SE-Attention U-Net: A Hybrid Loss-Optimized Model for Small Breast Lesion Segmentation in Mammography

SE-Attention U-Net: A Hybrid Loss-Optimized Model for Small Breast Lesion Segmentation in Mammography | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article SE-Attention U-Net: A Hybrid Loss-Optimized Model for Small Breast Lesion Segmentation in Mammography Maliheh Habibi This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6800663/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 10 You are reading this latest preprint version Abstract Breast cancer remains a leading cause of mortality among women worldwide, with early detection via mammography significantly improving patient outcomes. Automated segmentation of mammographic lesions using deep learning can enhance diagnostic efficiency; however, existing methods face critical challenges: (1) severe class imbalance (< 2% foreground pixels), (2) small lesion sizes (3–15 mm), and (3) limited annotated datasets, which hinder clinical applicability. To overcome these limitations, we propose SE-Attention U-Net, a hybrid framework featuring squeeze-and-excitation blocks for adaptive feature refinement, attention gates to focus on salient regions, and a novel loss function explicitly designed to address extreme class imbalance. We evaluated our approach on the publicly available CBIS-DDSM dataset, a widely recognized benchmark in mammography research. Our model achieved state-of-the-art performance with a Dice coefficient of 91.00%, Jaccard coefficient of 86.01%, accuracy of 99.54%, precision of 97.97%, sensitivity of 97.08%, and an F1-score of 97.53%. These results demonstrate robust lesion localization and minimized false positives, outperforming existing methods. The proposed framework shows significant potential for clinical integration, providing radiologists with a reliable tool for early and accurate segmentation of small breast lesions. Biological sciences/Cancer Health sciences/Medical research Health sciences/Oncology Mammographic lesion localization Lesion segmentation U-Net architecture Attention mechanisms squeeze-and-excitation blocks Class imbalance Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 1. Introduction Breast cancer is a major global health concern and one of the leading causes of cancer-related mortality among women. Early detection through screening is vital for improving patient outcomes, with mammography serving as the primary diagnostic tool. However, accurately segmenting small masses in mammograms remains a significant challenge due to tissue overlap, image noise, and low contrast, which often lead to missed diagnoses or delayed treatment. The task is further complicated by the extreme class imbalance between the tumor (foreground) and healthy tissue (background) pixels, as well as the irregular and subtle appearance of small masses. Despite advancements in imaging technology, developing reliable methods for precise delineation of these tiny anomalies is essential to enhance early detection and improve clinical outcomes [ 1 ]. Recent advancements in deep learning, especially convolutional neural networks (CNNs), hold great promise in tackling these challenges. Multiple studies [ 2 – 6 ] have demonstrated the potential of deep learning to significantly enhance medical image analysis. However, these approaches predominantly focus on larger masses and typically rely on extensive annotated datasets, which are often scarce in practical clinical environments. Even the U-Net architecture, widely adopted for biomedical segmentation [ 7 ], underperforms due to inherent tradeoffs between maintaining spatial information and resolving uncertain tumor interfaces or minor intensity variations [ 8 ]. In medical image analysis, enhancement techniques like Histogram Equalization (HE) and its variants—such as Global HE (GHE)[ 9 ], Local HE (LHE) [ 10 ], Adaptive HE (AHE)[ 11 , 12 ], and Contrast-Limited Adaptive Histogram Equalization (CLAHE) [ 9 , 13 – 15 ]—have been widely adopted. However, CLAHE has emerged as a preferred method due to its ability to enhance local contrast adaptively while constraining noise amplification through histogram clipping. This makes CLAHE indispensable for early breast cancer detection, where preserving fine details and avoiding artifacts are paramount. Despite advances in CNN architectures, the integration of attention mechanisms for enhanced feature extraction in mammographic image analysis remains underdeveloped [ 16 ]. While attention modules like Squeeze-and-Excitation (SE) blocks have proven effective in boosting segmentation accuracy by enabling channel-wise feature recalibration [ 17 ] and improving sensitivity to subtle features [ 18 ], their application to mammography - particularly for detecting small, irregular masses - has not been fully explored. Given the shortcomings of existing methods, this research aims to address these critical gaps through the proposal of an enhanced U-Net architecture that integrates several key components designed to improve the segmentation of small masses in mammograms. Our primary contributions include: Enhanced Preprocessing : A robust preprocessing pipeline that significantly improves image quality, utilizing adaptive histogram equalization [ 14 , 19 , 20 ] and Gaussian denoising techniques. This approach has been shown to effectively enhance contrast and reduce noise in medical imaging. Hybrid Attention Architecture : The implementation of a hybrid attention architecture combining channel-wise squeeze-excitation blocks with spatial attention gates, specifically optimized for detecting masses smaller than 15 mm. This architecture has been proven to refine feature extraction and boost accuracy in challenging segmentation tasks [ 17 , 18 ]. Novel Stable Extreme Loss Function : A dynamically weighted loss function that combines focal loss and adaptive Dice loss to address the significant class imbalances inherent in mammographic datasets[ 21 ]. By effectively penalizing misclassifications of small masses, this loss function enhances the model's ability to segment small, irregularly shaped masses accurately. Medical Augmentation Pipeline : To improve model generalization and address data limitations, we developed a comprehensive augmentation approach combining lesion-specific and whole-image transformations. For mass regions, we applied carefully boundary-aware flips (horizontal/vertical), limited rotations (± 15°), and localized contrast enhancement. Whole-image augmentations incorporated random flips (50% probability) and controlled brightness variations (± 10) [ 20 ]. This dual strategy enhances feature learning while preserving diagnostic relevance of mammographic findings. By leveraging these innovative methodologies, we aim to establish a new standard for automated mammographic analysis, ultimately providing healthcare professionals with a more reliable and accurate tool for breast cancer detection and diagnosis. This research not only improves segmentation capabilities for small masses but also contributes to the broader field of medical imaging by addressing significant limitations present in existing models. 2. Related Works In recent years, the challenge of small mass segmentation in mammography has attracted significant attention, leading to numerous studies exploring various deep learning methodologies. However, despite the advancements made, substantial gaps remain in managing class imbalances, enhancing feature extraction, and generating realistic synthetic data suitable for model training. Early convolutional approaches demonstrated promise for lesion detection. Still, they were fundamentally limited in handling small masses under 15 mm due to resolution constraints [ 1 ]. The introduction of U-Net architectures represented a pivotal development, leveraging innovative skip connections to preserve critical spatial details [ 7 ]. However, these architectures continued to struggle with the extreme class imbalance that is inherent to mammographic datasets [ 22 ]. In the realm of preprocessing, adaptive histogram equalization methods, such as CLAHE, became standard for contrast enhancement [ 13 – 15 , 19 ], while subsequent refinements incorporated tissue-specific parameter optimization that better preserved subtle lesion boundaries. Denoising techniques further evolved from basic non-local means filters [ 23 ] to more sophisticated hybrid approaches that combine wavelet transforms with edge preservation [ 24 , 25 ], although none specifically targeted the unique noise profiles typically found around small masses. Recent advancements have emphasized the use of attention mechanisms in deep learning frameworks. Research by [ 17 , 18 ] revealed the benefits of squeeze-and-excitation blocks and spatial attention for refining feature extraction. Subsequent hybrid architectures, such as those presented by Lie et al. [ 26 ] and Si et al. [ 27 ], demonstrated the complementary benefits of integrating channel and spatial attention mechanisms. However, adaptations specific to mammography remain underexplored. Our innovative hybrid attention architecture combines channel-wise squeeze-excitation blocks with spatial attention gates specifically optimized for detecting masses smaller than 15 mm. This integration significantly enhances segmentation accuracy in our model. Additionally, we present a novel Stable Extreme Loss function that dynamically adjusts weights by combining boundary-aware Dice loss [ 28 , 29 ], mass-presence weighted focal loss [ 30 , 31 ], and edge consistency loss. This approach effectively addresses the extreme class imbalance encountered in mammographic datasets and improves segmentation performance. By tackling these critical issues, our work establishes a new standard for automated mammographic analysis, offering healthcare professionals a more reliable and accurate tool for breast cancer detection and diagnosis. Through enhanced preprocessing, advanced attention mechanisms, and finely-tuned loss functions, we believe our approach not only bridges existing gaps in the literature but also paves the way for future advances in medical image analysis. 3. Materials and methodology This section outlines the comprehensive methodology employed in our paper to enhance the segmentation of small masses in mammographic images. As illustrated in Fig. 1 , our approach incorporates advanced preprocessing techniques, a uniquely tailored network architecture, and robust training strategies aimed at improving the detection and segmentation of small masses in mammograms. Our proposed method is evaluated using the CBIS-DDSM dataset [ 32 , 33 ]. The Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM) is a publicly available, curated dataset designed to facilitate research in medical image analysis, particularly for breast cancer detection. It is an improved and standardized version of the original DDSM dataset. The CBIS-DDSM comprises 2,620 scanned film mammograms—1,566 cases with calcifications and 1,054 with masses—each with bilateral views (CC and MLO), as illustrated in Figs. 2 and 3 . Every image is meticulously annotated by radiologists, including lesion type (calcification/mass), pathology labels (benign/malignant), bounding boxes, and pixel-wise segmentation masks. Calcifications tend to display fine-grained patterns (e.g., clustered or diffuse), while masses vary in morphology (e.g., spiculated or circumscribed), enabling robust training for both detection and characterization. The inclusion of paired CC (craniocaudal) and MLO (mediolateral oblique) views per case provides comprehensive spatial context, which is critical for reducing false positives and enhancing diagnostic accuracy. By leveraging CBIS-DDSM’s dual-pathology annotations and bilateral imaging, our work addresses segmentation challenges across both calcifications and masses, with particular focus on small lesions. Each mammogram is paired with pixel-level annotations (ROIs) for masses and calcifications, along with radiologist-assessed BI-RADS ratings, pathology labels (benign/malignant), and lesion diameters—ranging from 3 to 30 mm, with most masses under 15 mm. 3.1 Image Pre-processing Our preprocessing pipeline combines normalization, Adaptive Histogram Equalization (AHE) [ 22 ], Gaussian denoising, and edge sharpening to address key challenges in mammographic imaging, including low contrast, noise artifacts, and soft tissue ambiguity. By adaptively enhancing contrast, reducing noise, and accentuating structural details, this approach ensures high-quality input for downstream deep learning tasks, thereby improving feature extraction and segmentation accuracy. The four-stage workflow (Fig. 1 ) effectively optimizes image quality, as illustrated in Fig. 4 . Normalization Mammographic images often exhibit inter-scanner intensity variations due to differences in acquisition protocols, equipment, or digitization processes. To standardize input data and mitigate these inconsistencies, we apply min-max normalization, rescaling each image’s pixel intensities to a fixed range of [0, 255]. This range aligns with standard 8-bit grayscale representations, ensuring compatibility with conventional deep learning frameworks. Adaptive Contrast Enhancement To enhance the visibility of subtle lesions obscured by dense breast tissue, we applied Contrast Limited Adaptive Histogram Equalization (CLAHE) [ 14 ]. This method improves local contrast by dividing the image into non-overlapping 16×16 tile grids and performing histogram equalization within each region, constrained by a clip limit of 3.0 to prevent noise amplification. Unlike global histogram equalization, CLAHE adapts to local intensity variations, making it particularly effective for mammograms where small masses or microcalcifications may exhibit low contrast against heterogeneous backgrounds. The clip limit ensures balanced enhancement by redistributing only the histogram bins exceeding the threshold, thereby preserving anatomical details while mitigating artificial artifacts. Structure-Preserving Denoising To enhance the signal-to-noise ratio while preserving critical anatomical structures, we implemented Non-Local Means (NLM) [ 23 ] Denoising with optimized parameters (decay parameter h = 20, search window size = 21×21 pixels). The selected parameters were carefully tuned to achieve an optimal balance between noise suppression and preservation of subtle pathological features, such as spiculated margins or microcalcifications. The NLM algorithm's effectiveness stems from its ability to maintain structural integrity while reducing noise. Edge Sharpening To improve the delineation of lesion boundaries and subtle mass margins critical for accurate diagnosis, we implemented a conservative edge sharpening approach using a 3×3 high-pass kernel. This final preprocessing step selectively enhances high-frequency components while maintaining strict control over potential artifacts through careful parameterization. The sharpening process, applied after denoising but prior to network input, provides an optimal balance between feature enhancement and natural appearance. 