementary Table 4. The results demonstrated that the fusion model exhibited superior performance compared with the single-sequence model. In contrast, the predictive performance of the 2D DL model was poor, and the AUCs of the optimal machine learning classifiers were 0.890 and 0.686 in the training and test sets, respectively. Table 2 Performance of the 2.5D ResNet50 model based on comprehensive features (CE-T1WI and FLAIR) in the training and testing sets. Model ACC AUC 95%CI SEN SPN PPV NPV Cohort LR 0.967 0.994 0.991–0.998 0.948 0.976 0.952 0.974 Train LR 0.906 0.961 0.938–0.984 0.859 0.929 0.859 0.929 Test SVM 0.916 0.973 0.962–0.984 0.834 0.956 0.905 0.920 Train SVM 0.923 0.944 0.914–0.975 0.859 0.955 0.904 0.931 Test RF 0.872 0.946 0.927–0.966 0.620 0.998 0.993 0.840 Train RF 0.899 0.952 0.926–0.978 0.717 0.990 0.973 0.875 Test Abbreviations: RF , random forest; SVM , support vector machine; LR , logistic regression; AUC , area under the curve; ACC , accuracy; SEN , sensitivity; SPE , specificity; PPV , positive predictive value; NPV , negative predictive value Table 3 Performance of the 2D ResNet50 model based on comprehensive features (CE-T1WI and FLAIR) in the training and testing sets. Model ACC AUC 95%CI SEN SPN PPV NPV Cohort LR 0.723 0.753 0.715–0.791 0.581 0.795 0.586 0.791 Train LR 0.650 0.645 0.576–0.713 0.515 0.717 0.477 0.747 Test SVM 0.736 0.752 0.714–0.790 0.532 0.838 0.622 0.782 Train SVM 0.683 0.646 0.577–0.715 0.434 0.808 0.531 0.741 Test RF 0.803 0.890 0.866–0.915 0.773 0.819 0.681 0.878 Train RF 0.660 0.686 0.622–0.751 0.525 0.727 0.491 0.754 Test Abbreviations: RF , random forest; SVM , support vector machine; LR , logistic regression; AUC , area under the curve; ACC , accuracy; SEN , sensitivity; SPE , specificity; PPV , positive predictive value; NPV , negative predictive value Grad-CAM To explore the recognition capabilities of DL models on diverse samples, we utilized the gradient-weighted class activation mapping (Grad-CAM) technique for visualization 19 . Figure 5 demonstrates the application of Grad-CAM, which shows the activations in the final convolutional layer that are pertinent to predicting BM types. This visualization highlights image regions that substantially influence the model's decision-making process, thereby enhancing our understanding of its interpretability. Discussion This study compared DL features derived from 2D versus 2.5D data within multiparametric MRI for classifying BMs from lung, breast, and gastrointestinal cancers. Our approach employed seven adjacent slices acquired from both the superior-inferior and anterior-posterior directions. DL features were extracted from each slice via three pretrained CNNs. A multiple instance learning (MIL) framework was then applied, treating each patient’s collection of MRI slices as a "bag" and individual slices as "instances." The aggregated features from each bag served as input to machine learning classifiers. The aggregated features from each bag served as input to machine learning classifiers. This strategy synergizes the powerful feature extraction capabilities of DL, the contextual integration strengths of MIL, and the stability of traditional machine learning, enabling more accurate and reliable BM classification. In the test set, the optimal 2.5D classifier achieved an accuracy of 0.906 and an AUC of 0.961, outperforming the optimal 2D classifier (accuracy: 0.660; AUC: 0.686). This demonstrates the superior classification performance of the 2.5D DL model, indicating that 2.5D data provide greater spatial contextual information. Thus, we conclude that 2.5D-based DL models enhance the differentiation of BM types, facilitating the identification of primary tumor sites. DL is an essential branch of machine learning and has demonstrated remarkable potential in medical image analysis through its multilayer artificial neural network architecture, which simulates the human cognitive system 20 . In recent years, this technology has been extensively investigated and clinically validated in the field of brain tumor classification research. Grossman et al 21 . employed three ImageNet-pretrained CNNs (VGG19, Xception, InceptionV3) to differentiate breast cancer (BC) and lung cancer (LC) BMs via CE-T1WI. Their 3D CNN models achieved high diagnostic efficacy (accuracy: 0.82–0.85). Similarly, in a subsequent study, Gultekin et al 22 . used three pretrained CNNs and texture-based 2D and 3D tumor image features to distinguish LC and BC BMs. The study results showed that the overall performance of the CNN architectures using 3D ROIs as inputs was greater than that of the architectures based on texture features, with the 3D Xception architecture achieving good performance (accuracy: 0.85; AUC: 0.84). In broader tumor classification research, Kumar et al. 23 implemented multiple architectures (ResNet, AlexNet, U-Net, VGG-16) for MRI-based brain tumor categorization. Their modified ResNet50 achieved exceptional accuracy (benign: 0.993; malignant: 0.984). The variation in performance among the different models can be attributed to disparities in the network's internal architecture 24 . To select the best CNNs for diagnosing BMs, this study developed three commonly used diagnostic models based on different neural network architectures, including ResNet50, ResNet101, and DenseNet121. The pretrained ResNet50 model performed well in the BM classification task, achieving an average accuracy of 0.899. ResNet-50 uses residual learning to reduce gradient dispersion and precision loss in deep networks, which can improve model accuracy while accelerating the training of neural networks. Another notable feature of ResNet-50 is the use of a global average pooling layer, which takes the average value of all the pixels in each feature map as the output of that feature map, which reduces the number of model parameters and lowers the risk of overfitting 25 . In the field of image segmentation via machine learning models, 2D and 3D approaches represent the two primary methodologies. While 3D CNNs capture richer contextual information and generally outperform 2D CNNs do, they require significantly more computational resources, including GPU memory and processing time 12 . To address the limitations of both 2D and 3D CNNs, the 2.5D segmentation method has been proposed and recommended in numerous studies 26 , 27 . This approach combines the computational efficiency of 2D CNNs with the enhanced spatial context capture capabilities inherent to 3D CNNs 12 . Recent research has demonstrated the excellent performance of 2.5D segmentation methods across various tasks 28 – 30 . For example, Takao et al. 29 developed a 2.5D DL model for the automatic detection of BMs, which uses three consecutive slices as inputs to predict the central slice. Their comparative analysis revealed superior overall performance relative to a conventional 2D model using a single slice. Similarly, Xiong et al. 30 . investigated the value of DL models in distinguishing bone islands from osteoblastic bone metastases. Their findings indicated that a 2.5D DL model employing three-layer CT images as input outperformed 2D models in classifying sclerotic bone lesions. This model achieved high performance both internally (AUC, 0.996) and on two external validation sets (AUC, 0.958; AUC, 0.952). These results are consistent with the findings of the current study. Herein, the largest cross-section of the ROI was designated the primary slice, followed by the extraction of 2D images from multiple adjacent slices to form a 2.5D dataset incorporating partial 3D structural information. The results demonstrate effective performance in the differential diagnosis of three BM types via the 2.5D DL model, which also surpassed the performance of the 2D model. However, future research should aim to develop a comprehensive 3D model based on multisequence MRI to rigorously compare the performance differences and efficacy among various segmentation models for BM classification tasks. To validate the effectiveness of the proposed 2.5D DL model, we compared it with relevant studies in recent years. The analysis in Table 4 reveals that most studies have constructed classification models using radiomic features, whereas only Gultekin et al. 22 , Jiao et al. 31 , and Lyu et al. 32 have explored the classification performance of 2D or 3D DL models. Notably, Lyu et al. 32 established an end-to-end DL model using a large dataset, which improved the model's generalization ability and stability by optimizing the training process and reducing manual intervention. The work of Jiao et al. 31 is also noteworthy. By using the ResNet18 model to classify breast cancer and gastrointestinal BMs and validating the model with datasets from multiple centers, they enhanced the model's generalization ability, achieving an AUC value of 0.848. However, the accuracy performance was moderate (ACC = 0.700). Several radiomic studies reported in the literature have successfully linked MR/CT image features to the classification of BMs 33 – 38 . However, these studies have focused mainly on binary classification. In contrast, our method outperforms traditional radiomic methods in both Ortiz-Ramón's 36 three-class classification (accuracy of 0.906 vs 0.873) and Kniep's 35 five-class classification (accuracy of 0.906 vs 0.820). Compared with radiomics, our 2.5D DL model has several advantages in distinguishing BMs. First, the ROIs selection method based on 2.5D DL only needs to encompass the entire lesion instead of precisely delineating the lesion edge layer by layer, as in radiomics. Second, our study recruited a greater number of patients. A larger sample size can guarantee the reliability of a better classification model. Nevertheless, this study and most other studies have a single-center retrospective design. The generalizability of the model remains to be verified, which is the direction that future research needs to address. Table 4 Summary of the application of machine learning methods in the classification of brain metastases in recent years. Paper Single/Multi-Center Sample Size Primary cancer Optimal classifier AUC ACC Our study Single-Center 328 LC vs BC vs GIC 2.5D ResNet50 0.952 0.899 Gultekin et al 2023[22] Single-Center 143 LC vs BC 2D Xception 3D Xception 0.800 0.850 0.830 0.845 Jiao et al [31]2023 Multi-Center 214 BC vs GIC 3D ResNet 18 0.848 0.700 Lyu et al [32]2022 Single-Center 1582 LC vs BC vs Melanoma vs Renal vs Other 3D cycle-GAN 0.878 - Shi et al [33] 2023 Single-Center 160 LC vs BC 3D LR 0.778 0.843 0.717 0.792 Cao et al [34]2022 Single-Center 78 LC vs BC SVM 0.805 - Kniep et al [35]2019 Single-Center 189 SCLC vs BC vs melanoma vs GIC vs NSCLC 3D RF 0.64–0.82 - Ortiz-Ramón et al [36]2018 Single-Center 38 LC vs melanoma vs BC 3D RF 0.873 ± 0.064 - Ortiz-Ramón et al [37]2017 Single-Center 29 LC vs BC 2D SVM 0.953 ± 0.061 - Ortiz-Ramón et al [38]2017 Single-Center 30 LC vs Melanoma 2D KNN 3DNB 0.890 ± 0.085 0.947 ± 0.067 - - Abbreviations: RF , random forest; AUC , area under the curve; ACC , accuracy; LC , lung cancer; BC , breast cancer; GIC , gastrointestinal cancer; SVM , support vector machine; k-NN , k-nearest neighbors; LR , logistic regression; NB , naïve Bayes; NSCLC , non-small cell lung cancer; SCLC , small cell lung cancer This study has the following limitations. First, the study employed a retrospective, single-center design and lacked external data to validate the model's efficacy. Second, we only considered metastases from three primary sites, necessitating further research to collect data on other types of metastases. Additionally, although LC are widely regarded as the main