3.2 Data Augmentation To increase the variability of the training dataset and enhance model robustness, we employed various data augmentation techniques, focusing on both global and lesion-specific modifications. For mass-centric augmentation, applied only to lesion-containing regions, we performed safe flips (horizontal and vertical) with boundary checks to prevent lesion cropping, small-angle rotations within ± 15° to maintain anatomical integrity, and localized contrast enhancement using adaptive histogram equalization to improve the visibility of mass regions. For global augmentation, all images were randomly flipped (by 50%) along both vertical and horizontal axes, and brightness adjustments within a controlled range to simulate different imaging conditions. These augmentation strategies collectively aimed to improve the model’s generalization capability and address class imbalance in the CBIS-DDSM dataset. 3.3 SE-Attention U-Net: An Enhanced Architecture with Dual Attention Mechanisms According the foundational U-Net architecture [ 7 ], we propose an advanced U-Net variant with dual attention mechanisms and a specialized loss function for mammographic mass segmentation through three key innovations. As shown in Fig. 5 , our encoder pathway incorporates residual convolutional blocks paired with squeeze-and-excitation (SE) modules, enabling both local feature extraction and global channel-wise feature recalibration. Each SE block dynamically emphasizes diagnostically relevant features while suppressing less informative channels through learned excitation weights. Spatial attention gates are integrated at skip connections between the encoder and decoder, which selectively emphasize mass-containing regions while suppressing irrelevant background areas, particularly crucial for maintaining boundary precision in sub-15mm lesions. The decoder pathway utilizes transposed convolutions with halved channel depth at each up-sampling stage, systematically recovering spatial resolution while integrating attention-weighted features from corresponding encoder levels. A strategic 0.5 dropout rate enhances generalization capability without compromising feature retention. Throughout the architecture, we maintain dimensional consistency through symmetric padding, ensuring the final segmentation map precisely aligns with input mammogram dimensions. This comprehensive design achieves three critical objectives: (1) preservation of fine structural details through residual learning and attention mechanisms, (2) adaptive feature enhancement via SE blocks, and (3) computationally efficient processing through optimized channel depth reduction in the decoder pathway. The complete system, trained with our novel Stable Extreme Loss Function. 3.4 Training Protocol with Novel Stable Extreme Loss Function Our training process employs a carefully designed three-phase strategy to optimize performance for mammographic mass segmentation. In the initial phase, we utilize standard Dice loss to establish baseline feature extraction capabilities, allowing the network to learn fundamental segmentation patterns. The model then transitions to our proposed Stable Extreme Loss Function (SELF) in the second phase, which specifically addresses three critical challenges. In the final phase, we employ transfer learning by freezing the encoder layers while fine-tuning the decoder components with SELF. This strategic approach preserved the learned hierarchical feature representations while optimizing the segmentation-specific architecture. This comprehensive training approach, combined with Adam optimization (β₁=0.9, β₂=0.999) and early stopping based on validation metrics, enables robust segmentation performance across all mass sizes. Novel Stable Extreme Loss Function (SELF) We propose a novel loss function to address three fundamental challenges in small mass mammographic segmentation: (1) extreme class imbalance where foreground pixels constitute less than 2% of the image area, (2) precise boundary localization for sub-15mm mass, and (3) numerical stability during optimization. SELF combines three strategically weighted components: a boundary-aware Dice loss (60%), mass-weighted focal loss (30%), and edge consistency loss (10%) - each targeting distinct aspects of the segmentation problem. Boundary-Aware Dice Loss Boundary-aware Dice loss is a novel technique developed to improve segmentation performance in medical imaging by overcoming the shortcomings of standard Dice loss[ 34 – 36 ]. This approach focuses on precisely delineating boundaries. It assigns higher weights to pixels located near the boundaries of organs or lesions. Our boundary-aware Dice loss (Eq. 1 ) extends the conventional formulation by integrating three key innovations to improve margin delineation. $$\:{L}_{B-Dice}=1-\frac{2.\sum\:_{i=1}^{N}{w}_{i}.\left({y}_{i}{\widehat{y}}_{i}\right)+ϵ}{\sum\:_{i=1}^{N}{w}_{i}.\left({y}_{i}+{\widehat{y}}_{i}\right)+ϵ}+\lambda\:{‖{\nabla\:}^{2}y-{\nabla\:}^{2}\widehat{y}‖}_{2}$$ 1 First, a 3×3 Laplacian kernel explicitly enhances sensitivity to mass boundaries by amplifying gradient signals at edge voxels (Eq. 2 ): $$\:{w}_{i}=1+\alpha\:.\left|{Laplacian\left(y\right)}_{i}\right|$$ 2 \(\:Laplacian\left(y\right)\) \(\:3\times\:3\) edge detection kernel output \(\:\alpha\:=5.0\) Empirically determined edge emphasis factor Second, dynamic foreground weighting adjusts pixel-wise (Eq. 3 ) contributions based on batch-specific mass prevalence, preventing small-mass features from being overwhelmed by background dominance: $$\:{w}_{i}\leftarrow\:{w}_{i}.\frac{\beta\:}{{\text{B}\text{a}\text{t}\text{c}\text{h}}_{\text{p}\text{r}\text{e}\text{v}\text{a}\text{l}\text{e}\text{n}\text{c}\text{e}}+ϵ}\:\:\:\:\:\:\:\:\text{w}\text{h}\text{e}\text{r}\text{e}\:\:\:\:{\text{B}\text{a}\text{t}\text{c}\text{h}}_{\text{p}\text{r}\text{e}\text{v}\text{a}\text{l}\text{e}\text{n}\text{c}\text{e}}=\frac{{N}_{fg}}{{N}_{total}}$$ 3 \(\:\beta\:\) =0.1: Normalization constant \(\:{N}_{fg}\) Foreground pixels in batch \(\:{N}_{total}\) Total pixels in batch Third, a second-order gradient term (Eq. 4 ) enforces geometric consistency, regularizing irregular contour formations. This yields improvement in margin sharpness compared to standard implementations, particularly crucial for sub-15mm lesions where boundary precision directly impacts diagnostic accuracy. $$\:\lambda\:{‖{\nabla\:}^{2}y-{\nabla\:}^{2}\widehat{y}‖}_{2}$$ 4 \(\:{\nabla\:}^{2}\) Laplacian operator for contour smoothness \(\:\lambda\:\) Geometric consistency weight \(\:ϵ={10}^{-7}\) Numerical stability constant This formulation improves margin sharpness for sub-15mm lesions. Mass-Weighted Focal Loss Mass-weighted focal loss represents an enhanced variant of focal loss[ 37 ], designed to address class imbalance in various medical imaging tasks[ 31 , 38 ]. To address extreme class imbalance (< 2% foreground), we augment the focal loss with two stabilization mechanisms (Eq. 5 , 6 ): 1. Adaptive Class Weighting It dynamically scales foreground loss contributions based on real-time batch statistics to counteract extreme class imbalance The foreground weighting factor \(\:{w}_{fg}\:\) ​ dynamically adjusts to batch-specific class distributions: $$\:{w}_{fg}={min}\left(50,\frac{1}{{\text{B}\text{a}\text{t}\text{c}\text{h}}_{\text{p}\text{r}\text{e}\text{v}\text{a}\text{l}\text{e}\text{n}\text{c}\text{e}}+ϵ}\right)$$ 5 \(\:ϵ={10}^{-7}\) : Ensures numerical stability for mass-free batches Cap (50×): Prevents dominance of ultra-rare masses (e.g., batches with < 0.02% prevalence) This adaptation ensures proportional gradient contributions from small masses while maintaining optimization stability. 2. Stabilized Parameterization We modify the focal loss formulation with three key stabilizers ( \(\:{w}_{fg},\:\:\gamma\:,\:\:{p}_{t})\) : $$\:{L}_{focal}={w}_{fg}.\left[-{\left(1-{p}_{t}\right)}^{\gamma\:}\text{log}\left({p}_{t}\right)\right],\:\:\:\:\:\:\:where\:\:\:{p}_{t}=clip\:(p,ϵ,1-ϵ)$$ 6 We reduce the focusing parameter to γ = 3.0 (vs. standard γ = 4.0) to soften gradients for tiny masses, while double clipping probability estimates \(\:(p\in\:\left[{10}^{-6},1-{10}^{-6}\right])\) guarantees numerical stability during backpropagation. Edge Consistency Loss The edge consistency loss enhances boundary precision by enforcing alignment between predicted and ground-truth mass margins[ 39 ]. We compute the L1 distance between edges extracted via 3×3 Laplacian filtering from both the segmentation output ( \(\:\widehat{E}\) ) and ground truth ( \(\:E\) ) as shown in Eq. 7: \(\:{L}_{edg}=\frac{1}{N}\sum\:_{i=1}^{N}\left|{\widehat{E}}_{i}-{E}_{i}\right|,\:\:\:\:\:\:\:\:where\:\:\:\widehat{E}=Laplacian\:({y}_{pred}),\:\:\:\:\:E=Laplacian({y}_{true}\) ) (7) This term addresses three critical clinical requirements: Margin Sharpness, reduces false positives at mass boundaries, and maintains topological consistency in irregular masses. 3.5 Evaluation Metrics To comprehensively evaluate the proposed model for medical image segmentation and classification, we employ metrics assessing both pixel-level classification accuracy and region-wise segmentation overlap. These metrics address class imbalance and spatial delineation challenges inherent to medical datasets (e.g., tumors occupying small regions). Classification Metrics These metrics evaluate the model’s global ability to correctly classify pixels, emphasizing robustness to class imbalance: 1. Accuracy Accuracy provides an overall measure of how well the model correctly classifies pixels in the image, combining correct positive and negative predictions. It is intuitive but can be misleading when class distributions are imbalanced, especially in medical images where abnormalities may occupy a small region. It is defined as Eq. 8 : $$\:Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$ 8 Where: TP (True Positives): Pixels correctly identified as part of the abnormality (e.g., tumor). TN (True Negatives): Pixels correctly identified as normal/background. FP (False Positives): Normal pixels incorrectly classified as abnormal. FN (False Negatives): Abnormal pixels incorrectly classified as normal. 2. Precision Precision indicates the proportion of positive identifications that were actually correct. It focuses on the reliability of positive detections. High precision means fewer false positives, which is critical to avoid unnecessary further procedures. It is calculated as Eq. 9 : $$\:Precision=\frac{TP}{TP+FP}$$ 9 3. Sensitivity (Recall) Sensitivity measures the model's ability to correctly identify all actual abnormal pixels. A high recall is essential to minimize missed detections, reducing the risk of overlooking disease. It is computed as Eq. 10 : $$\:Sensitivity=\frac{TP}{TP+FN}$$ 10 4. F1 Score The F1 Score (Eq. 11 ) is the harmonic mean of precision and recall, offering a single metric that balances both false positives and false negatives. It is particularly useful when the data is imbalanced and when one needs to balance precision and recall. The F1-score is mathematically equivalent to the Dice coefficient (Segmentation Metrics) but is included here for consistency with general classification literature. A higher F1 score indicates a better balance of precision and recall: $$\:F1\:Score=2\times\:\:\:\frac{Precision\times\:Sensitivity}{Precision+Sensitivity}$$ 11 Segmentation Metrics These metrics quantify spatial overlap between predicted and ground-truth masks, emphasizing anatomical delineation: 1. Dice Coefficient The Dice Coefficient (Eq. 12 ) (also called Sørensen–Dice index) is specifically designed for measuring the overlap between the predicted and ground truth binary masks in image segmentation tasks. Dice coefficient values range from 0 (no overlap) to 1 (perfect overlap): $$\:Dice\:Coefficient=\:\frac{2\times\:\:|X\cap\:Y|}{\left|X\right|+\left|Y\right|}=\frac{2.\:TP}{2.\:TP+FP+FN}$$ 12 X (Ground Truth) : The manually annotated (true) segmentation mask. Y (Prediction) : The model's predicted segmentation mask. ∣X∩Y∣ (TP) : Pixels correctly predicted as part of the mass. ∣X∣ (Ground Truth Size) : Total positives in the true mask (TP + FN). ∣Y∣ (Prediction Size) : Total positives in the predicted mask (TP + FP). 