source of BMs, the significant discrepancies in sample sizes among various categories in this study pose a problem that cannot be overlooked 39 . This unbalanced sample distribution is likely to impose restrictions on the model's capacity for identifying and classifying samples of minority classes. To address this challenge, future studies should consider employing specific data processing techniques, such as the synthetic minority oversampling technique (SMOTE), to handle the issue of unbalanced sample sizes more effectively 40 . Conclusion In summary, the MRI-based 2.5D deep learning model can serve as a sensitive and specific diagnostic tool for distinguishing between different pathological types of BMs. This innovative method not only helps improve the diagnostic level of BMs but also provides an important basis for subsequent personalized treatment and prognosis evaluation. Abbreviations BMs brain metastases CE-T1WI contrast-enhanced T1-weighted imaging DL deep learning 2D two-dimensional 3D three-dimensional FLAIR fluid-attenuated inversion recovery images MIL multi-instance learning RF random forest SVM support vector machine LR logistic regression LC Lung cancer BC breast cancer GIC gastrointestinal cancer AUC area under the curve 95% CI 95% confidence interval Grad-CAM gradient-weighted class activation mapping CNNs Convolutional neural networks Declarations Ethics approval and consent to participate This study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of The Second Affiliated Hospital of Soochow University. The ethics committee waived the requirement for informed consent due to the retrospective nature of the study using anonymized data. Consent for publication Not applicable. Data availability The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request. Competing interests The authors declare no competing interests. Funding This work was supported by Suzhou Science and Technology Development Plan Project in China [grant numbers SLJ2022009]; Suzhou Science and Technology Development Plan Project (Medical and Health Technology Innovation) in China [grant numbers SYSD2022112]; Exploratory Scientific Research Grant Project of the Second Affiliated Hospital of Soochow University in China [grant numbers SDFEYBS2425]. Authors' contributions Bo Zhang. conceived and designed the research framework. Jinling Zhu. served as the major contributor to manuscript drafting.Jixuan Deng. performed data acquisition and curation. Ruizhe Xu. and Li Zou. developed the deep learning models and performed the statistical analyses. Xin Xie.,Ye Tian., and Wu Cai. offered technical supervision and methodological refinement during model construction. All authors reviewed the manuscript. Acknowledgements The authors sincerely thank Platform Onekey AI for Python technology of the study. References Balestrino R, Rudà R, Soffietti R. Brain metastasis from unknown primary tumour: moving from old retrospective studies to clinical trials on targeted agents. Cancers (Basel) 2020;12 Lamba N, Wen PY, Aizer AA. Epidemiology of brain metastases and leptomeningeal disease. Neuro Oncol 2021;23:1447-1456 Suh JH, Kotecha R, Chao ST, et al. Current approaches to the management of brain metastases. Nat Rev Clin Oncol 2020;17:279-299 Rassy E, Zanaty M, Azoury F, et al. Advances in the management of brain metastases from cancer of unknown primary. Future Oncol 2019;15:2759-2768 Wolpert F, Weller M, Berghoff AS, et al. Diagnostic value of (18)F-fluordesoxyglucose positron emission tomography for patients with brain metastasis from unknown primary site. Eur J Cancer 2018;96:64-72 DeVries DA, Lagerwaard F, Zindler J, et al. Performance sensitivity analysis of brain metastasis stereotactic radiosurgery outcome prediction using MRI radiomics. Sci Rep 2022;12:20975 Cagney DN, Martin AM, Catalano PJ, et al. Incidence and prognosis of patients with brain metastases at diagnosis of systemic malignancy: a population-based study. Neuro Oncol 2017;19:1511-1521 Bae S, An C, Ahn SS, et al. Robust performance of deep learning for distinguishing glioblastoma from single brain metastasis using radiomic features: model development and validation. Sci Rep 2020;10:12110 Zhang H, Zhang H, Zhang Y, et al. Deep learning radiomics for the assessment of telomerase reverse transcriptase promoter mutation status in patients with glioblastoma using multiparametric MRI. J Magn Reson Imaging 2023;58:1441-1451 Tulum G. Novel radiomic features versus deep learning: differentiating brain metastases from pathological lung cancer types in small datasets. Br J Radiol 2023;96:20220841 Tian Y, Xue F, Lambo R, et al. Fully automated functional region annotation of liver via a 2.5D class-aware deep neural network with spatial adaptation. Comput Methods Programs Biomed 2021;200:105818 Zhang Y, Liao Q, Ding L, et al. Bridging 2D and 3D segmentation networks for computation-efficient volumetric medical image segmentation: An empirical study of 2.5D solutions. Comput Med Imaging Graph 2022;99:102088 Yin S, Peng Q, Li H, et al. Multi-instance deep learning of ultrasound imaging data for pattern classification of congenital abnormalities of the kidney and urinary tract in children. Urology 2020;142 Xiao Y, Liang F, Liu B. A Transfer Learning-Based Multi-Instance Learning Method With Weak Labels. IEEE Trans Cybern 2022;52:287-300 Zhu J, Zou L, Xie X, et al. 2.5D deep learning based on multiparameter MRI to differentiate primary lung cancer pathological subtypes in patients with brain metastases. Eur J Radiol 2024;180:111712 Moradmand H, Aghamiri SMR, Ghaderi R. Impact of image preprocessing methods on reproducibility of radiomic features in multimodal magnetic resonance imaging in glioblastoma. J Appl Clin Med Phys 2020;21:179-190 Wang J, Cai L, Peng J, et al. A novel multiple instance learning method based on extreme learning machine. Comput Intell Neurosci 2015;2015:405890 Astorino A, Fuduli A, Veltri P, et al. Melanoma detection by means of multiple instance learning. Interdiscip Sci 2020;12:24-31 Zhang H, Ogasawara K. Grad-CAM-Based Explainable Artificial Intelligence Related to Medical Text Processing. Bioengineering (Basel) 2023;10 Lee JG, Jun S, Cho YW, et al. Deep Learning in Medical Imaging: General Overview. Korean J Radiol 2017;18:570-584 Grossman R, Haim O, Abramov S, et al. Differentiating small-cell lung cancer from non-small cell lung cancer brain metastases based on MRI using efficientnet and transfer learning approach. Technol Cancer Res Treat 2021;20:15330338211004919 Gultekin MA, Peker AA, Oktay AB, et al. Differentiation of lung and breast cancer brain metastases: Comparison of texture analysis and deep convolutional neural networks. J Clin Ultrasound 2023;51:1579-1586 Kumar S, Choudhary S, Jain A, et al. Brain tumor classification using deep neural network and transfer learning. Brain Topogr 2023;36:305-318 Yu Q, Ning Y, Wang A, et al. Deep learning-assisted diagnosis of benign and malignant parotid tumors based on contrast-enhanced CT: a multicenter study. Eur Radiol 2023;33:6054-6065 He K, Zhang X, Ren S, et al. Deep residual learning for image recognition. IEEE 2016 Mzoughi H, Njeh I, Wali A, et al. Deep multi-scale 3D convolutional neural network (cnn) for mri gliomas brain tumor classification. J Digit Imaging 2020;33:903-915 Avesta A, Hossain S, Lin M, et al. Comparing 3D, 2.5D, and 2D approaches to brain image autosegmentation. Bioengineering (Basel) 2023;10 Huang L, Zhao Z, An L, et al. 2.5D transfer deep learning model for segmentation of contrast-enhancing lesions on brain magnetic resonance imaging of multiple sclerosis and neuromyelitis optica spectrum disorder. Quant Imaging Med Surg 2024;14:273-290 Takao H, Amemiya S, Kato S, et al. Deep-learning 2.5-dimensional single-shot detector improves the performance of automated detection of brain metastases on contrast-enhanced CT. Neuroradiology 2022;64:1511-1518 Xiong Y, Guo W, Liang Z, et al. Deep learning-based diagnosis of osteoblastic bone metastases and bone islands in computed tomograph images: a multicenter diagnostic study. Eur Radiol 2023;33:6359-6368 Jiao T, Li F, Cui Y, et al. Deep learning with an attention mechanism for differentiating the origin of brain metastasis using MR images. J Magn Reson Imaging 2023;58:1624-1635 Lyu Q, Namjoshi SV, McTyre E, et al. A transformer-based deep-learning approach for classifying brain metastases into primary organ sites using clinical whole-brain MRI images. Patterns (N Y) 2022;3:100613 Shi J, Chen H, Wang X, et al. Using Radiomics to Differentiate brain metastases from lung cancer versus breast cancer, including predicting epidermal growth factor receptor and human epidermal growth factor receptor 2 status. J Comput Assist Tomogr 2023;47:924-933 Cao G, Zhang J, Lei X, et al. Differentiating primary tumors for brain metastasis with integrated radiomics from multiple imaging modalities. Dis Markers 2022;2022:5147085 Kniep HC, Madesta F, Schneider T, et al. Radiomics of brain MRI: utility in prediction of metastatic tumor type. Radiology 2019;290:479-487 Ortiz-Ramón R, Larroza A, Ruiz-España S, et al. Classifying brain metastases by their primary site of origin using a radiomics approach based on texture analysis: a feasibility study. Eur Radiol 2018;28:4514-4523 Ortiz-Ramon R, Larroza A, Arana E, et al. Identifying the primary site of origin of MRI brain metastases from lung and breast cancer following a 2D radiomics approach. 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017); 2017 Ortiz-Ramon R, Larroza A, Arana E, et al. A radiomics evaluation of 2D and 3D MRI texture features to classify brain metastases from lung cancer and melanoma. Annu Int Conf IEEE Eng Med Biol Soc 2017;2017:493-496 Shi W, Tanzhu G, Chen L, et al. Radiotherapy in preclinical models of brain metastases: a review and recommendations for future studies. Int J Biol Sci 2024;20:765-783 Dablain D, Krawczyk B, Chawla NV. DeepSMOTE: fusing deep learning and smote for imbalanced data. IEEE Trans Neural Netw Learn Syst 2023;34:6390-6404 Additional Declarations No competing interests reported. Supplementary Files Supplementary.docx Cite Share Download PDF Status: Posted Version 1 posted 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-7210303","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":518243257,"identity":"c0ea2f12-3265-4bb1-9f39-6930e514670a","order_by":0,"name":"Jinling Zhu","email":"","orcid":"","institution":"Second Affiliated Hospital of Soochow University Department of Radiology","correspondingAuthor":false,"prefix":"","firstName":"Jinling","middleName":"","lastName":"Zhu","suffix":""},{"id":518243258,"identity":"5d7a2118-3e36-48db-956b-53ccea8fecf8","order_by":1,"name":"Jixuan Deng","email":"","orcid":"","institution":"Second Affiliated Hospital of Soochow University Department of Radiology","correspondingAuthor":false,"prefix":"","firstName":"Jixuan","middleName":"","lastName":"Deng","suffix":""},{"id":518243259,"identity":"203ef863-eaf6-418a-8514-5afc15a45899","order_by":2,"name":"Ruizhe Xu","email":"","orcid":"","institution":"Second Affiliated Hospital of Soochow University Department of Radiotherapy \u0026 Oncology","correspondingAuthor":false,"prefix":"","firstName":"Ruizhe","middleName":"","lastName":"Xu","suffix":""},{"id":518243260,"identity":"55105f01-4587-4757-a920-1f62503d3f37","order_by":3,"name":"Li Zou","email":"","orcid":"","institution":"Second Affiliated Hospital of Soochow University Department of Radiotherapy \u0026 Oncology","correspondingAuthor":false,"prefix":"","firstName":"Li","middleName":"","lastName":"Zou","suffix":""},{"id":518243261,"identity":"b61fe7b8-b955-4759-bedd-acc70ee424f8","order_by":4,"name":"Xin