2. Jaccard coefficient (Intersection over Union, IoU) The Jaccard coefficient, also known as Intersection over Union (IoU), quantifies the overlap between the predicted and ground truth masks as the ratio of their intersection to their union as shown in Eq. 13 . It ranges from 0 to 1, where 1 indicates perfect overlap. IoU is more stringent than the Dice score, rewarding higher precision in spatial overlap: $$\:Jaccard\:Coefficient=\frac{|X\cap\:Y|}{|X\cup\:Y|}=\frac{TP}{TP+FP+FN}$$ 13 4. Experimental Results In this section, we evaluate the performance of the proposed method against existing approaches using the widely recognized CBIS-DDSM dataset. Performance is assessed using standard evaluation metrics, and visual comparisons are provided to highlight key differences. The experiments utilized mammogram images from the CBIS-DDSM dataset, which were resized to 224×224 pixels and processed through an advanced preprocessing pipeline that included adaptive histogram equalization, denoising, and edge enhancement. The rigorous training and evaluation process, combined with the enhanced model architecture and preprocessing techniques, establish a robust solution for small mass segmentation in mammography—a critical step in accurate breast cancer detection. The dataset was randomly divided into training, validation, and testing sets, with 70% allocated for training. The remaining 30% was evenly split between validation and testing sets. The proposed algorithm was applied to the mass segmentation task, and the results are illustrated in Fig. 6 . To comprehensively evaluate the performance of the proposed approach, we present results of several relevant metrics. Figure 7 illustrates the classification-related metrics, including accuracy, precision, sensitivity, and F1-score. These metrics provide insights into the model's ability to correctly identify and classify lesions, reflecting its robustness and reliability. To illustrate the performance of the proposed segmentation model, we present both the Dice coefficient and Jaccard coefficient for the training and validation phases in Fig. 8 . We compare the performance of our method with several state-of-the-art approaches in terms of accuracy, precision, sensitivity, F1-score, Dice coefficient and Jaccard coefficient as detailed in Table 1 and Table 2 . Table 1 Comparison of the proposed method with existing methods in terms of accuracy, precision, sensitivity and F1-score metrics on CBIS-DSSM dataset. Methods Sensitivity (%) Accuracy (%) Precision (%) F1-score (%) Sun et al. [ 40 ] 84.90 - - - Hou et al. [ 41 ] 85.60 - - - Ramesh et al.[ 42 ] 79 - - - El-Banby et al.[ 20 ] 90.58 - - 87.98 Khan et al. [ 43 ] 75.41 69.98 - - Duggento et al. [ 44 ] 84.40 71.19 - - Rajalakshmi et al. [ 42 ] 84.1 - - - Liao and Aagaard [ 45 ] (ResNet-50 (448×448)) 62.96 - 68.00 65.38 Oza et al. [ 46 ] - 91.96 - - Iqbal and Sharif [ 47 ] - - - 77.10 Our proposed model 97.08 99.54 97.97 97.53 Table 2 Comparison of the proposed and state-of-the-art methods in terms of Dice coefficient and Jaccard coefficient on CBIS-DSSM dataset. Methods Dice coefficient (%) Jaccard coefficient (IoU) Sun et al. [ 40 ] 81.80 - Hou et al. [ 41 ] 86.30 - Ramesh et al.[ 42 ] 82.90 - El-Banby et al.[ 20 ] 87.98 - Zhang et al.[ 48 ] 58.10 41.96 Rajalakshmi et al. [ 42 ] 82.90 85.70 Baccouche et al. 89.52 80.02 Chen et al. [ 49 ] 82.16 - Tsochatzidis et al. 72.20 56.50 Fazilov et al. [ 50 ] 77.82 65.17 Jin et al. [ 51 ] 85.89 - Our proposed model Mean: 91.00 Max: 96.55 Mean: 86.01 Max: 93.36 The proposed model demonstrates superior performance across both classification and segmentation tasks on the CBIS-DSSM dataset, outperforming existing state-of-the-art methods in all evaluated metrics. For classification (Table 1 ), the model achieves an accuracy of 99.54%, precision of 97.97%, sensitivity of 97.08%, and F1-score of 97.53%. In segmentation (Table 2 ), the model achieves a Dice coefficient of 91.00% and Jaccard coefficient (IoU) of 86.01%, surpassing all competing methods. These results demonstrate that our model not only provides highly accurate classification predictions but also delivers precise segmentation masks, enabling reliable lesion localization alongside diagnostic decisions. The simultaneous excellence in both tasks suggests our approach effectively captures both global and local features in medical images, making it particularly valuable for clinical applications requiring comprehensive analysis. 5. Conclusion and future directions Our study introduces a SE-Attention U-Net framework that significantly advances mammographic mass segmentation through three key innovations: (1) adaptive feature refinement with squeeze-and-excitation blocks, (2) attention-guided spatial prioritization of small lesions, and (3) a stable loss function optimized for severe class imbalance. Rigorous multi-level evaluation on the CBIS-DDSM dataset confirms the clinical viability of our model. At the pixel level, the model achieves a Dice coefficient of 0.966 and a Jaccard index of 0.914, demonstrating exceptional boundary accuracy—particularly useful for irregularly shaped masses and critical for surgical planning and radiotherapy. At the region level, the model attains 97.08% sensitivity and a 97.53% F1-score, indicating reliable lesion-wise diagnosis—crucial for reducing missed cancers during screening. Notably, with 99.54% accuracy and 97.97% precision, our approach exhibits robustness against false positives, effectively addressing a key limitation of existing CAD systems. For future work, we plan to explore the integration of multimodal imaging data, such as combining mammograms with ultrasound or MRI to enhance diagnostic accuracy. Incorporating radiomics features could further improve lesion characterization and aid in personalized treatment planning. Additionally, evaluating our framework on diverse datasets beyond CBIS-DDSM will help verify its generalizability across different populations and imaging protocols. Furthermore, extending our approach to process 3D mammographic or tomographic images may provide more comprehensive spatial context and improve segmentation precision, especially for complex lesions. Declarations Competing Interests The author declares no competing interests. Author Contribution The author M.H. conceptualized the study, designed the methodology, performed the analysis, wrote the manuscript, and prepared all figures/tables. Data Availability The dataset used in this study is the publicly available CBIS-DDSM (Curated Breast Imaging Subset of the Digital Database for Screening Mammography). Access to the CBIS-DDSM dataset can be obtained from https://www.kaggle.com/datasets/awsaf49/cbis-ddsm-breast-cancer-image-dataset . All relevant data supporting the findings of this study are available from the author, Maliheh Habibi, at [email protected] , upon reasonable request. References Zhang, Y. et al. From single to universal: tiny lesion detection in medical imaging. Artif. Intell. Rev. 57 (8), 192 (2024). Wang, J., Wang, S. & Zhang, Y. Deep learning on medical image analysis. CAAI Trans. Intell. Technol. 10 (1), 1–35 (2025). Esteva, A. et al. Dermatologist-level classification of skin cancer with deep neural networks. nature 542 (7639), 115–118 (2017). Litjens, G. et al. A survey on deep learning in medical image analysis. Med. Image. Anal. 42 , 60–88 (2017). Abdulwahhab, A. H. et al. A review on medical image applications based on deep learning techniques. J. Image Graphics . 12 (3), 215–227 (2024). Oluseyi, O. M., Udofot, A. I. & Bassey, E. Enhancing Medical Image Diagnosis Using Convolutional Neural Network and Transfer Learning. Ronneberger, O., Fischer, P. & Brox, T. U-net: Convolutional networks for biomedical image segmentation . in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5–9 , proceedings, part III 18. 2015. Springer. (2015). Avazov, K. et al. Dynamic Focus on Tumor Boundaries: A Lightweight U-Net for MRI Brain Tumor Segmentation. Bioengineering 11 (12), 1302 (2024). Salem, N., Malik, H. & Shams, A. Medical image enhancement based on histogram algorithms. Procedia Comput. Sci. 163 , 300–311 (2019). Patel, S. et al. Comparative study on histogram equalization techniques for medical image enhancement . in Soft Computing for Problem Solving: SocProS 2018, Volume 1 . Springer. (2020). Pizer, S. M. et al. Adaptive histogram equalization for automatic contrast enhancement of medical images. in Application of Optical Instrumentation in Medicine XIV and Picture Archiving and Communication Systems (Spie, 1986). Vidyasaraswathi, H. & Hanumantharaju, M. Review of various histogram based medical image enhancement techniques . in Proceedings of the International Conference on Advanced Research in Computer Science Engineering & Technology (ICARCSET 2015) . 2015. 2015. (2015). Haddadi, Y. R., Mansouri, B. & Khodja, F. Z. I. A novel medical image enhancement algorithm based on CLAHE and pelican optimization . Multimedia Tools Appl. , : pp. 1–20. (2024). Li, L., Si, Y. & Jia, Z. Medical image enhancement based on CLAHE and unsharp masking in NSCT domain. J. Med. Imaging Health Inf. 8 (3), 431–438 (2018). Sharma, R. & Kamra, A. A review on CLAHE based enhancement techniques . in 6th International Conference on Contemporary Computing and Informatics (IC3I) . 2023. IEEE. 2023. IEEE. (2023). Yan, L. J. et al. A hybrid deep learning model for mammographic breast cancer detection: Multi-autoencoder and attention mechanisms. J. Radiation Res. Appl. Sci. 18 (3), 101578 (2025). Hu, J., Shen, L. & Sun, G. Squeeze-and-excitation networks . in Proceedings of the IEEE conference on computer vision and pattern recognition . (2018). Woo, S. et al. Cbam: Convolutional block attention module . in Proceedings of the European conference on computer vision (ECCV) . (2018). Zuiderveld, K. J. Contrast limited adaptive histogram equalization. Graphics gems . 4 (1), 474–485 (1994). El-Banby, G. M. et al. Automated abnormalities detection in mammography using deep learning. Complex. Intell. Syst. 10 (5), 7279–7295 (2024). Islam, M. R. et al. Enhancing Semantic Segmentation with Adaptive Focal Loss: A Novel Approach. arXiv preprint arXiv:2407.09828, (2024). Rahman, M. A. & Wang, Y. Optimizing Intersection-Over-Union in Deep Neural Networks for Image Segmentation (Springer International Publishing, 2016). Buades, A., Coll, B. & Morel, J. M. A non-local algorithm for image denoising . in IEEE computer society conference on computer vision and pattern recognition (CVPR'05) . 2005. Ieee. 2005. Ieee. (2005). Kumar, M. & Diwakar, M. Edge preservation based CT image denoising using wavelet and curvelet transforms . in Proceedings of Fifth International Conference on Soft Computing for Problem Solving: SocProS 2015, Volume 1 . Springer. (2016). Wang, Y. & Zhou, H. Total Variation Wavelet-Based Medical Image Denoising. Int. J. Biomed. Imaging . 2006 (1), 089095 (2006). Lei, D. et al. A spatiotemporal fusion method based on multiscale feature extraction and spatial channel attention mechanism. Remote Sens. 14 (3), 461 (2022). Si, Y. et al. SCSA: Exploring the synergistic effects between spatial and channel attention. Neurocomputing 634 , 129866 (2025). Abbasi, S. et al. Improved A-Line and B-Line Detection in Lung Ultrasound Using Deep Learning with Boundary-Aware Dice Loss. Bioengineering 12 (3), 311 (2025). Ngoc, M. Ô. V. et al. Introducing the Boundary-Aware loss for deep image segmentation . in British Machine Vision Conference (BMVC) 2021 . (2021). Mahmoodi, N. et al. Automatically weighted focal loss for imbalance learning. Neural Comput. Appl. 37 (5), 4035–4052 (2025). Qin, R. et al. Weighted focal loss: An effective loss function to overcome unbalance problem of chest X-ray14 . in IOP conference series: Materials science and engineering . IOP Publishing. (2018). Lee, R. S. et al. A curated mammography data set for use in computer-aided detection and diagnosis research. Sci. data . 4 (1), 1–9 (2017). Clark, K. et al. The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository. J. Digit. Imaging . 26 , 1045–1057 (2013). Kim, M. & Lee, B. D. A simple generic method for effective boundary extraction in medical image segmentation. IEEE Access. 9 , 103875–103884 (2021). Zhang, Y. et al. A 3D + 2D CNN approach incorporating boundary loss for stroke lesion segmentation . in Machine Learning in Medical Imaging: 11th International Workshop, MLMI 2020, Held in Conjunction with MICCAI 2020, Lima, Peru, October 4, 2020, Proceedings 11 . Springer. (2020). Shi, P. et al. Centerline boundary dice loss for vascular segmentation . in International Conference on Medical Image Computing and Computer-Assisted Intervention . Springer. (2024). Seth, A. The Application of Focal Loss in various Domains: A Survey . Int. J. Res. Appl. Sci. Eng. Technol. (IJRASET) , (2321–9653 ). (2020). Aliniya, P. et al. Improved loss function for mass segmentation in mammography images using density and mass size. J. Imaging . 10 (1), 20 (2024). Qiu, Y. et al. Semi-ECNet: Edge-Consistency Based Semi-Supervised Retinal Vessel Segmentation Network . in 2024 IEEE International Symposium on Biomedical Imaging (ISBI) . IEEE. (2024). Sun, H. et al. AUNet: attention-guided dense-upsampling networks for breast mass segmentation in whole mammograms. Phys. Med. Biol. 65 (5), 055005 (2020). Hou, X. et al. Mass segmentation for whole mammograms via attentive multi-task learning framework. Phys. Med. Biol. 66 (10), 105015 (2021). Rajalakshmi, N. R. et al. Deeply supervised u-net for mass segmentation in digital mammograms. Int. J. Imaging Syst. Technol. 31 (1), 59–71 (2021). Khan, H. N. et al. Multi-view feature fusion based four views model for mammogram classification using convolutional neural network. IEEE Access. 7 , 165724–165733 (2019). Duggento, A. et al. An ad hoc random initialization deep neural network architecture for discriminating malignant breast cancer lesions in mammographic images 2019p. 5982834 (Contrast media & molecular imaging, 2019). 1. Liao, L. & Aagaard, E. M. An open codebase for enhancing transparency in deep learning-based breast cancer diagnosis utilizing CBIS-DDSM data. Sci. Rep. 14 (1), 27318 (2024). Oza, P. R., Sharma, P. & Patel, S. A transfer representation learning approach for breast cancer diagnosis from mammograms using efficientnet models. Scalable Computing: Pract. Experience . 