Xie","email":"","orcid":"","institution":"Second Affiliated Hospital of Soochow University Department of Radiology","correspondingAuthor":false,"prefix":"","firstName":"Xin","middleName":"","lastName":"Xie","suffix":""},{"id":518243263,"identity":"bc7dc3db-8517-4fa3-a832-a575dfa90b41","order_by":5,"name":"Ye Tian","email":"","orcid":"","institution":"Second Affiliated Hospital of Soochow University Department of Radiotherapy \u0026 Oncology","correspondingAuthor":false,"prefix":"","firstName":"Ye","middleName":"","lastName":"Tian","suffix":""},{"id":518243264,"identity":"389aa4d2-ad4e-4b1c-91fa-a1b55e4d850d","order_by":6,"name":"Wu Cai","email":"","orcid":"","institution":"Second Affiliated Hospital of Soochow University Department of Radiology","correspondingAuthor":false,"prefix":"","firstName":"Wu","middleName":"","lastName":"Cai","suffix":""},{"id":518243265,"identity":"8491cf45-9793-4247-805e-adbdc1dc8248","order_by":7,"name":"Bo Zhang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA50lEQVRIiWNgGAWjYDACCSBmbGBgYGNmPvjgg4GNHPFa+NjZkg1nFKQZE69Fjp/HTJrnw+FEgjrkZzc/e/h1hw3QYQwG0jYGzAkM7IePbsCnhXHOMXNj2TNpIC0JxjkGbHkMPGlpN/BpYZZIMJOWbDsM0nIgOceAp5hBgscMrxY2ifRvUC2MDYctDCQSGwhp4ZHIMZP8CNbCzNjMYGBAWIuERE6ZNCPYL0B7egwSjNkI+UV+Rvo2yZ/AEJPvP//9x48//+X42Q8fw6sFBJh5GBjqG+C+I6QcBBh/EKNqFIyCUTAKRi4AAIZ7P6KOCL9ZAAAAAElFTkSuQmCC","orcid":"","institution":"Second Affiliated Hospital of Soochow University Department of Radiology","correspondingAuthor":true,"prefix":"","firstName":"Bo","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2025-07-25 04:38:36","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7210303/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7210303/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":91988053,"identity":"0451109b-cf77-4d44-87af-a3098e5a5a61","added_by":"auto","created_at":"2025-09-23 12:12:50","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":117123,"visible":true,"origin":"","legend":"","description":"","filename":"MainDocument.docx","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/776f239a7d793fcfbec69df1.docx"},{"id":91987974,"identity":"6bc156f7-7a9b-423f-8f56-2688f5e07c98","added_by":"auto","created_at":"2025-09-23 12:12:45","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28704,"visible":true,"origin":"","legend":"","description":"","filename":"Table.docx","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/8bdabb3e510a92bcc2c7a52e.docx"},{"id":91988068,"identity":"78f46b35-a186-4385-9a33-b6d0a3382c87","added_by":"auto","created_at":"2025-09-23 12:12:51","extension":"tif","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11685360,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.tif","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/9d6781b1b4520bbbff0abf0f.tif"},{"id":91987878,"identity":"78e5f8ed-ea5c-4569-a008-caff45af36da","added_by":"auto","created_at":"2025-09-23 12:12:27","extension":"tiff","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1958876,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.tiff","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/229528bc7d65834e404a7cad.tiff"},{"id":91987918,"identity":"efea50d0-c512-47b4-b073-5bed1c918e9d","added_by":"auto","created_at":"2025-09-23 12:12:35","extension":"tif","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":25908948,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.tif","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/fe88ce22b19c2258e9b5a711.tif"},{"id":91987939,"identity":"02a3c1c4-b2bf-4023-9e2d-da4dda55b16f","added_by":"auto","created_at":"2025-09-23 12:12:41","extension":"tif","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15209300,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.tif","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/c1880fda6a7cd8330e0f818b.tif"},{"id":91988011,"identity":"1a79aeb0-00b6-4f41-917e-8dde83f0c469","added_by":"auto","created_at":"2025-09-23 12:12:47","extension":"tif","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16452996,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.tif","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/2592d7cb4cc4a6b0a67ebd82.tif"},{"id":91987899,"identity":"112d9bca-0309-4907-9fa8-b5872a81bd2c","added_by":"auto","created_at":"2025-09-23 12:12:30","extension":"json","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9250,"visible":true,"origin":"","legend":"","description":"","filename":"4f51ea2ddd404c3eb596f3b58d7558d5.json","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/a07daa94dd5b6fd9e5d86cae.json"},{"id":91988050,"identity":"0a99634d-c8d9-4149-bd2d-5f80c05fec80","added_by":"auto","created_at":"2025-09-23 12:12:50","extension":"docx","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":36075,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/20c2a6fd5f2f2a3ddfe99a8a.docx"},{"id":91987916,"identity":"02e814f9-b1c7-4f5e-a20e-8d92c749afb7","added_by":"auto","created_at":"2025-09-23 12:12:35","extension":"xml","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":116753,"visible":true,"origin":"","legend":"","description":"","filename":"4f51ea2ddd404c3eb596f3b58d7558d51enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/1b6be54cf299f0fd06c4b66a.xml"},{"id":91987970,"identity":"2207effa-7e61-4fdd-ac9d-52b59fb77efe","added_by":"auto","created_at":"2025-09-23 12:12:45","extension":"tif","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11685360,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.tif","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/60870c27019e66a633d61ad0.tif"},{"id":91987936,"identity":"28182a32-fead-44c2-96ad-78ecb65808cf","added_by":"auto","created_at":"2025-09-23 12:12:40","extension":"tiff","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1958876,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.tiff","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/9f7ff4a415a6a6619dd81bfa.tiff"},{"id":91987903,"identity":"dae61478-a53f-42fb-821b-94b4c627b151","added_by":"auto","created_at":"2025-09-23 12:12:31","extension":"tif","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":25908948,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.tif","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/c68829cd5716d653ee6fdeaf.tif"},{"id":91988352,"identity":"51733ef5-cb69-4f0e-900f-c16d74856543","added_by":"auto","created_at":"2025-09-23 12:20:51","extension":"tif","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15209300,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.tif","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/bf008c09b2ead2a9b9f1077b.tif"},{"id":91987934,"identity":"3cd516d2-c571-4e02-b0d7-86e58de53f5c","added_by":"auto","created_at":"2025-09-23 12:12:39","extension":"tif","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16452996,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.tif","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/13cd671328f116a3c2b2f186.tif"},{"id":91987873,"identity":"70c7fb9a-4df4-4cba-ba63-fd2990e824f9","added_by":"auto","created_at":"2025-09-23 12:12:23","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1061726,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/6d33107bd4ecef559a29d580.png"},{"id":91987943,"identity":"f16c7069-a8bf-4436-9e0c-3619012e78c1","added_by":"auto","created_at":"2025-09-23 12:12:41","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":245407,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/f9d258b72d357fa81b3d84a1.png"},{"id":91987949,"identity":"83acda79-3bb6-438b-8cb8-9592ceffc64f","added_by":"auto","created_at":"2025-09-23 12:12:42","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":152765,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/1ab3f7f8f087e41a8c3dc023.png"},{"id":91987905,"identity":"4980c933-f8a8-41d5-968e-636fe9145e53","added_by":"auto","created_at":"2025-09-23 12:12:32","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":139106,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/2625f918a111980dc7b9767b.png"},{"id":91988042,"identity":"f2f067f0-22e9-444a-99a9-a98dc054cae2","added_by":"auto","created_at":"2025-09-23 12:12:48","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3893105,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/1651fc5ff65bab53835690c3.png"},{"id":91987928,"identity":"60a93968-4e48-42c6-af7f-904394590213","added_by":"auto","created_at":"2025-09-23 12:12:38","extension":"xml","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":113729,"visible":true,"origin":"","legend":"","description":"","filename":"4f51ea2ddd404c3eb596f3b58d7558d51structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/70cd0095ac8eec681b54aafe.xml"},{"id":91988073,"identity":"b9e7609a-9335-4a9f-aee7-3955a2299c0b","added_by":"auto","created_at":"2025-09-23 12:12:51","extension":"html","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":125429,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/3ec1031a3f12655ac6053788.html"},{"id":91988346,"identity":"5a32d797-ebdd-45d5-a20e-980a71b5a450","added_by":"auto","created_at":"2025-09-23 12:20:41","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":845701,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of study population with inclusion and exclusion criteria.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/f67c6e76353fb6fae529ded0.png"},{"id":91988066,"identity":"80bd093f-cc68-4367-a863-b75dc1228457","added_by":"auto","created_at":"2025-09-23 12:12:51","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":3689330,"visible":true,"origin":"","legend":"\u003cp\u003eComprehensive workflow of the 2.5D deep learning model.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/21fe972abbb327dd392fba25.png"},{"id":91987911,"identity":"0efc61cf-0c40-4ba6-9765-98f68318fd2c","added_by":"auto","created_at":"2025-09-23 12:12:35","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1757888,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves of 2.5D ResNet50 in training and testing sets. Class 0: gastrointestinal cancer; Class 1: breast cancer; Class 2: lung cancer.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/9e4d6c65a55735981d0d25b3.png"},{"id":91987924,"identity":"65fda6e0-8cd7-4e42-967d-5fae447b59ed","added_by":"auto","created_at":"2025-09-23 12:12:36","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1412087,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves of 2D ResNet50 in training and testing sets. Class 0: gastrointestinal cancer; Class 1: breast cancer; Class 2: lung cancer.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/7c314104139ed98b189d6537.png"},{"id":91988079,"identity":"f919ac76-25ff-4a02-86e7-b3fe54959c73","added_by":"auto","created_at":"2025-09-23 12:12:51","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":4876126,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization of a representative patient case. Red regions indicate areas of higher predictive weight.\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/18d20a4214e2a255e75f8367.png"},{"id":103481836,"identity":"08539430-9c21-4066-929e-97237cd47c89","added_by":"auto","created_at":"2026-02-26 08:12:38","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":12914349,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/56620f1a-393d-4c1f-bce6-9d2c3309df46.pdf"},{"id":91987895,"identity":"fa6a4ae4-7239-497c-8f9f-3d7cf9f5d640","added_by":"auto","created_at":"2025-09-23 12:12:29","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":36075,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-7210303/v1/e136872d5a9947b9346c72aa.