23 (2), 51–58 (2022). Iqbal, A. & Sharif, M. BTS-ST: Swin transformer network for segmentation and classification of multimodality breast cancer images. Knowl. Based Syst. 267 , 110393 (2023). Zhang, B. et al. A method for breast mass segmentation using image augmentation with SAM and receptive field expansion . in Proceedings of the 12th International Conference on Computing and Pattern Recognition . 2023. 2023. (2023). Chen, J. et al. A novel multi-scale adversarial networks for precise segmentation of x-ray breast mass. IEEE access. 8 , 103772–103781 (2020). Fazilov, S. et al. Segmentation in mammographic images based on the U-NET model . in AIP Conference Proceedings . AIP Publishing. (2024). Jin, S. et al. A Breast Mass Image Segmentation Method Based on Improved UNet 3 + Network . in IECON -49th Annual Conference of the IEEE Industrial Electronics Society . 2023. IEEE. 2023. IEEE. (2023). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 08 Jan, 2026 Reviews received at journal 04 Jan, 2026 Reviews received at journal 17 Oct, 2025 Reviewers agreed at journal 09 Oct, 2025 Reviewers agreed at journal 07 Oct, 2025 Reviewers invited by journal 07 Oct, 2025 Editor assigned by journal 15 Sep, 2025 Editor invited by journal 13 Jun, 2025 Submission checks completed at journal 12 Jun, 2025 First submitted to journal 02 Jun, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6800663","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":531398713,"identity":"2eccf423-09f3-434e-bb1e-ee3fb1daaa70","order_by":0,"name":"Maliheh Habibi","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+UlEQVRIiWNgGAWjYFAC5obDDAwHGAwYeBgOMxjYwIQl8GhhRNGSRpwWZpgWIOMwYWfpth9sPFxQcYfBnP3swcMFBefzzCUSGD/8YLDIx6XF7Exiw+EZZ54xWPbkJRyeYXC72HJGArNkD4OEZQMuLQeAWnjbgL44kGNwmMfgduKGGwkM0kC/GOC05fxDoJZ/QC3n34C0nANpYf6NV8sNkC3AQDO4AbblAEgLG35bbgBt4TkG0vIO5JfkYoMzD9ssewzwOSz58GeeGpDDcg9/Lvhjl2dwPPnwjR8VdTi1wEB9A5SRAIopBgaCGpBAAglqR8EoGAWjYIQAAD/qYLqbe+ioAAAAAElFTkSuQmCC","orcid":"","institution":"Payame Noor University","correspondingAuthor":true,"prefix":"","firstName":"Maliheh","middleName":"","lastName":"Habibi","suffix":""}],"badges":[],"createdAt":"2025-06-02 09:23:28","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6800663/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6800663/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":93914382,"identity":"e65663fa-7041-401c-945d-e7cbc5f0c7b5","added_by":"auto","created_at":"2025-10-20 08:39:03","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1521459,"visible":true,"origin":"","legend":"","description":"","filename":"SmallLesionSegmentationinMammographyFinalV2.docx","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/c67482436938ea2c4e18c4af.docx"},{"id":93914447,"identity":"34ad20b8-3b2c-4613-a9f3-3119f7183778","added_by":"auto","created_at":"2025-10-20 08:39:10","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3707,"visible":true,"origin":"","legend":"","description":"","filename":"9f0795514e2649a4a4290170d7a611e6.json","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/a1d33d41b04a0ea13a28d3c9.json"},{"id":93917015,"identity":"e7b6a9b4-b12f-4dbc-ba24-554629ea9737","added_by":"auto","created_at":"2025-10-20 08:55:10","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":118763,"visible":true,"origin":"","legend":"","description":"","filename":"9f0795514e2649a4a4290170d7a611e61enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/0e47c834d5d374c171684c0e.xml"},{"id":93914412,"identity":"581fc9f3-cde8-4267-ad83-24907db9ebb7","added_by":"auto","created_at":"2025-10-20 08:39:08","extension":"jpeg","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":407362,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/87bc17c6c471d4b1e3be24c8.jpeg"},{"id":93914416,"identity":"ed1866f0-7bd5-447e-801f-9319119afa38","added_by":"auto","created_at":"2025-10-20 08:39:09","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":871,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/72594b126dc7b49fbbb44616.png"},{"id":93914407,"identity":"3589a3f1-4834-4e5d-93a7-470303d16c40","added_by":"auto","created_at":"2025-10-20 08:39:07","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":315836,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage11.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/d6eb88276cb91bcb40c268de.jpeg"},{"id":93914391,"identity":"6fb1d0fd-ccb4-4157-b63b-e637d86c92f1","added_by":"auto","created_at":"2025-10-20 08:39:05","extension":"jpeg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":478801,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage12.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/1140f94515709e039d850a51.jpeg"},{"id":93914419,"identity":"68b6b616-f5c7-4c83-b528-2321c2c8efe6","added_by":"auto","created_at":"2025-10-20 08:39:09","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":213011,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/a5b06caa3e387f2a75e162bc.png"},{"id":93914400,"identity":"448d09da-b286-49fa-8bce-d95d2e09b975","added_by":"auto","created_at":"2025-10-20 08:39:06","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":30791,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/bd4557574249f379808d6f8f.png"},{"id":93915874,"identity":"dbcca588-55b6-4ffe-a0b8-7c5d8f0fd8b6","added_by":"auto","created_at":"2025-10-20 08:47:09","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":65365,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/2404986cefc6aabc88e9d386.png"},{"id":93914448,"identity":"c9b70041-1bad-4e49-a993-fae07080bb50","added_by":"auto","created_at":"2025-10-20 08:39:10","extension":"jpeg","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":177547,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/7cd2102b9b2659570da2534c.jpeg"},{"id":93914437,"identity":"ed4e3cf0-77f9-47f4-9211-ff6d31610570","added_by":"auto","created_at":"2025-10-20 08:39:10","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":20120,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/4b815500ce71d5e5e7aeb99c.png"},{"id":93914410,"identity":"f016f4bc-5493-4a8e-bda0-45303f8dca08","added_by":"auto","created_at":"2025-10-20 08:39:08","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16239,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/75641c4ddf83adab978b9da4.png"},{"id":93915877,"identity":"8dada6a0-84fd-4076-a6b6-20943a932674","added_by":"auto","created_at":"2025-10-20 08:47:10","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":22877,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/5ccac1bcf6e5aa19d3677d57.png"},{"id":93914404,"identity":"2f946d57-84ac-4749-9e53-0e824cbc4414","added_by":"auto","created_at":"2025-10-20 08:39:06","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":23797,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/d73413435b871152c036cb26.png"},{"id":93915880,"identity":"199d35b6-c72d-42f2-8513-0ea7bea473e2","added_by":"auto","created_at":"2025-10-20 08:47:10","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":853,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/fa9490eae7dfbb5784e536f3.png"},{"id":93915873,"identity":"6d7b2d58-a8ac-40e1-8d84-5364d4efea2b","added_by":"auto","created_at":"2025-10-20 08:47:09","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":854,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/185fd9b647c580d7d1d82e27.png"},{"id":93915871,"identity":"9d86d2d4-3516-40e3-98ba-c3ae6c40399f","added_by":"auto","created_at":"2025-10-20 08:47:06","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":874,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/99b0f6e36f8a92396caece14.png"},{"id":93914381,"identity":"fefca49e-0c5e-48b9-a116-beb17126a167","added_by":"auto","created_at":"2025-10-20 08:39:03","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":66801,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/fe147330981fd0f3cced4096.png"},{"id":93915872,"identity":"aca3e6e3-ffee-43ad-9a89-a90a9a393bd3","added_by":"auto","created_at":"2025-10-20 08:47:06","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":509,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/afc192cf93888eabe19f76e7.png"},{"id":93914397,"identity":"195a8b09-950d-46ba-b95e-3c3f03ffbdc2","added_by":"auto","created_at":"2025-10-20 08:39:06","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":527570,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/f3512c2d0c14134ad8c3d5e6.png"},{"id":93914411,"identity":"358f0a1f-d251-4c7c-a1e1-3fbad2576f3f","added_by":"auto","created_at":"2025-10-20 08:39:08","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":116335,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/76053b48b87c38bdee7d0133.png"},{"id":93914439,"identity":"f2af982e-b3ed-466c-a978-27e383c418fb","added_by":"auto","created_at":"2025-10-20 08:39:10","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":43130,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/698fdb99c18b9fae776edb1c.png"},{"id":93914403,"identity":"7b74bab4-fe11-4559-a5b3-cd934fecdb1e","added_by":"auto","created_at":"2025-10-20 08:39:06","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8150,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/d1d2cfb9d7e6549310c7b657.png"},{"id":93914434,"identity":"83488ac5-ea58-4bb9-a810-a742af778622","added_by":"auto","created_at":"2025-10-20 08:39:09","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":20781,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/23ac07f0edb8950f787ceb37.png"},{"id":93914435,"identity":"e0861a53-4777-45e5-b29c-6406c3679744","added_by":"auto","created_at":"2025-10-20 08:39:09","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":234288,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/62240adf86daee174a0c072a.png"},{"id":93914406,"identity":"09fcf19d-d688-40e2-9c54-f1e7582e24ef","added_by":"auto","created_at":"2025-10-20 08:39:07","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16959,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/1f18a5bd37e5af3709c1af40.png"},{"id":93914393,"identity":"111d709c-e6a2-4b6c-896c-2213bfb41c87","added_by":"auto","created_at":"2025-10-20 08:39:05","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":13821,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/563036897fbafb9e75b1a430.png"},{"id":93914408,"identity":"555eb139-6bb0-4a79-a74c-b0174968f385","added_by":"auto","created_at":"2025-10-20 08:39:07","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":20695,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/503d9759a00a974d79b86899.png"},{"id":93915875,"identity":"2206ff35-1150-47b7-81c9-74fb80c99024","added_by":"auto","created_at":"2025-10-20 08:47:10","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":22899,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/9cf6495f0f791141a7a12a74.png"},{"id":93914414,"identity":"96b5f7ee-3dc2-4577-a92a-929ae0440cf9","added_by":"auto","created_at":"2025-10-20 08:39:09","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":498,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/4ff37eac8a1c958a0f6dc897.png"},{"id":93917014,"identity":"a5293179-ac17-45d1-b862-45bf4f75ef81","added_by":"auto","created_at":"2025-10-20 08:55:10","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":503,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/3c0fdeceefff68990664d910.png"},{"id":93914441,"identity":"7780df74-c20f-42b6-8eba-3b7b24f08a10","added_by":"auto","created_at":"2025-10-20 08:39:10","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":506,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/6a385832aa10b186e0bb34ce.png"},{"id":93914445,"identity":"3c1a4b3a-b623-49d8-a3b3-5bca0e692e40","added_by":"auto","created_at":"2025-10-20 08:39:10","extension":"xml","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":118367,"visible":true,"origin":"","legend":"","description":"","filename":"9f0795514e2649a4a4290170d7a611e61structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/b2dde7b9e808dbf6f89fe623.xml"},{"id":93914399,"identity":"a29e1140-756d-4a9d-97ef-55f55b0a4a9f","added_by":"auto","created_at":"2025-10-20 08:39:06","extension":"html","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":131008,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/4db4f76c881eb6bfb27904b5.html"},{"id":93914436,"identity":"b93d40b4-8883-4f05-a5ac-eb7779f2e1ee","added_by":"auto","created_at":"2025-10-20 08:39:09","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":114602,"visible":true,"origin":"","legend":"\u003cp\u003eThe proposed pipeline for small Lesion segmentation\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/b56f5c14434ba297e7dc6d0d.jpg"},{"id":93914396,"identity":"00c0db26-30e9-4bae-8c18-5327adf48781","added_by":"auto","created_at":"2025-10-20 08:39:06","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":52073,"visible":true,"origin":"","legend":"\u003cp\u003eCalcifications in CC and MLO Views with Radiologist-Annotated ROI.\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/1ac193d0515df4a0a2dceeef.jpg"},{"id":93914413,"identity":"4bcf6b6d-d9c3-4d42-9e97-56f4b4b14d9c","added_by":"auto","created_at":"2025-10-20 08:39:09","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":50003,"visible":true,"origin":"","legend":"\u003cp\u003eMass in CC and MLO Views with Radiologist-Annotated ROI.\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/40e638c15dcc127d69c0d4ec.jpg"},{"id":93914417,"identity":"7ce8d2f6-fda0-41fe-b29d-a7d5ee389896","added_by":"auto","created_at":"2025-10-20 08:39:09","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":78284,"visible":true,"origin":"","legend":"\u003cp\u003ePreprocessing pipeline output: Comparison of original (top) and enhanced (bottom) mammograms from CBIS-DDSM.\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/7c758be7a0360f7dce4ea074.jpg"},{"id":93915878,"identity":"2e18d95b-025d-4ead-a80a-5b098910e5f3","added_by":"auto","created_at":"2025-10-20 08:47:10","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":137986,"visible":true,"origin":"","legend":"\u003cp\u003eEnhanced U-Net Architecture with Attention Mechanisms and SE lock\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/e70ee4d0dfd9b24c78ad7f8b.jpg"},{"id":93914451,"identity":"590dbcd4-cd0d-408c-8290-47bf958b9377","added_by":"auto","created_at":"2025-10-20 08:39:11","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":138244,"visible":true,"origin":"","legend":"\u003cp\u003eExamples of prediction in the CBIS-DDSM dataset. First column: original images; second column: ground truth ROI; and third column: predicted ROI by the proposed method (Dice and Jaccard score).\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/bce177f77a82cec320c0d5d0.jpg"},{"id":93914409,"identity":"e4cf78d9-7a03-44fd-a4cb-43ca8d5a7720","added_by":"auto","created_at":"2025-10-20 08:39:08","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":100361,"visible":true,"origin":"","legend":"\u003cp\u003eClassification performance metrics. Comparative results of the proposed model on mass detection, evaluated through accuracy, precision, sensitivity (recall), and F1-score.\u003c/p\u003e","description":"","filename":"7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/755c3d749f978d83f1bac4a0.jpg"},{"id":93914401,"identity":"3f0f210e-172c-40d7-8354-89f8ff293f51","added_by":"auto","created_at":"2025-10-20 08:39:06","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":63220,"visible":true,"origin":"","legend":"\u003cp\u003eSegmentation consistency analysis. Training/validation curves for Dice coefficient (left) and Jaccard coefficient (IoU) (right), highlighting the model's stability in preserving precise mass boundary characterization during model optimization.\u003c/p\u003e","description":"","filename":"8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/d987b5a94301d6d4cf539307.jpg"},{"id":93917016,"identity":"421ede3e-c80d-43ca-87a1-e5dd417e0d53","added_by":"auto","created_at":"2025-10-20 08:55:16","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1751754,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6800663/v1/200effa9-51e4-4d97-8eb1-a2c0238c2eee.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"SE-Attention U-Net: A Hybrid Loss-Optimized Model for Small Breast Lesion Segmentation in Mammography ","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eBreast cancer is a major global health concern and one of the leading causes of cancer-related mortality among women. Early detection through screening is vital for improving patient outcomes, with mammography serving as the primary diagnostic tool. However, accurately segmenting small masses in mammograms remains a significant challenge due to tissue overlap, image noise, and low contrast, which often lead to missed diagnoses or delayed treatment. The task is further complicated by the extreme class imbalance between the tumor (foreground) and healthy tissue (background) pixels, as well as the irregular and subtle appearance of small masses. Despite advancements in imaging technology, developing reliable methods for precise delineation of these tiny anomalies is essential to enhance early detection and improve clinical outcomes [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eRecent advancements in deep learning, especially convolutional neural networks (CNNs), hold great promise in tackling these challenges. Multiple studies [\u003cspan additionalcitationids=\"CR3 CR4 CR5\" citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] have demonstrated the potential of deep learning to significantly enhance medical image analysis. However, these approaches predominantly focus on larger masses and typically rely on extensive annotated datasets, which are often scarce in practical clinical environments. Even the U-Net architecture, widely adopted for biomedical segmentation [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], underperforms due to inherent tradeoffs between maintaining spatial information and resolving uncertain tumor interfaces or minor intensity variations [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eIn medical image analysis, enhancement techniques like Histogram Equalization (HE) and its variants\u0026mdash;such as Global HE (GHE)[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], Local HE (LHE) [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], Adaptive HE (AHE)[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], and Contrast-Limited Adaptive Histogram Equalization (CLAHE) [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan additionalcitationids=\"CR14\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u0026mdash;have been widely adopted. However, CLAHE has emerged as a preferred method due to its ability to enhance local contrast adaptively while constraining noise amplification through histogram clipping. This makes CLAHE indispensable for early breast cancer detection, where preserving fine details and avoiding artifacts are paramount.\u003c/p\u003e\u003cp\u003eDespite advances in CNN architectures, the integration of attention mechanisms for enhanced feature extraction in mammographic image analysis remains underdeveloped [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. While attention modules like Squeeze-and-Excitation (SE) blocks have proven effective in boosting segmentation accuracy by enabling channel-wise feature recalibration [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] and improving sensitivity to subtle features [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], their application to mammography - particularly for detecting small, irregular masses - has not been fully explored.\u003c/p\u003e\u003cp\u003eGiven the shortcomings of existing methods, this research aims to address these critical gaps through the proposal of an enhanced U-Net architecture that integrates several key components designed to improve the segmentation of small masses in mammograms. Our primary contributions include:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eEnhanced Preprocessing\u003c/b\u003e: A robust preprocessing pipeline that significantly improves image quality, utilizing adaptive histogram equalization [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] and Gaussian denoising techniques. This approach has been shown to effectively enhance contrast and reduce noise in medical imaging.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eHybrid Attention Architecture\u003c/b\u003e: The implementation of a hybrid attention architecture combining channel-wise squeeze-excitation blocks with spatial attention gates, specifically optimized for detecting masses smaller than 15 mm. This architecture has been proven to refine feature extraction and boost accuracy in challenging segmentation tasks [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e].\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eNovel Stable Extreme Loss Function\u003c/b\u003e: A dynamically weighted loss function that combines focal loss and adaptive Dice loss to address the significant class imbalances inherent in mammographic datasets[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. By effectively penalizing misclassifications of small masses, this loss function enhances the model's ability to segment small, irregularly shaped masses accurately.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eMedical Augmentation Pipeline\u003c/b\u003e: To improve model generalization and address data limitations, we developed a comprehensive augmentation approach combining lesion-specific and whole-image transformations. For mass regions, we applied carefully boundary-aware flips (horizontal/vertical), limited rotations (\u0026plusmn;\u0026thinsp;15\u0026deg;), and localized contrast enhancement. Whole-image augmentations incorporated random flips (50% probability) and controlled brightness variations (\u0026plusmn;\u0026thinsp;10) [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. This dual strategy enhances feature learning while preserving diagnostic relevance of mammographic findings.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eBy leveraging these innovative methodologies, we aim to establish a new standard for automated mammographic analysis, ultimately providing healthcare professionals with a more reliable and accurate tool for breast cancer detection and diagnosis. This research not only improves segmentation capabilities for small masses but also contributes to the broader field of medical imaging by addressing significant limitations present in existing models.\u003c/p\u003e"},{"header":"2. Related Works","content":"\u003cp\u003eIn recent years, the challenge of small mass segmentation in mammography has attracted significant attention, leading to numerous studies exploring various deep learning methodologies. However, despite the advancements made, substantial gaps remain in managing class imbalances, enhancing feature extraction, and generating realistic synthetic data suitable for model training.\u003c/p\u003e\u003cp\u003eEarly convolutional approaches demonstrated promise for lesion detection. Still, they were fundamentally limited in handling small masses under 15 mm due to resolution constraints [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. The introduction of U-Net architectures represented a pivotal development, leveraging innovative skip connections to preserve critical spatial details [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. However, these architectures continued to struggle with the extreme class imbalance that is inherent to mammographic datasets [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. In the realm of preprocessing, adaptive histogram equalization methods, such as CLAHE, became standard for contrast enhancement [\u003cspan additionalcitationids=\"CR14\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], while subsequent refinements incorporated tissue-specific parameter optimization that better preserved subtle lesion boundaries. Denoising techniques further evolved from basic non-local means filters [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] to more sophisticated hybrid approaches that combine wavelet transforms with edge preservation [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], although none specifically targeted the unique noise profiles typically found around small masses.\u003c/p\u003e\u003cp\u003eRecent advancements have emphasized the use of attention mechanisms in deep learning frameworks. Research by [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] revealed the benefits of squeeze-and-excitation blocks and spatial attention for refining feature extraction. Subsequent hybrid architectures, such as those presented by Lie et al. [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] and Si et al. [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e], demonstrated the complementary benefits of integrating channel and spatial attention mechanisms. However, adaptations specific to mammography remain underexplored. Our innovative hybrid attention architecture combines channel-wise squeeze-excitation blocks with spatial attention gates specifically optimized for detecting masses smaller than 15 mm. This integration significantly enhances segmentation accuracy in our model.\u003c/p\u003e\u003cp\u003eAdditionally, we present a novel Stable Extreme Loss function that dynamically adjusts weights by combining boundary-aware Dice loss [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e], mass-presence weighted focal loss [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], and edge consistency loss. This approach effectively addresses the extreme class imbalance encountered in mammographic datasets and improves segmentation performance.\u003c/p\u003e\u003cp\u003eBy tackling these critical issues, our work establishes a new standard for automated mammographic analysis, offering healthcare professionals a more reliable and accurate tool for breast cancer detection and diagnosis. Through enhanced preprocessing, advanced attention mechanisms, and finely-tuned loss functions, we believe our approach not only bridges existing gaps in the literature but also paves the way for future advances in medical image analysis.\u003c/p\u003e"},{"header":"3. Materials and methodology","content":"\u003cp\u003eThis section outlines the comprehensive methodology employed in our paper to enhance the segmentation of small masses in mammographic images. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, our approach incorporates advanced preprocessing techniques, a uniquely tailored network architecture, and robust training strategies aimed at improving the detection and segmentation of small masses in mammograms.\u003c/p\u003e\u003cp\u003eOur proposed method is evaluated using the CBIS-DDSM dataset [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. The Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM) is a publicly available, curated dataset designed to facilitate research in medical image analysis, particularly for breast cancer detection. It is an improved and standardized version of the original DDSM dataset. The CBIS-DDSM comprises 2,620 scanned film mammograms\u0026mdash;1,566 cases with calcifications and 1,054 with masses\u0026mdash;each with bilateral views (CC and MLO), as illustrated in Figs.