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eComparison of 2D and 2.5D deep learning features based on multi-parametric magnetic resonance imaging for brain metastases classification\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eBrain metastases (BMs) may be responsible for neurological symptoms in patients with undiagnosed malignancies\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Approximately 10\u0026ndash;40% of cancer patients develop BMs during the course of the illness\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. The common sources of their primary tumors, which usually originate from BMs, are lung cancer (LC), breast cancer (BC), melanoma, and gastrointestinal cancer (GIC)\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Approximately 15% of patients with BMs have an unknown primary tumor \u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Patients with BMs typically present with severe neurological symptoms and have a poor prognosis \u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Early detection of BMs can help reduce mortality and treatment-related toxicity \u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. However, invasive biopsies impose a significant burden on patients, particularly those in poor physical condition. Performing extensive whole-body imaging or using other definitive identification methods may unnecessarily delay individualized treatment of patients with BMs. This underscores the importance and urgency of developing a noninvasive diagnostic method for BMs.\u003c/p\u003e\u003cp\u003eIn recent years, deep learning (DL), especially the convolutional neural networks (CNNs) architecture, has achieved remarkable success in the field of medical image processing. DL can automatically extract deeper information from medical images without relying on predefined definitions from human experts. In some medical tasks, DL features show greater potential than hand-crafted radiomic features do \u003csup\u003e\u003cspan additionalcitationids=\"CR9\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. However, most existing studies focus on extracting features from two-dimensional (2D) or three-dimensional (3D) regions of interest (ROIs). The significant advantage of 3D segmentation over 2D segmentation lies in its ability to utilize 3D spatial information comprehensively. However, 3D segmentation also faces several limitations, such as high computational costs for 3D networks, substantial GPU memory consumption, and numerous parameters that can potentially lead to overfitting \u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. In contrast, the 2.5D method employs the stacking of adjacent slices of the lesion as input rather than the entire lesion volume, making it an effective alternative \u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. In this study, a 2.5D DL classification model was developed utilizing multi-instance learning (MIL) methodology \u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. This approach focuses on processing all 2D magnetic resonance imaging (MRI) slices from the same subject to generate an overall classification score. Unlike traditional supervised learning, MIL operates on training data organized as labeled \"bags\", where each bag contains multiple unlabeled instances \u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. The goal of MIL is to learn a classifier from labeled positive and negative bags and predict the category of unknown bags, which is suitable for image classification tasks. In our previous research, we validated the practicality of the 2.5D method in differentiating the pathological types of BMs from those of LC \u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. Building upon these findings, the current study aims to further explore the applicability of the 2.5D method in distinguishing among the LC, GIC, and BC of BMs.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cp\u003e\u003cb\u003ePatients\u003c/b\u003e\u003c/p\u003e\u003cp\u003e This single-center retrospective study was approved by the ethics committee of our hospital, and written informed consent was not needed. The MRI and clinical data of patients with BMs were collected in our hospital between December 2010 and April 2023. The inclusion criteria were as follows: (1) BMs confirmed by histopathology or clinical and imaging follow-up; (2) no previous treatment or surgery for BMs; and (3) the pathological type of the primary malignant tumor was determined by pathological examination, and (4) there was only one primary tumor. The exclusion criteria were as follows: (1) metastatic lesion diameter less than 5 mm; (2) images affected by artifacts; and (3) lack of any required sequence of CE-T1WI and FLAIR images. The detailed inclusion and exclusion criteria as well as the participant enrollment process are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Thus, a total of 328 patients were included in the study. Among all patients, 32 had breast cancer, 37 had gastrointestinal cancer, and 259 had lung cancer. The study randomly divided the samples into a training cohort, comprising 70% (N\u0026thinsp;=\u0026thinsp;229) of the data, and a testing cohort, containing the remaining 30% (N\u0026thinsp;=\u0026thinsp;99). The clinical features of the patients with BMs we collected included age, sex, number of tumors, maximum tumor diameter, and maximum diameter of edema.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eImage Acquisition and Preprocessing\u003c/b\u003e\u003c/p\u003e\u003cp\u003eAll images were acquired from an image archiving and communication system (PACS) and saved in digital imaging and communication in medicine (DICOM) file format for subsequent analysis. RIs were obtained via 1.5 T MRI and 3.0 T MRI. The MRI sequences included FLAIR and CE-T1WI. The acquisition parameters for the different machines are given in Supplementary Table\u0026nbsp;1.\u003c/p\u003e\u003cp\u003eThe dataset comprises two imaging sequences: FLAIR and CE-T1WI. The region of interest (ROI) was meticulously delineated via ITK-SNAP (version 3.8.0; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.itksnap.org\u003c/span\u003e\u003cspan address=\"http://www.itksnap.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) by a junior radiologist blinded to the diagnostic and clinical information, followed by review and revision by an experienced senior radiologist. To enhance precision and consistency in medical image analysis, the resolution was uniformly adjusted to 1 mm\u0026times;1 mm\u0026times;1 mm, optimizing image quality for subsequent analytical processes 16. Automated rigid alignment was performed to achieve spatial positional alignment of corresponding anatomical structures between the CE-T1WI and FLAIR sequences. All raw images underwent N4 bias field correction to address intensity nonuniformity.\u003c/p\u003e\u003cp\u003e\u003cb\u003eData generation\u003c/b\u003e\u003c/p\u003e\u003cp\u003eA method was developed to assemble a series of 2D images by extracting adjacent slices along the superior-inferior and anterior-posterior axes relative to a central slice. Hyperparameter optimization revealed that the use of seven-layer 2D images optimally balances lesion information richness with computational efficiency. Adjacent slices at positions\u0026thinsp;\u0026plusmn;\u0026thinsp;1, \u0026plusmn;2, and \u0026plusmn;\u0026thinsp;4 relative to the central slice were selected, resulting in seven 2D images per patient. These images, centered on the maximal cross-sectional slice of the ROI, encompass partial three-dimensional structural data, hence termed 2.5D data. Conversely, the single maximum cross-section of the ROI is defined as 2D data. The cropping was performed via the OKT-crop_max_roi tool from the OnekeyAI Platform, with parameters configured to capture extended cross-sectional contexts of the ROI by including slices at +\u0026thinsp;1, +2, +\u0026thinsp;4, -1, -2, and \u0026minus;\u0026thinsp;4. The study design and pipeline are illustrated in Fig.\u0026nbsp;2.\u003c/p\u003e\u003cp\u003e\u003cb\u003eDevelopment of the DL Model\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThis study integrated the generated 2.5D data within a transfer learning framework. The efficacy of several established deep learning architectures\u0026mdash;DenseNet121, ResNet50, and ResNet101\u0026mdash;was evaluated. These models, pretrained on the ImageNet Large Scale Visual Recognition Challenge 2012 (ILSVRC-2012) dataset, were employed for analysis. To ensure intensity uniformity across the dataset, grayscale values of selected slices underwent min\u0026ndash;max normalization, scaled to the range [-1, 1]. Each cropped subregion image was subsequently resized to 224 \u0026times; 224 pixels via nearest neighbor interpolation to match the input requirements of the chosen models.\u003c/p\u003e\u003cp\u003eGiven dataset constraints, the learning rate was systematically optimized to enhance model generalizability via a cosine decay strategy. The implemented learning rate parameters were as follows:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:{\\eta\\:}_{t}={\\eta\\:}_{min}^{i}+\\frac{1}{2}\\left({\\eta\\:}_{max}^{i}-{\\eta\\:}_{min}^{i}\\right)\\left(1+cos\\left(\\frac{{T}_{cur}}{{T}_{i}}\\pi\\:\\right)\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe minimum learning rate, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\eta\\:}_{min}^{i}\\)\u003c/span\u003e\u003c/span\u003e, is set to 0, whereas the maximum learning rate, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\eta\\:}_{max}^{i}\\)\u003c/span\u003e\u003c/span\u003e, is set to 0.01. The parameter \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{i}\\)\u003c/span\u003e\u003c/span\u003e denotes the number of iteration epochs. Other hyperparameters are configured as follows: the optimizer is stochastic gradient descent (SGD), and the loss function used is softmax cross entropy. Details can be found in the Supplementary material.\u003c/p\u003e\u003cp\u003e\u003cb\u003eMulti-Instance Learning Fusion\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThis study implemented two fusion techniques embedded within the MIL framework. Under the MIL paradigm, a bag is labeled positive if it contains\u0026thinsp;\u0026ge;\u0026thinsp;1 positive instance, whereas it is assigned a negative label only when all instances are negative\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. Under this assumption, the MIL paradigm is well suited for image classification because images (bags) are generally classified according to some of their subregions (instances)\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. Details can be found in the Supplementary materials:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003ePredictive likelihood histogram (PLH)\u003c/b\u003e: Using 2.5D DL models, we generated histograms representing the distribution of predictive probabilities and labels across each slice in the 2.5D images, encapsulating image features.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eBag of Words (BoW)\u003c/b\u003e: The complete image was segmented into slices, with probabilities and predictions extracted from each. This process yielded 2 * 7 predictive results per sample (derived from both 2.5D and multimodel analyses). These results were treated analogously to word frequencies within a document, and TF-IDF weighting was applied to characterize the features.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eFeature Fusion\u003c/b\u003e: We combined features from PLH and BoW with radiomic features (based on CE-T1WI and FLAIR) to create a comprehensive feature set. This integrated approach fuses diverse data sources to represent image characteristics efficiently. These features were then used in machine learning algorithms to build models that enhance classification performance.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eModel Construction and Validation\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eClinical Signature\u003c/strong\u003e\u003cp\u003eUnivariate analyses were performed on clinical features via the same models applied to the 2.5D DL data to identify features associated with BM classification. A clinical model was subsequently constructed utilizing features that demonstrated statistical significance (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003e2.5D and 2D deep learning signatures\u003c/b\u003e: The feature selection process is described in the supplementary material. The combined features were input into machine learning algorithms similar to those used in radiomic feature modeling to develop the 2.5DMIL_Rad signature and 2D signature. Within the training set, 5-fold cross-validation was employed alongside grid search for hyperparameter optimization. To assess the effectiveness of multisequence fusion, an identical modeling approach was applied to evaluate performance via exclusively CE-T1WI data and exclusively FLAIR data.\u003c/p\u003e\u003cp\u003eReceiver operating characteristic (ROC) curves were plotted, and the area under the curve (AUC), accuracy, sensitivity and specificity were calculated to evaluate the performance of various classification models.\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eStatistical analysis\u003c/h2\u003e\u003cp\u003eThe normality of the clinical features was assessed via the Shapiro‒Wilk test. Continuous variables were evaluated for significance via analysis of variance (ANOVA), whereas categorical variables were analyzed with chi-square (χ\u0026sup2;) tests. All the data analyses were conducted via Python 3.7.12 on the OnekeyAI platform v3.1.8. For statistical analyses, we used statsmodels v0.13.2, and for radiomic feature extraction, we employed PyRadiomics v3.0.1. The machine learning algorithms were implemented via scikit-learn 1.0.2. All the DL models were developed via PyTorch 1.11.0, with CUDA 11.3.1 and cuDNN 8.2.1 for hardware acceleration.\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cb\u003eClinical baseline characteristics\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e presents the clinical characteristics of the patients in the two data cohorts. According to the inclusion and exclusion criteria, 328 patients were ultimately included in this study, including 208 males and 120 females, with an age range of 30\u0026ndash;85 years and an average age of 63.17\u0026thinsp;\u0026plusmn;\u0026thinsp;10.54 years. The training group included 229 patients, and the test group included 99 patients. Age, sex, number of tumors, maximum tumor diameter, and maximum edema diameter were not significantly different between the two cohorts (P\u0026thinsp;\u0026gt;\u0026thinsp;0.05).\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\u003eBaseline characteristics of the training and test cohorts.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\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=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFeature_name\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eALL\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003etrain\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003ePvalue\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAge\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e63.17\u0026thinsp;\u0026plusmn;\u0026thinsp;10.54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e62.88\u0026thinsp;\u0026plusmn;\u0026thinsp;10.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e63.85\u0026thinsp;\u0026plusmn;\u0026thinsp;11.57\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.247\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMaximun_diameter_of_edema\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e40.95\u0026thinsp;\u0026plusmn;\u0026thinsp;28.29\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e39.02\u0026thinsp;\u0026plusmn;\u0026thinsp;27.53\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e45.43\u0026thinsp;\u0026plusmn;\u0026thinsp;29.63\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.077\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNumber_of_tumors\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e4.14\u0026thinsp;\u0026plusmn;\u0026thinsp;5.29\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e4.05\u0026thinsp;\u0026plusmn;\u0026thinsp;5.48\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e4.33\u0026thinsp;\u0026plusmn;\u0026thinsp;4.83\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.296\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMaximun_tumor_diameter\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e23.25\u0026thinsp;\u0026plusmn;\u0026thinsp;11.64\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e22.92\u0026thinsp;\u0026plusmn;\u0026thinsp;12.14\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e24.02\u0026thinsp;\u0026plusmn;\u0026thinsp;10.40\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.145\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSex\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=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.353\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003emale\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e208(63.41)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e141(61.57)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e67(67.68)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFemale\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e120(36.59)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e88(38.43)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e32(32.32)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003ePerformance of the 2.5D and 2D Deep Learning Model\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe performance of each model on the basis of a single sequence can be seen in Supplementary Table\u0026nbsp;2 and Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Among the three DL models, ResNet50 achieved the best overall classification performance. Therefore, we used the feature set obtained from the ResNet50 model to construct RF, SVM, and LR classification models via the scikitlearn machine learning library. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e indicate the performance of ResNet50 on the 2.5D and 2D DL signatures. Figure\u0026nbsp;3 and Fig.\u0026nbsp;4 show the AUCs for the 2.5D and 2D models in the cohorts. For the 2.5D DL model, the LR model had a high AUC of 0.994 (95% CI: 0.991\u0026ndash;0.998) in the training cohort, indicating excellent discrimination. The AUC was slightly lower at 0.961 (95% CI: 0.938\u0026ndash;0.984) in the test cohort but still indicated good generalizability. The performance of the 2.5D ResNet50 model on the basis of a single sequence (CE-T1WI or FLAIR) is presented in Supplementary Table\u0026nbsp;4. The results demonstrated that the fusion model exhibited superior performance compared with the single-sequence model. In contrast, the predictive performance of the 2D DL model was poor, and the AUCs of the optimal machine learning classifiers were 0.890 and 0.686 in the training and test sets, respectively.\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\u003ePerformance of the 2.5D ResNet50 model based on comprehensive features (CE-T1WI and FLAIR) in the training and testing sets.\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=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\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\u003eACC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003e95%CI\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSEN\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eSPN\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003ePPV\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eNPV\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eCohort\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.967\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.994\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.991\u0026ndash;0.998\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.948\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.976\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.952\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.974\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.906\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.961\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.938\u0026ndash;0.984\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.859\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.929\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.859\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.929\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSVM\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\u003e0.973\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.962\u0026ndash;0.984\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.834\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.956\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.905\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.920\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSVM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.923\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.944\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.914\u0026ndash;0.975\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.859\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.955\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.904\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.931\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.872\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.946\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.927\u0026ndash;0.966\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.620\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.998\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.993\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.840\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.899\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.952\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.926\u0026ndash;0.978\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.717\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.990\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.973\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.875\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"9\"\u003eAbbreviations: \u003cb\u003eRF\u003c/b\u003e, random forest; \u003cb\u003eSVM\u003c/b\u003e, support vector machine; \u003cb\u003eLR\u003c/b\u003e, logistic regression; \u003cb\u003eAUC\u003c/b\u003e, area under the curve; \u003cb\u003eACC\u003c/b\u003e, accuracy; \u003cb\u003eSEN\u003c/b\u003e, sensitivity; \u003cb\u003eSPE\u003c/b\u003e, specificity; \u003cb\u003ePPV\u003c/b\u003e, positive predictive value; \u003cb\u003eNPV\u003c/b\u003e, negative predictive value\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\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\u003ePerformance of the 2D ResNet50 model based on comprehensive features (CE-T1WI and FLAIR) in the training and testing sets.