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Every image is meticulously annotated by radiologists, including lesion type (calcification/mass), pathology labels (benign/malignant), bounding boxes, and pixel-wise segmentation masks. Calcifications tend to display fine-grained patterns (e.g., clustered or diffuse), while masses vary in morphology (e.g., spiculated or circumscribed), enabling robust training for both detection and characterization. The inclusion of paired CC (craniocaudal) and MLO (mediolateral oblique) views per case provides comprehensive spatial context, which is critical for reducing false positives and enhancing diagnostic accuracy. By leveraging CBIS-DDSM\u0026rsquo;s dual-pathology annotations and bilateral imaging, our work addresses segmentation challenges across both calcifications and masses, with particular focus on small lesions. Each mammogram is paired with pixel-level annotations (ROIs) for masses and calcifications, along with radiologist-assessed BI-RADS ratings, pathology labels (benign/malignant), and lesion diameters\u0026mdash;ranging from 3 to 30 mm, with most masses under 15 mm.\u003c/p\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e3.1 Image Pre-processing\u003c/h2\u003e\u003cp\u003eOur preprocessing pipeline combines normalization, Adaptive Histogram Equalization (AHE) [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e], Gaussian denoising, and edge sharpening to address key challenges in mammographic imaging, including low contrast, noise artifacts, and soft tissue ambiguity. By adaptively enhancing contrast, reducing noise, and accentuating structural details, this approach ensures high-quality input for downstream deep learning tasks, thereby improving feature extraction and segmentation accuracy. The four-stage workflow (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) effectively optimizes image quality, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cb\u003eNormalization\u003c/b\u003e\u003c/p\u003e\u003cp\u003eMammographic images often exhibit inter-scanner intensity variations due to differences in acquisition protocols, equipment, or digitization processes. To standardize input data and mitigate these inconsistencies, we apply min-max normalization, rescaling each image\u0026rsquo;s pixel intensities to a fixed range of [0, 255]. This range aligns with standard 8-bit grayscale representations, ensuring compatibility with conventional deep learning frameworks.\u003c/p\u003e\u003cp\u003e\u003cb\u003eAdaptive Contrast Enhancement\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo enhance the visibility of subtle lesions obscured by dense breast tissue, we applied Contrast Limited Adaptive Histogram Equalization (CLAHE) [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. This method improves local contrast by dividing the image into non-overlapping 16\u0026times;16 tile grids and performing histogram equalization within each region, constrained by a clip limit of 3.0 to prevent noise amplification. Unlike global histogram equalization, CLAHE adapts to local intensity variations, making it particularly effective for mammograms where small masses or microcalcifications may exhibit low contrast against heterogeneous backgrounds. The clip limit ensures balanced enhancement by redistributing only the histogram bins exceeding the threshold, thereby preserving anatomical details while mitigating artificial artifacts.\u003c/p\u003e\u003cp\u003e\u003cb\u003eStructure-Preserving Denoising\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo enhance the signal-to-noise ratio while preserving critical anatomical structures, we implemented Non-Local Means (NLM) [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] Denoising with optimized parameters (decay parameter h\u0026thinsp;=\u0026thinsp;20, search window size\u0026thinsp;=\u0026thinsp;21\u0026times;21 pixels). The selected parameters were carefully tuned to achieve an optimal balance between noise suppression and preservation of subtle pathological features, such as spiculated margins or microcalcifications. The NLM algorithm's effectiveness stems from its ability to maintain structural integrity while reducing noise.\u003c/p\u003e\u003cp\u003e\u003cb\u003eEdge Sharpening\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo improve the delineation of lesion boundaries and subtle mass margins critical for accurate diagnosis, we implemented a conservative edge sharpening approach using a 3\u0026times;3 high-pass kernel. This final preprocessing step selectively enhances high-frequency components while maintaining strict control over potential artifacts through careful parameterization. The sharpening process, applied after denoising but prior to network input, provides an optimal balance between feature enhancement and natural appearance.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e3.2 Data Augmentation\u003c/h2\u003e\u003cp\u003eTo increase the variability of the training dataset and enhance model robustness, we employed various data augmentation techniques, focusing on both global and lesion-specific modifications. For mass-centric augmentation, applied only to lesion-containing regions, we performed safe flips (horizontal and vertical) with boundary checks to prevent lesion cropping, small-angle rotations within \u0026plusmn;\u0026thinsp;15\u0026deg; to maintain anatomical integrity, and localized contrast enhancement using adaptive histogram equalization to improve the visibility of mass regions. For global augmentation, all images were randomly flipped (by 50%) along both vertical and horizontal axes, and brightness adjustments within a controlled range to simulate different imaging conditions. These augmentation strategies collectively aimed to improve the model\u0026rsquo;s generalization capability and address class imbalance in the CBIS-DDSM dataset.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e3.3 SE-Attention U-Net: An Enhanced Architecture with Dual Attention Mechanisms\u003c/h2\u003e\u003cp\u003eAccording the foundational U-Net architecture [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], we propose an advanced U-Net variant with dual attention mechanisms and a specialized loss function for mammographic mass segmentation through three key innovations. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, our encoder pathway incorporates residual convolutional blocks paired with squeeze-and-excitation (SE) modules, enabling both local feature extraction and global channel-wise feature recalibration. Each SE block dynamically emphasizes diagnostically relevant features while suppressing less informative channels through learned excitation weights. Spatial attention gates are integrated at skip connections between the encoder and decoder, which selectively emphasize mass-containing regions while suppressing irrelevant background areas, particularly crucial for maintaining boundary precision in sub-15mm lesions. The decoder pathway utilizes transposed convolutions with halved channel depth at each up-sampling stage, systematically recovering spatial resolution while integrating attention-weighted features from corresponding encoder levels. A strategic 0.5 dropout rate enhances generalization capability without compromising feature retention. Throughout the architecture, we maintain dimensional consistency through symmetric padding, ensuring the final segmentation map precisely aligns with input mammogram dimensions. This comprehensive design achieves three critical objectives: (1) preservation of fine structural details through residual learning and attention mechanisms, (2) adaptive feature enhancement via SE blocks, and (3) computationally efficient processing through optimized channel depth reduction in the decoder pathway. The complete system, trained with our novel Stable Extreme Loss Function.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e3.4 Training Protocol with Novel Stable Extreme Loss Function\u003c/h2\u003e\u003cp\u003eOur training process employs a carefully designed three-phase strategy to optimize performance for mammographic mass segmentation. In the initial phase, we utilize standard Dice loss to establish baseline feature extraction capabilities, allowing the network to learn fundamental segmentation patterns. The model then transitions to our proposed Stable Extreme Loss Function (SELF) in the second phase, which specifically addresses three critical challenges. In the final phase, we employ transfer learning by freezing the encoder layers while fine-tuning the decoder components with SELF. This strategic approach preserved the learned hierarchical feature representations while optimizing the segmentation-specific architecture. This comprehensive training approach, combined with Adam optimization (β₁=0.9, β₂=0.999) and early stopping based on validation metrics, enables robust segmentation performance across all mass sizes.\u003c/p\u003e\u003cp\u003e\u003cb\u003eNovel Stable Extreme Loss Function (SELF)\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe propose a novel loss function to address three fundamental challenges in small mass mammographic segmentation: (1) extreme class imbalance where foreground pixels constitute less than 2% of the image area, (2) precise boundary localization for sub-15mm mass, and (3) numerical stability during optimization. SELF combines three strategically weighted components: a boundary-aware Dice loss (60%), mass-weighted focal loss (30%), and edge consistency loss (10%) - each targeting distinct aspects of the segmentation problem.\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eBoundary-Aware Dice Loss\u003c/b\u003e\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eBoundary-aware Dice loss is a novel technique developed to improve segmentation performance in medical imaging by overcoming the shortcomings of standard Dice loss[\u003cspan additionalcitationids=\"CR35\" citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. This approach focuses on precisely delineating boundaries. It assigns higher weights to pixels located near the boundaries of organs or lesions.\u003c/p\u003e\u003cp\u003eOur boundary-aware Dice loss (Eq.\u0026nbsp;\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) extends the conventional formulation by integrating three key innovations to improve margin delineation.\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:{L}_{B-Dice}=1-\\frac{2.\\sum\\:_{i=1}^{N}{w}_{i}.\\left({y}_{i}{\\widehat{y}}_{i}\\right)+ϵ}{\\sum\\:_{i=1}^{N}{w}_{i}.\\left({y}_{i}+{\\widehat{y}}_{i}\\right)+ϵ}+\\lambda\\:{‖{\\nabla\\:}^{2}y-{\\nabla\\:}^{2}\\widehat{y}‖}_{2}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eFirst, a 3\u0026times;3 Laplacian kernel explicitly enhances sensitivity to mass boundaries by amplifying gradient signals at edge voxels (Eq.\u0026nbsp;\u003cspan refid=\"Equ2\" class=\"InternalRef\"\u003e2\u003c/span\u003e):\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:{w}_{i}=1+\\alpha\\:.\\left|{Laplacian\\left(y\\right)}_{i}\\right|$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Laplacian\\left(y\\right)\\)\u003c/span\u003e\u003c/span\u003e\u003c/strong\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:3\\times\\:3\\)\u003c/span\u003e\u003c/span\u003eedge detection kernel output\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\alpha\\:=5.0\\)\u003c/span\u003e\u003c/span\u003e\u003c/strong\u003e\u003cp\u003eEmpirically determined edge emphasis factor\u003c/p\u003e\u003c/p\u003e\u003cp\u003eSecond, dynamic foreground weighting adjusts pixel-wise (Eq.\u0026nbsp;\u003cspan refid=\"Equ3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) contributions based on batch-specific mass prevalence, preventing small-mass features from being overwhelmed by background dominance:\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:{w}_{i}\\leftarrow\\:{w}_{i}.\\frac{\\beta\\:}{{\\text{B}\\text{a}\\text{t}\\text{c}\\text{h}}_{\\text{p}\\text{r}\\text{e}\\text{v}\\text{a}\\text{l}\\text{e}\\text{n}\\text{c}\\text{e}}+ϵ}\\:\\:\\:\\:\\:\\:\\:\\:\\text{w}\\text{h}\\text{e}\\text{r}\\text{e}\\:\\:\\:\\:{\\text{B}\\text{a}\\text{t}\\text{c}\\text{h}}_{\\text{p}\\text{r}\\text{e}\\text{v}\\text{a}\\text{l}\\text{e}\\text{n}\\text{c}\\text{e}}=\\frac{{N}_{fg}}{{N}_{total}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\beta\\:\\)\u003c/span\u003e\u003c/span\u003e=0.1: Normalization constant\u003c/p\u003e\u003cp\u003e\u003cstrong\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{N}_{fg}\\)\u003c/span\u003e\u003c/span\u003e\u003c/strong\u003e\u003cp\u003eForeground pixels in batch\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{N}_{total}\\)\u003c/span\u003e\u003c/span\u003e\u003c/strong\u003e\u003cp\u003eTotal pixels in batch\u003c/p\u003e\u003c/p\u003e\u003cp\u003eThird, a second-order gradient term (Eq.\u0026nbsp;\u003cspan refid=\"Equ4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) enforces geometric consistency, regularizing irregular contour formations. This yields improvement in margin sharpness compared to standard implementations, particularly crucial for sub-15mm lesions where boundary precision directly impacts diagnostic accuracy.\u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$$\\:\\lambda\\:{‖{\\nabla\\:}^{2}y-{\\nabla\\:}^{2}\\widehat{y}‖}_{2}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\nabla\\:}^{2}\\)\u003c/span\u003e\u003c/span\u003e\u003c/strong\u003e\u003cp\u003eLaplacian operator for contour smoothness\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\lambda\\:\\)\u003c/span\u003e\u003c/span\u003e\u003c/strong\u003e\u003cp\u003eGeometric consistency weight\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:ϵ={10}^{-7}\\)\u003c/span\u003e\u003c/span\u003e\u003c/strong\u003e\u003cp\u003eNumerical stability constant\u003c/p\u003e\u003c/p\u003e\u003cp\u003eThis formulation improves margin sharpness for sub-15mm lesions.\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eMass-Weighted Focal Loss\u003c/b\u003e\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eMass-weighted focal loss represents an enhanced variant of focal loss[\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e], designed to address class imbalance in various medical imaging tasks[\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. To address extreme class imbalance (\u0026lt;\u0026thinsp;2% foreground), we augment the focal loss with two stabilization mechanisms (Eq.