\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=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\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\u003eACC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003e95%CI\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSEN\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eSPN\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003ePPV\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eNPV\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eCohort\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.723\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.753\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.715\u0026ndash;0.791\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.581\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.795\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.586\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.791\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.650\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.645\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.576\u0026ndash;0.713\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.515\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.717\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.477\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.747\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSVM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.736\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.752\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.714\u0026ndash;0.790\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.532\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.838\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.622\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.782\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSVM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.683\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.646\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.577\u0026ndash;0.715\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.434\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.808\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.531\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.741\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.803\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.890\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.866\u0026ndash;0.915\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.773\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.819\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.681\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.878\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.660\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.686\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.622\u0026ndash;0.751\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.525\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.727\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.491\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.754\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"9\"\u003eAbbreviations: \u003cb\u003eRF\u003c/b\u003e, random forest; \u003cb\u003eSVM\u003c/b\u003e, support vector machine; \u003cb\u003eLR\u003c/b\u003e, logistic regression; \u003cb\u003eAUC\u003c/b\u003e, area under the curve; \u003cb\u003eACC\u003c/b\u003e, accuracy; \u003cb\u003eSEN\u003c/b\u003e, sensitivity; \u003cb\u003eSPE\u003c/b\u003e, specificity; \u003cb\u003ePPV\u003c/b\u003e, positive predictive value; \u003cb\u003eNPV\u003c/b\u003e, negative predictive value\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eGrad-CAM\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo explore the recognition capabilities of DL models on diverse samples, we utilized the gradient-weighted class activation mapping (Grad-CAM) technique for visualization\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e5\u003c/span\u003e demonstrates the application of Grad-CAM, which shows the activations in the final convolutional layer that are pertinent to predicting BM types. This visualization highlights image regions that substantially influence the model's decision-making process, thereby enhancing our understanding of its interpretability.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study compared DL features derived from 2D versus 2.5D data within multiparametric MRI for classifying BMs from lung, breast, and gastrointestinal cancers. Our approach employed seven adjacent slices acquired from both the superior-inferior and anterior-posterior directions. DL features were extracted from each slice via three pretrained CNNs. A multiple instance learning (MIL) framework was then applied, treating each patient\u0026rsquo;s collection of MRI slices as a \"bag\" and individual slices as \"instances.\" The aggregated features from each bag served as input to machine learning classifiers. The aggregated features from each bag served as input to machine learning classifiers. This strategy synergizes the powerful feature extraction capabilities of DL, the contextual integration strengths of MIL, and the stability of traditional machine learning, enabling more accurate and reliable BM classification. In the test set, the optimal 2.5D classifier achieved an accuracy of 0.906 and an AUC of 0.961, outperforming the optimal 2D classifier (accuracy: 0.660; AUC: 0.686). This demonstrates the superior classification performance of the 2.5D DL model, indicating that 2.5D data provide greater spatial contextual information. Thus, we conclude that 2.5D-based DL models enhance the differentiation of BM types, facilitating the identification of primary tumor sites.\u003c/p\u003e\u003cp\u003eDL is an essential branch of machine learning and has demonstrated remarkable potential in medical image analysis through its multilayer artificial neural network architecture, which simulates the human cognitive system \u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. In recent years, this technology has been extensively investigated and clinically validated in the field of brain tumor classification research. Grossman et al \u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. employed three ImageNet-pretrained CNNs (VGG19, Xception, InceptionV3) to differentiate breast cancer (BC) and lung cancer (LC) BMs via CE-T1WI. Their 3D CNN models achieved high diagnostic efficacy (accuracy: 0.82\u0026ndash;0.85). Similarly, in a subsequent study, Gultekin et al \u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. used three pretrained CNNs and texture-based 2D and 3D tumor image features to distinguish LC and BC BMs. The study results showed that the overall performance of the CNN architectures using 3D ROIs as inputs was greater than that of the architectures based on texture features, with the 3D Xception architecture achieving good performance (accuracy: 0.85; AUC: 0.84). In broader tumor classification research, Kumar et al. \u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e implemented multiple architectures (ResNet, AlexNet, U-Net, VGG-16) for MRI-based brain tumor categorization. Their modified ResNet50 achieved exceptional accuracy (benign: 0.993; malignant: 0.984). The variation in performance among the different models can be attributed to disparities in the network's internal architecture \u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. To select the best CNNs for diagnosing BMs, this study developed three commonly used diagnostic models based on different neural network architectures, including ResNet50, ResNet101, and DenseNet121. The pretrained ResNet50 model performed well in the BM classification task, achieving an average accuracy of 0.899. ResNet-50 uses residual learning to reduce gradient dispersion and precision loss in deep networks, which can improve model accuracy while accelerating the training of neural networks. Another notable feature of ResNet-50 is the use of a global average pooling layer, which takes the average value of all the pixels in each feature map as the output of that feature map, which reduces the number of model parameters and lowers the risk of overfitting \u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eIn the field of image segmentation via machine learning models, 2D and 3D approaches represent the two primary methodologies. While 3D CNNs capture richer contextual information and generally outperform 2D CNNs do, they require significantly more computational resources, including GPU memory and processing time \u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. To address the limitations of both 2D and 3D CNNs, the 2.5D segmentation method has been proposed and recommended in numerous studies \u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. This approach combines the computational efficiency of 2D CNNs with the enhanced spatial context capture capabilities inherent to 3D CNNs \u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. Recent research has demonstrated the excellent performance of 2.5D segmentation methods across various tasks \u003csup\u003e\u003cspan additionalcitationids=\"CR29\" citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e. For example, Takao et al.\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e developed a 2.5D DL model for the automatic detection of BMs, which uses three consecutive slices as inputs to predict the central slice. Their comparative analysis revealed superior overall performance relative to a conventional 2D model using a single slice. Similarly, Xiong et al. \u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e. investigated the value of DL models in distinguishing bone islands from osteoblastic bone metastases. Their findings indicated that a 2.5D DL model employing three-layer CT images as input outperformed 2D models in classifying sclerotic bone lesions. This model achieved high performance both internally (AUC, 0.996) and on two external validation sets (AUC, 0.958; AUC, 0.952). These results are consistent with the findings of the current study. Herein, the largest cross-section of the ROI was designated the primary slice, followed by the extraction of 2D images from multiple adjacent slices to form a 2.5D dataset incorporating partial 3D structural information. The results demonstrate effective performance in the differential diagnosis of three BM types via the 2.5D DL model, which also surpassed the performance of the 2D model. However, future research should aim to develop a comprehensive 3D model based on multisequence MRI to rigorously compare the performance differences and efficacy among various segmentation models for BM classification tasks.\u003c/p\u003e\u003cp\u003eTo validate the effectiveness of the proposed 2.5D DL model, we compared it with relevant studies in recent years. The analysis in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e reveals that most studies have constructed classification models using radiomic features, whereas only Gultekin et al.\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e, Jiao et al.\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e, and Lyu et al.\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e have explored the classification performance of 2D or 3D DL models. Notably, Lyu et al.\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e established an end-to-end DL model using a large dataset, which improved the model's generalization ability and stability by optimizing the training process and reducing manual intervention. The work of Jiao et al.