\u0026nbsp;\u003cspan refid=\"Equ5\" class=\"InternalRef\"\u003e5\u003c/span\u003e,\u003cspan refid=\"Equ6\" class=\"InternalRef\"\u003e6\u003c/span\u003e):\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003e1. Adaptive Class Weighting\u003c/h3\u003e\n\u003cp\u003eIt dynamically scales foreground loss contributions based on real-time batch statistics to counteract extreme class imbalance The foreground weighting factor \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{w}_{fg}\\:\\)\u003c/span\u003e\u003c/span\u003e​ dynamically adjusts to batch-specific class distributions:\u003cdiv id=\"Equ5\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ5\" name=\"EquationSource\"\u003e\n$$\\:{w}_{fg}={min}\\left(50,\\frac{1}{{\\text{B}\\text{a}\\text{t}\\text{c}\\text{h}}_{\\text{p}\\text{r}\\text{e}\\text{v}\\text{a}\\text{l}\\text{e}\\text{n}\\text{c}\\text{e}}+ϵ}\\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:ϵ={10}^{-7}\\)\u003c/span\u003e\u003c/span\u003e: Ensures numerical stability for mass-free batches\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eCap (50\u0026times;): Prevents dominance of ultra-rare masses (e.g., batches with \u0026lt;\u0026thinsp;0.02% prevalence)\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThis adaptation ensures proportional gradient contributions from small masses while maintaining optimization stability.\u003c/p\u003e\n\u003ch3\u003e2. Stabilized Parameterization\u003c/h3\u003e\n\u003cp\u003eWe modify the focal loss formulation with three key stabilizers (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{w}_{fg},\\:\\:\\gamma\\:,\\:\\:{p}_{t})\\)\u003c/span\u003e\u003c/span\u003e:\u003cdiv id=\"Equ6\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ6\" name=\"EquationSource\"\u003e\n$$\\:{L}_{focal}={w}_{fg}.\\left[-{\\left(1-{p}_{t}\\right)}^{\\gamma\\:}\\text{log}\\left({p}_{t}\\right)\\right],\\:\\:\\:\\:\\:\\:\\:where\\:\\:\\:{p}_{t}=clip\\:(p,ϵ,1-ϵ)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e6\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eWe reduce the focusing parameter to γ\u0026thinsp;=\u0026thinsp;3.0 (vs. standard γ\u0026thinsp;=\u0026thinsp;4.0) to soften gradients for tiny masses, while double clipping probability estimates \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:(p\\in\\:\\left[{10}^{-6},1-{10}^{-6}\\right])\\)\u003c/span\u003e\u003c/span\u003e guarantees numerical stability during backpropagation.\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eEdge Consistency Loss\u003c/b\u003e\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThe edge consistency loss enhances boundary precision by enforcing alignment between predicted and ground-truth mass margins[\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. We compute the L1 distance between edges extracted via 3\u0026times;3 Laplacian filtering from both the segmentation output (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\widehat{E}\\)\u003c/span\u003e\u003c/span\u003e) and ground truth (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:E\\)\u003c/span\u003e\u003c/span\u003e) as shown in Eq.\u0026nbsp;7:\u003c/p\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{L}_{edg}=\\frac{1}{N}\\sum\\:_{i=1}^{N}\\left|{\\widehat{E}}_{i}-{E}_{i}\\right|,\\:\\:\\:\\:\\:\\:\\:\\:where\\:\\:\\:\\widehat{E}=Laplacian\\:({y}_{pred}),\\:\\:\\:\\:\\:E=Laplacian({y}_{true}\\)\u003c/span\u003e\u003c/span\u003e) (7)\u003c/p\u003e\u003cp\u003eThis term addresses three critical clinical requirements: Margin Sharpness, reduces false positives at mass boundaries, and maintains topological consistency in irregular masses.\u003c/p\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e\u003cb\u003e3.5 Evaluation Metrics\u003c/b\u003e\u003c/h2\u003e\u003cp\u003eTo comprehensively evaluate the proposed model for medical image segmentation and classification, we employ metrics assessing both pixel-level classification accuracy and region-wise segmentation overlap. These metrics address class imbalance and spatial delineation challenges inherent to medical datasets (e.g., tumors occupying small regions).\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eClassification Metrics\u003c/b\u003e\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThese metrics evaluate the model\u0026rsquo;s global ability to correctly classify pixels, emphasizing robustness to class imbalance:\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003e1. Accuracy\u003c/h3\u003e\n\u003cp\u003eAccuracy provides an overall measure of how well the model correctly classifies pixels in the image, combining correct positive and negative predictions. It is intuitive but can be misleading when class distributions are imbalanced, especially in medical images where abnormalities may occupy a small region. It is defined as Eq.\u0026nbsp;\u003cspan refid=\"Equ7\" class=\"InternalRef\"\u003e8\u003c/span\u003e:\u003cdiv id=\"Equ7\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ7\" name=\"EquationSource\"\u003e\n$$\\:Accuracy=\\frac{TP+TN}{TP+TN+FP+FN}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e8\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eWhere:\u003c/p\u003e\u003cp\u003eTP (True Positives): Pixels correctly identified as part of the abnormality (e.g., tumor).\u003c/p\u003e\u003cp\u003eTN (True Negatives): Pixels correctly identified as normal/background.\u003c/p\u003e\u003cp\u003eFP (False Positives): Normal pixels incorrectly classified as abnormal.\u003c/p\u003e\u003cp\u003eFN (False Negatives): Abnormal pixels incorrectly classified as normal.\u003c/p\u003e\n\u003ch3\u003e2. Precision\u003c/h3\u003e\n\u003cp\u003ePrecision indicates the proportion of positive identifications that were actually correct. It focuses on the reliability of positive detections. High precision means fewer false positives, which is critical to avoid unnecessary further procedures. It is calculated as Eq.\u0026nbsp;\u003cspan refid=\"Equ8\" class=\"InternalRef\"\u003e9\u003c/span\u003e:\u003cdiv id=\"Equ8\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ8\" name=\"EquationSource\"\u003e\n$$\\:Precision=\\frac{TP}{TP+FP}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e9\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\n\u003ch3\u003e3. Sensitivity (Recall)\u003c/h3\u003e\n\u003cp\u003eSensitivity measures the model's ability to correctly identify all actual abnormal pixels. A high recall is essential to minimize missed detections, reducing the risk of overlooking disease. It is computed as Eq.\u0026nbsp;\u003cspan refid=\"Equ9\" class=\"InternalRef\"\u003e10\u003c/span\u003e:\u003cdiv id=\"Equ9\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ9\" name=\"EquationSource\"\u003e\n$$\\:Sensitivity=\\frac{TP}{TP+FN}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e10\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\n\u003ch3\u003e4. F1 Score\u003c/h3\u003e\n\u003cp\u003eThe F1 Score (Eq.\u0026nbsp;\u003cspan refid=\"Equ10\" class=\"InternalRef\"\u003e11\u003c/span\u003e) is the harmonic mean of precision and recall, offering a single metric that balances both false positives and false negatives. It is particularly useful when the data is imbalanced and when one needs to balance precision and recall. The F1-score is mathematically equivalent to the Dice coefficient (Segmentation Metrics) but is included here for consistency with general classification literature. A higher F1 score indicates a better balance of precision and recall:\u003cdiv id=\"Equ10\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ10\" name=\"EquationSource\"\u003e\n$$\\:F1\\:Score=2\\times\\:\\:\\:\\frac{Precision\\times\\:Sensitivity}{Precision+Sensitivity}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e11\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eSegmentation Metrics\u003c/b\u003e\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThese metrics quantify spatial overlap between predicted and ground-truth masks, emphasizing anatomical delineation:\u003c/p\u003e\n\u003ch3\u003e1. Dice Coefficient\u003c/h3\u003e\n\u003cp\u003eThe Dice Coefficient (Eq.\u0026nbsp;\u003cspan refid=\"Equ11\" class=\"InternalRef\"\u003e12\u003c/span\u003e) (also called S\u0026oslash;rensen\u0026ndash;Dice index) is specifically designed for measuring the overlap between the predicted and ground truth binary masks in image segmentation tasks. Dice coefficient values range from 0 (no overlap) to 1 (perfect overlap):\u003cdiv id=\"Equ11\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ11\" name=\"EquationSource\"\u003e\n$$\\:Dice\\:Coefficient=\\:\\frac{2\\times\\:\\:|X\\cap\\:Y|}{\\left|X\\right|+\\left|Y\\right|}=\\frac{2.\\:TP}{2.\\:TP+FP+FN}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e12\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eX (Ground Truth) : The manually annotated (true) segmentation mask.\u003c/p\u003e\u003cp\u003eY (Prediction) : The model's predicted segmentation mask.\u003c/p\u003e\u003cp\u003e∣X\u0026cap;Y∣ (TP) : Pixels correctly predicted as part of the mass.\u003c/p\u003e\u003cp\u003e∣X∣ (Ground Truth Size) : Total positives in the true mask (TP\u0026thinsp;+\u0026thinsp;FN).\u003c/p\u003e\u003cp\u003e∣Y∣ (Prediction Size) : Total positives in the predicted mask (TP\u0026thinsp;+\u0026thinsp;FP).\u003c/p\u003e\n\u003ch3\u003e2. Jaccard coefficient (Intersection over Union, IoU)\u003c/h3\u003e\n\u003cp\u003eThe Jaccard coefficient, also known as Intersection over Union (IoU), quantifies the overlap between the predicted and ground truth masks as the ratio of their intersection to their union as shown in Eq.\u0026nbsp;\u003cspan refid=\"Equ12\" class=\"InternalRef\"\u003e13\u003c/span\u003e. It ranges from 0 to 1, where 1 indicates perfect overlap. IoU is more stringent than the Dice score, rewarding higher precision in spatial overlap:\u003cdiv id=\"Equ12\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ12\" name=\"EquationSource\"\u003e\n$$\\:Jaccard\\:Coefficient=\\frac{|X\\cap\\:Y|}{|X\\cup\\:Y|}=\\frac{TP}{TP+FP+FN}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e13\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e"},{"header":"4. Experimental Results","content":"\u003cp\u003eIn this section, we evaluate the performance of the proposed method against existing approaches using the widely recognized CBIS-DDSM dataset. Performance is assessed using standard evaluation metrics, and visual comparisons are provided to highlight key differences.\u003c/p\u003e\u003cp\u003eThe experiments utilized mammogram images from the CBIS-DDSM dataset, which were resized to 224\u0026times;224 pixels and processed through an advanced preprocessing pipeline that included adaptive histogram equalization, denoising, and edge enhancement. The rigorous training and evaluation process, combined with the enhanced model architecture and preprocessing techniques, establish a robust solution for small mass segmentation in mammography\u0026mdash;a critical step in accurate breast cancer detection.\u003c/p\u003e\u003cp\u003eThe dataset was randomly divided into training, validation, and testing sets, with 70% allocated for training. The remaining 30% was evenly split between validation and testing sets. The proposed algorithm was applied to the mass segmentation task, and the results are illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo comprehensively evaluate the performance of the proposed approach, we present results of several relevant metrics. Figure\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e illustrates the classification-related metrics, including accuracy, precision, sensitivity, and F1-score. These metrics provide insights into the model's ability to correctly identify and classify lesions, reflecting its robustness and reliability. To illustrate the performance of the proposed segmentation model, we present both the Dice coefficient and Jaccard coefficient for the training and validation phases in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eWe compare the performance of our method with several state-of-the-art approaches in terms of accuracy, precision, sensitivity, F1-score, Dice coefficient and Jaccard coefficient as detailed in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison of the proposed method with existing methods in terms of accuracy, precision, sensitivity and F1-score metrics on CBIS-DSSM dataset.\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=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMethods\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSensitivity (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAccuracy (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003ePrecision (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eF1-score (%)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSun et al. [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e84.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHou et al. [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e85.60\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRamesh et al.[\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e79\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eEl-Banby et al.[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e90.58\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e87.98\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eKhan et al. [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e75.41\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e69.98\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDuggento et al. [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e84.40\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e71.19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRajalakshmi et al. [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e84.