\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e is also noteworthy. By using the ResNet18 model to classify breast cancer and gastrointestinal BMs and validating the model with datasets from multiple centers, they enhanced the model's generalization ability, achieving an AUC value of 0.848. However, the accuracy performance was moderate (ACC\u0026thinsp;=\u0026thinsp;0.700). Several radiomic studies reported in the literature have successfully linked MR/CT image features to the classification of BMs \u003csup\u003e\u003cspan additionalcitationids=\"CR34 CR35 CR36 CR37\" citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e. However, these studies have focused mainly on binary classification. In contrast, our method outperforms traditional radiomic methods in both Ortiz-Ram\u0026oacute;n's \u003csup\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003ethree-class classification (accuracy of 0.906 vs 0.873) and Kniep's \u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003efive-class classification (accuracy of 0.906 vs 0.820). Compared with radiomics, our 2.5D DL model has several advantages in distinguishing BMs. First, the ROIs selection method based on 2.5D DL only needs to encompass the entire lesion instead of precisely delineating the lesion edge layer by layer, as in radiomics. Second, our study recruited a greater number of patients. A larger sample size can guarantee the reliability of a better classification model. Nevertheless, this study and most other studies have a single-center retrospective design. The generalizability of the model remains to be verified, which is the direction that future research needs to address.\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\u003eSummary of the application of machine learning methods in the classification of brain metastases in recent years.\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=\"char\" char=\".\" 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=\"char\" char=\".\" 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\u003ePaper\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSingle/Multi-Center\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSample Size\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003ePrimary cancer\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eOptimal classifier\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eACC\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOur study\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSingle-Center\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e328\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eLC vs BC vs GIC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2.5D ResNet50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.952\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.899\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGultekin et al 2023[22]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSingle-Center\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e143\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eLC vs BC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2D Xception\u003c/p\u003e\u003cp\u003e3D Xception\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.800\u003c/p\u003e\u003cp\u003e0.850\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.830\u003c/p\u003e\u003cp\u003e0.845\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eJiao et al [31]2023\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMulti-Center\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e214\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eBC vs GIC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e3D ResNet 18\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.848\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.700\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLyu et al [32]2022\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSingle-Center\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1582\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eLC vs BC vs Melanoma vs Renal vs Other\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e3D cycle-GAN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.878\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eShi et al [33]\u003c/p\u003e\u003cp\u003e2023\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSingle-Center\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e160\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eLC vs BC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e3D LR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.778\u003c/p\u003e\u003cp\u003e0.843\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.717\u003c/p\u003e\u003cp\u003e0.792\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCao et al [34]2022\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSingle-Center\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e78\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eLC vs BC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSVM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.805\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eKniep et al [35]2019\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSingle-Center\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e189\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSCLC vs BC vs melanoma vs GIC vs NSCLC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e3D RF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.64\u0026ndash;0.82\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOrtiz-Ram\u0026oacute;n et al [36]2018\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSingle-Center\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e38\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eLC vs melanoma vs BC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e3D RF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.873\u0026thinsp;\u0026plusmn;\u0026thinsp;0.064\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOrtiz-Ram\u0026oacute;n\u003c/p\u003e\u003cp\u003eet al [37]2017\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSingle-Center\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e29\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eLC vs BC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2D SVM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.953\u0026thinsp;\u0026plusmn;\u0026thinsp;0.061\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOrtiz-Ram\u0026oacute;n\u003c/p\u003e\u003cp\u003eet al [38]2017\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSingle-Center\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e30\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eLC vs Melanoma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2D KNN\u003c/p\u003e\u003cp\u003e3DNB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.890\u0026thinsp;\u0026plusmn;\u0026thinsp;0.085\u003c/p\u003e\u003cp\u003e0.947\u0026thinsp;\u0026plusmn;\u0026thinsp;0.067\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e-\u003c/p\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"7\"\u003eAbbreviations: \u003cb\u003eRF\u003c/b\u003e, random forest; \u003cb\u003eAUC\u003c/b\u003e, area under the curve; \u003cb\u003eACC\u003c/b\u003e, accuracy; \u003cb\u003eLC\u003c/b\u003e, lung cancer; \u003cb\u003eBC\u003c/b\u003e, breast cancer; \u003cb\u003eGIC\u003c/b\u003e, gastrointestinal cancer; \u003cb\u003eSVM\u003c/b\u003e, support vector machine; \u003cb\u003ek-NN\u003c/b\u003e, k-nearest neighbors; \u003cb\u003eLR\u003c/b\u003e, logistic regression; \u003cb\u003eNB\u003c/b\u003e, na\u0026iuml;ve Bayes; \u003cb\u003eNSCLC\u003c/b\u003e, non-small cell lung cancer; \u003cb\u003eSCLC\u003c/b\u003e, small cell lung cancer\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThis study has the following limitations. First, the study employed a retrospective, single-center design and lacked external data to validate the model's efficacy. Second, we only considered metastases from three primary sites, necessitating further research to collect data on other types of metastases. Additionally, although LC are widely regarded as the main source of BMs, the significant discrepancies in sample sizes among various categories in this study pose a problem that cannot be overlooked \u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e. This unbalanced sample distribution is likely to impose restrictions on the model's capacity for identifying and classifying samples of minority classes. To address this challenge, future studies should consider employing specific data processing techniques, such as the synthetic minority oversampling technique (SMOTE), to handle the issue of unbalanced sample sizes more effectively\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn summary, the MRI-based 2.5D deep learning model can serve as a sensitive and specific diagnostic tool for distinguishing between different pathological types of BMs. This innovative method not only helps improve the diagnostic level of BMs but also provides an important basis for subsequent personalized treatment and prognosis evaluation.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eBMs \u0026nbsp; brain metastases\u003c/p\u003e\n\u003cp\u003eCE-T1WI \u0026nbsp; contrast-enhanced T1-weighted imaging\u003c/p\u003e\n\u003cp\u003eDL \u0026nbsp; deep learning\u003c/p\u003e\n\u003cp\u003e2D \u0026nbsp; two-dimensional\u003c/p\u003e\n\u003cp\u003e3D \u0026nbsp; three-dimensional\u003c/p\u003e\n\u003cp\u003eFLAIR \u0026nbsp; fluid-attenuated inversion recovery images\u003c/p\u003e\n\u003cp\u003eMIL \u0026nbsp; multi-instance learning\u003c/p\u003e\n\u003cp\u003eRF \u0026nbsp; random forest\u003c/p\u003e\n\u003cp\u003eSVM \u0026nbsp; support vector machine\u003c/p\u003e\n\u003cp\u003eLR \u0026nbsp; logistic regression\u003c/p\u003e\n\u003cp\u003eLC \u0026nbsp; Lung cancer\u003c/p\u003e\n\u003cp\u003eBC \u0026nbsp; breast cancer\u003c/p\u003e\n\u003cp\u003eGIC \u0026nbsp; gastrointestinal cancer\u003c/p\u003e\n\u003cp\u003eAUC \u0026nbsp; area under the curve\u003c/p\u003e\n\u003cp\u003e95% CI \u0026nbsp; 95% confidence interval\u003c/p\u003e\n\u003cp\u003eGrad-CAM \u0026nbsp; \u0026nbsp;gradient-weighted class activation mapping\u003c/p\u003e\n\u003cp\u003eCNNs \u0026nbsp; Convolutional neural networks\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was conducted in accordance with the Declaration of Helsinki\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eand was approved by the Ethics Committee of The Second Affiliated Hospital of Soochow University. The ethics committee waived the requirement for informed consent due to the retrospective nature of the study using anonymized data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analyzed during the current study are available from\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ethe corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by Suzhou Science and Technology Development Plan Project in China [grant numbers SLJ2022009]; Suzhou Science and Technology Development Plan Project (Medical and Health Technology Innovation) in China [grant numbers SYSD2022112]; Exploratory Scientific Research Grant Project of the Second Affiliated Hospital of Soochow University in China [grant numbers SDFEYBS2425].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBo Zhang. conceived and designed the research framework. Jinling Zhu.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eserved as the major contributor to manuscript drafting.Jixuan Deng. performed data acquisition and curation. Ruizhe Xu. and Li Zou. developed the deep learning models and performed the statistical analyses. Xin Xie.,Ye Tian., and Wu Cai. offered technical supervision and methodological refinement during model construction. All authors reviewed the manuscript.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors sincerely thank Platform Onekey AI for Python technology of the study.\u003cstrong\u003e\u003cbr\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eBalestrino R, Rud\u0026agrave; R, Soffietti R. Brain metastasis from unknown primary tumour: moving from old retrospective studies to clinical trials on targeted agents. Cancers (Basel) 2020;12\u003c/li\u003e\n\u003cli\u003eLamba N, Wen PY, Aizer AA. Epidemiology of brain metastases and leptomeningeal disease. Neuro Oncol 2021;23:1447-1456\u003c/li\u003e\n\u003cli\u003eSuh JH, Kotecha R, Chao ST, et al. Current approaches to the management of brain metastases. Nat Rev Clin Oncol 2020;17:279-299\u003c/li\u003e\n\u003cli\u003eRassy E, Zanaty M, Azoury F, et al. Advances in the management of brain metastases from cancer of unknown primary. Future Oncol 2019;15:2759-2768\u003c/li\u003e\n\u003cli\u003eWolpert F, Weller M, Berghoff AS, et al. Diagnostic value of (18)F-fluordesoxyglucose positron emission tomography for patients with brain metastasis from unknown primary site. Eur J Cancer 2018;96:64-72\u003c/li\u003e\n\u003cli\u003eDeVries DA, Lagerwaard F, Zindler J, et al. Performance sensitivity analysis of brain metastasis stereotactic radiosurgery outcome prediction using MRI radiomics. Sci Rep 2022;12:20975\u003c/li\u003e\n\u003cli\u003eCagney DN, Martin AM, Catalano PJ, et al. Incidence and prognosis of patients with brain metastases at diagnosis of systemic malignancy: a population-based study. Neuro Oncol 2017;19:1511-1521\u003c/li\u003e\n\u003cli\u003eBae S, An C, Ahn SS, et al. Robust performance of deep learning for distinguishing glioblastoma from single brain metastasis using radiomic features: model development and validation. Sci Rep 2020;10:12110\u003c/li\u003e\n\u003cli\u003eZhang H, Zhang H, Zhang Y, et al. Deep learning radiomics for the assessment of telomerase reverse transcriptase promoter mutation status in patients with glioblastoma using multiparametric MRI. J Magn Reson Imaging 2023;58:1441-1451\u003c/li\u003e\n\u003cli\u003eTulum G. Novel radiomic features versus deep learning: differentiating brain metastases from pathological lung cancer types in small datasets. Br J Radiol 2023;96:20220841\u003c/li\u003e\n\u003cli\u003eTian Y, Xue F, Lambo R, et al. Fully automated functional region annotation of liver via a 2.5D class-aware deep neural network with spatial adaptation. Comput Methods Programs Biomed 2021;200:105818\u003c/li\u003e\n\u003cli\u003eZhang Y, Liao Q, Ding L, et al. Bridging 2D and 3D segmentation networks for computation-efficient volumetric medical image segmentation: An empirical study of 2.5D solutions. Comput Med Imaging Graph 2022;99:102088\u003c/li\u003e\n\u003cli\u003eYin S, Peng Q, Li H, et al. Multi-instance deep learning of ultrasound imaging data for pattern classification of congenital abnormalities of the kidney and urinary tract in children. Urology 2020;142\u003c/li\u003e\n\u003cli\u003eXiao Y, Liang F, Liu B. A Transfer Learning-Based Multi-Instance Learning Method With Weak Labels. IEEE Trans Cybern 2022;52:287-300\u003c/li\u003e\n\u003cli\u003eZhu J, Zou L, Xie X, et al. 2.5D deep learning based on multiparameter MRI to differentiate primary lung cancer pathological subtypes in patients with brain metastases. Eur J Radiol 2024;180:111712\u003c/li\u003e\n\u003cli\u003eMoradmand H, Aghamiri SMR, Ghaderi R. Impact of image preprocessing methods on reproducibility of radiomic features in multimodal magnetic resonance imaging in glioblastoma. J Appl Clin Med Phys 2020;21:179-190\u003c/li\u003e\n\u003cli\u003eWang J, Cai L, Peng J, et al. A novel multiple instance learning method based on extreme learning machine. Comput Intell Neurosci 2015;2015:405890\u003c/li\u003e\n\u003cli\u003eAstorino A, Fuduli A, Veltri P, et al. Melanoma detection by means of multiple instance learning. Interdiscip Sci 2020;12:24-31\u003c/li\u003e\n\u003cli\u003eZhang H, Ogasawara K. Grad-CAM-Based Explainable Artificial Intelligence Related to Medical Text Processing. Bioengineering (Basel) 2023;10\u003c/li\u003e\n\u003cli\u003eLee JG, Jun S, Cho YW, et al. Deep Learning in Medical Imaging: General Overview. Korean J Radiol 2017;18:570-584\u003c/li\u003e\n\u003cli\u003eGrossman R, Haim O, Abramov S, et al. Differentiating small-cell lung cancer from non-small cell lung cancer brain metastases based on MRI using efficientnet and transfer learning approach. Technol Cancer Res Treat 2021;20:15330338211004919\u003c/li\u003e\n\u003cli\u003eGultekin MA, Peker AA, Oktay AB, et al. Differentiation of lung and breast cancer brain metastases: Comparison of texture analysis and deep convolutional neural networks. J Clin Ultrasound 2023;51:1579-1586\u003c/li\u003e\n\u003cli\u003eKumar S, Choudhary S, Jain A, et al. Brain tumor classification using deep neural network and transfer learning. Brain Topogr 2023;36:305-318\u003c/li\u003e\n\u003cli\u003eYu Q, Ning Y, Wang A, et al. Deep learning-assisted diagnosis of benign and malignant parotid tumors based on contrast-enhanced CT: a multicenter study. Eur Radiol 2023;33:6054-6065\u003c/li\u003e\n\u003cli\u003eHe K, Zhang X, Ren S, et al. Deep residual learning for image recognition. IEEE 2016\u003c/li\u003e\n\u003cli\u003eMzoughi H, Njeh I, Wali A, et al. Deep multi-scale 3D convolutional neural network (cnn) for mri gliomas brain tumor classification. J Digit Imaging 2020;33:903-915\u003c/li\u003e\n\u003cli\u003eAvesta A, Hossain S, Lin M, et al. Comparing 3D, 2.5D, and 2D approaches to brain image autosegmentation. Bioengineering (Basel) 2023;10\u003c/li\u003e\n\u003cli\u003eHuang L, Zhao Z, An L, et al. 2.5D transfer deep learning model for segmentation of contrast-enhancing lesions on brain magnetic resonance imaging of multiple sclerosis and neuromyelitis optica spectrum disorder. Quant Imaging Med Surg 2024;14:273-290\u003c/li\u003e\n\u003cli\u003eTakao H, Amemiya S, Kato S, et al. Deep-learning 2.5-dimensional single-shot detector improves the performance of automated detection of brain metastases on contrast-enhanced CT. Neuroradiology 2022;64:1511-1518\u003c/li\u003e\n\u003cli\u003eXiong Y, Guo W, Liang Z, et al. Deep learning-based diagnosis of osteoblastic bone metastases and bone islands in computed tomograph images: a multicenter diagnostic study. Eur Radiol 2023;33:6359-6368\u003c/li\u003e\n\u003cli\u003eJiao T, Li F, Cui Y, et al. Deep learning with an attention mechanism for differentiating the origin of brain metastasis using MR images. J Magn Reson Imaging 2023;58:1624-1635\u003c/li\u003e\n\u003cli\u003eLyu Q, Namjoshi SV, McTyre E, et al. A transformer-based deep-learning approach for classifying brain metastases into primary organ sites using clinical whole-brain MRI images. Patterns (N Y) 2022;3:100613\u003c/li\u003e\n\u003cli\u003eShi J, Chen H, Wang X, et al. Using Radiomics to Differentiate brain metastases from lung cancer versus breast cancer, including predicting epidermal growth factor receptor and human epidermal growth factor receptor 2 status. J Comput Assist Tomogr 2023;47:924-933\u003c/li\u003e\n\u003cli\u003eCao G, Zhang J, Lei X, et al. Differentiating primary tumors for brain metastasis with integrated radiomics from multiple imaging modalities. Dis Markers 2022;2022:5147085\u003c/li\u003e\n\u003cli\u003eKniep HC, Madesta F, Schneider T, et al. Radiomics of brain MRI: utility in prediction of metastatic tumor type. Radiology 2019;290:479-487\u003c/li\u003e\n\u003cli\u003eOrtiz-Ram\u0026oacute;n R, Larroza A, Ruiz-Espa\u0026ntilde;a S, et al. Classifying brain metastases by their primary site of origin using a radiomics approach based on texture analysis: a feasibility study. Eur Radiol 2018;28:4514-4523\u003c/li\u003e\n\u003cli\u003eOrtiz-Ramon R, Larroza A, Arana E, et al. Identifying the primary site of origin of MRI brain metastases from lung and breast cancer following a 2D radiomics approach. 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017); 2017\u003c/li\u003e\n\u003cli\u003eOrtiz-Ramon R, Larroza A, Arana E, et al. A radiomics evaluation of 2D and 3D MRI texture features to classify brain metastases from lung cancer and melanoma. Annu Int Conf IEEE Eng Med Biol Soc 2017;2017:493-496\u003c/li\u003e\n\u003cli\u003eShi W, Tanzhu G, Chen L, et al. Radiotherapy in preclinical models of brain metastases: a review and recommendations for future studies. Int J Biol Sci 2024;20:765-783\u003c/li\u003e\n\u003cli\u003eDablain D, Krawczyk B, Chawla NV. DeepSMOTE: fusing deep learning and smote for imbalanced data. IEEE Trans Neural Netw Learn Syst 2023;34:6390-6404\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Brain Metastases, 2.5D Deep Learning, Multi-Instance Learning, Magnetic Resonance Imaging","lastPublishedDoi":"10.21203/rs.3.rs-7210303/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7210303/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eObjective\u003c/h2\u003e\u003cp\u003eThis study aims to compare deep learning (DL) features extracted from 2D and 2.5D data via multiparametric magnetic resonance imaging (MRI) to classify brain metastases (BMs) originating from lung cancer (LC), breast cancer (BC), and gastrointestinal cancer (GIC).\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eThis retrospective study analyzed MR images from 328 patients with brain metastases (BMs), which were randomly divided into training (N\u0026thinsp;=\u0026thinsp;229) and test (N\u0026thinsp;=\u0026thinsp;99) sets at a 7:3 ratio. From the primary lesion slice, we obtained adjacent slices in both the superior-inferior and anterior-posterior directions, constructing a series of two-dimensional (2D) images. DL features were extracted from these slices via pretrained convolutional neural networks (CNNs), including DenseNet121, ResNet50, and ResNet101. A multi-instance learning (MIL) framework was then applied to integrate features into a comprehensive representation. The 2D model, which uses the tumor\u0026rsquo;s maximal cross-sectional slice as input, followed an identical processing pipeline to that of the 2.5D model. All feature sets were evaluated via machine learning algorithms. Diagnostic performance was assessed via fivefold cross-validation, with accuracy and area under the curve (AUC) metrics quantified for analysis.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eThe best classification results were obtained from multiparametric MR images combined with ResNet50. In the test set, the accuracy and AUC of the optimal 2.5D model were 0.906 and 0.961 (95% confidence interval [CI], 0.926\u0026ndash;0.978), respectively. The accuracy and AUC of the optimal 2D model were 0.660 and 0.686 (95% CI, 0.622\u0026ndash;0.751), respectively.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eOn the basis of multiparametric MRI data, the 2.5D-based DL model is feasible for distinguishing the origins of BMs.\u003c/p\u003e","manuscriptTitle":"Comparison of 2D and 2.5D deep learning features based on multi-parametric magnetic resonance imaging for brain metastases classification","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-23 12:10:52","doi":"10.21203/rs.3.rs-7210303/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"36a5149f-630b-42fe-a280-fca012e21e09","owner":[],"postedDate":"September 23rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-02-26T08:11:51+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-23 12:10:52","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7210303","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7210303","identity":"rs-7210303","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}