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLiao and Aagaard [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]\u003c/p\u003e\u003cp\u003e(ResNet-50 (448\u0026times;448))\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e62.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e68.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e65.38\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOza et al. [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e91.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIqbal and Sharif [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e77.10\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOur proposed model\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e97.08\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e99.54\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e97.97\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e97.53\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\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\u003eComparison of the proposed and state-of-the-art methods in terms of Dice coefficient and Jaccard coefficient on CBIS-DSSM dataset.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\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\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMethods\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDice coefficient (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eJaccard coefficient (IoU)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSun et al. [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e81.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHou et al. [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e86.30\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRamesh et al.[\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e82.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eEl-Banby et al.[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e87.98\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eZhang et al.[\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e58.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e41.96\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRajalakshmi et al. [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e82.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e85.70\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBaccouche et al.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e89.52\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e80.02\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eChen et al. [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e82.16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTsochatzidis et al.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e72.20\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e56.50\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFazilov et al. [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e77.82\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e65.17\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eJin et al. [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e85.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOur proposed model\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003eMean: 91.00\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eMax: 96.55\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003eMean: 86.01\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eMax: 93.36\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe proposed model demonstrates superior performance across both classification and segmentation tasks on the CBIS-DSSM dataset, outperforming existing state-of-the-art methods in all evaluated metrics. For classification (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), the model achieves an accuracy of 99.54%, precision of 97.97%, sensitivity of 97.08%, and F1-score of 97.53%. In segmentation (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), the model achieves a Dice coefficient of 91.00% and Jaccard coefficient (IoU) of 86.01%, surpassing all competing methods. These results demonstrate that our model not only provides highly accurate classification predictions but also delivers precise segmentation masks, enabling reliable lesion localization alongside diagnostic decisions. The simultaneous excellence in both tasks suggests our approach effectively captures both global and local features in medical images, making it particularly valuable for clinical applications requiring comprehensive analysis.\u003c/p\u003e"},{"header":"5. Conclusion and future directions","content":"\u003cp\u003eOur study introduces a SE-Attention U-Net framework that significantly advances mammographic mass segmentation through three key innovations: (1) adaptive feature refinement with squeeze-and-excitation blocks, (2) attention-guided spatial prioritization of small lesions, and (3) a stable loss function optimized for severe class imbalance.\u003c/p\u003e\u003cp\u003eRigorous multi-level evaluation on the CBIS-DDSM dataset confirms the clinical viability of our model. At the pixel level, the model achieves a Dice coefficient of 0.966 and a Jaccard index of 0.914, demonstrating exceptional boundary accuracy\u0026mdash;particularly useful for irregularly shaped masses and critical for surgical planning and radiotherapy. At the region level, the model attains 97.08% sensitivity and a 97.53% F1-score, indicating reliable lesion-wise diagnosis\u0026mdash;crucial for reducing missed cancers during screening. Notably, with 99.54% accuracy and 97.97% precision, our approach exhibits robustness against false positives, effectively addressing a key limitation of existing CAD systems.\u003c/p\u003e\u003cp\u003eFor future work, we plan to explore the integration of multimodal imaging data, such as combining mammograms with ultrasound or MRI to enhance diagnostic accuracy. Incorporating radiomics features could further improve lesion characterization and aid in personalized treatment planning. Additionally, evaluating our framework on diverse datasets beyond CBIS-DDSM will help verify its generalizability across different populations and imaging protocols. Furthermore, extending our approach to process 3D mammographic or tomographic images may provide more comprehensive spatial context and improve segmentation precision, especially for complex lesions.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eCompeting Interests\u003c/h2\u003e\u003cp\u003eThe author declares no competing interests.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eThe author M.H. conceptualized the study, designed the methodology, performed the analysis, wrote the manuscript, and prepared all figures/tables.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe dataset used in this study is the publicly available CBIS-DDSM (Curated Breast Imaging Subset of the Digital Database for Screening Mammography). Access to the CBIS-DDSM dataset can be obtained from https://www.kaggle.com/datasets/awsaf49/cbis-ddsm-breast-cancer-image-dataset . All relevant data supporting the findings of this study are available from the author, Maliheh Habibi, at [email protected], upon reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eZhang, Y. et al. From single to universal: tiny lesion detection in medical imaging. \u003cem\u003eArtif. Intell. Rev.\u003c/em\u003e \u003cb\u003e57\u003c/b\u003e (8), 192 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang, J., Wang, S. \u0026amp; Zhang, Y. Deep learning on medical image analysis. \u003cem\u003eCAAI Trans. Intell. Technol.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e (1), 1\u0026ndash;35 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEsteva, A. et al. Dermatologist-level classification of skin cancer with deep neural networks. \u003cem\u003enature\u003c/em\u003e \u003cb\u003e542\u003c/b\u003e (7639), 115\u0026ndash;118 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLitjens, G. et al. A survey on deep learning in medical image analysis. \u003cem\u003eMed. Image. Anal.\u003c/em\u003e \u003cb\u003e42\u003c/b\u003e, 60\u0026ndash;88 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAbdulwahhab, A. H. et al. A review on medical image applications based on deep learning techniques. \u003cem\u003eJ. Image Graphics\u003c/em\u003e. \u003cb\u003e12\u003c/b\u003e (3), 215\u0026ndash;227 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOluseyi, O. M., Udofot, A. I. \u0026amp; Bassey, E. \u003cem\u003eEnhancing Medical Image Diagnosis Using Convolutional Neural Network and Transfer Learning.\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRonneberger, O., Fischer, P. \u0026amp; Brox, T. \u003cem\u003eU-net: Convolutional networks for biomedical image segmentation\u003c/em\u003e. in \u003cem\u003eMedical image computing and computer-assisted intervention\u0026ndash;MICCAI 2015: 18th international conference, Munich, Germany, October 5\u0026ndash;9\u003c/em\u003e, proceedings, part III 18. 2015. Springer. (2015).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAvazov, K. et al. Dynamic Focus on Tumor Boundaries: A Lightweight U-Net for MRI Brain Tumor Segmentation. \u003cem\u003eBioengineering\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e (12), 1302 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSalem, N., Malik, H. \u0026amp; Shams, A. Medical image enhancement based on histogram algorithms. \u003cem\u003eProcedia Comput. Sci.\u003c/em\u003e \u003cb\u003e163\u003c/b\u003e, 300\u0026ndash;311 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePatel, S. et al. \u003cem\u003eComparative study on histogram equalization techniques for medical image enhancement\u003c/em\u003e. in \u003cem\u003eSoft Computing for Problem Solving: SocProS 2018, Volume 1\u003c/em\u003e. Springer. (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePizer, S. M. et al. \u003cem\u003eAdaptive histogram equalization for automatic contrast enhancement of medical images. in Application of Optical Instrumentation in Medicine XIV and Picture Archiving and Communication Systems\u003c/em\u003e (Spie, 1986).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVidyasaraswathi, H. \u0026amp; Hanumantharaju, M. \u003cem\u003eReview of various histogram based medical image enhancement techniques\u003c/em\u003e. in \u003cem\u003eProceedings of the International Conference on Advanced Research in Computer Science Engineering \u0026amp; Technology (ICARCSET 2015)\u003c/em\u003e. 2015. 2015. (2015).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHaddadi, Y. R., Mansouri, B. \u0026amp; Khodja, F. Z. I. \u003cem\u003eA novel medical image enhancement algorithm based on CLAHE and pelican optimization\u003c/em\u003e. \u003cem\u003eMultimedia Tools Appl.\u003c/em\u003e, : pp. 1\u0026ndash;20. (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi, L., Si, Y. \u0026amp; Jia, Z. Medical image enhancement based on CLAHE and unsharp masking in NSCT domain. \u003cem\u003eJ. Med. Imaging Health Inf.\u003c/em\u003e \u003cb\u003e8\u003c/b\u003e (3), 431\u0026ndash;438 (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSharma, R. \u0026amp; Kamra, A. \u003cem\u003eA review on CLAHE based enhancement techniques\u003c/em\u003e. in \u003cem\u003e6th International Conference on Contemporary Computing and Informatics (IC3I)\u003c/em\u003e. 2023. IEEE. 2023. IEEE. (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYan, L. J. et al. A hybrid deep learning model for mammographic breast cancer detection: Multi-autoencoder and attention mechanisms. \u003cem\u003eJ. Radiation Res. Appl. Sci.\u003c/em\u003e \u003cb\u003e18\u003c/b\u003e (3), 101578 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHu, J., Shen, L. \u0026amp; Sun, G. \u003cem\u003eSqueeze-and-excitation networks\u003c/em\u003e. in \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition\u003c/em\u003e. (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWoo, S. et al. \u003cem\u003eCbam: Convolutional block attention module\u003c/em\u003e. in \u003cem\u003eProceedings of the European conference on computer vision (ECCV)\u003c/em\u003e. (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZuiderveld, K. J. Contrast limited adaptive histogram equalization. \u003cem\u003eGraphics gems\u003c/em\u003e. \u003cb\u003e4\u003c/b\u003e (1), 474\u0026ndash;485 (1994).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEl-Banby, G. M. et al. Automated abnormalities detection in mammography using deep learning. \u003cem\u003eComplex. Intell. Syst.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e (5), 7279\u0026ndash;7295 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eIslam, M. R. et al. \u003cem\u003eEnhancing Semantic Segmentation with Adaptive Focal Loss: A Novel Approach.\u003c/em\u003e arXiv preprint arXiv:2407.09828, (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRahman, M. A. \u0026amp; Wang, Y. \u003cem\u003eOptimizing Intersection-Over-Union in Deep Neural Networks for Image Segmentation\u003c/em\u003e (Springer International Publishing, 2016).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBuades, A., Coll, B. \u0026amp; Morel, J. M. \u003cem\u003eA non-local algorithm for image denoising\u003c/em\u003e. in \u003cem\u003eIEEE computer society conference on computer vision and pattern recognition (CVPR'05)\u003c/em\u003e. 2005. Ieee. 2005. Ieee. (2005).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKumar, M. \u0026amp; Diwakar, M. \u003cem\u003eEdge preservation based CT image denoising using wavelet and curvelet transforms\u003c/em\u003e. in \u003cem\u003eProceedings of Fifth International Conference on Soft Computing for Problem Solving: SocProS 2015, Volume 1\u003c/em\u003e. Springer. (2016).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang, Y. \u0026amp; Zhou, H. Total Variation Wavelet-Based Medical Image Denoising. \u003cem\u003eInt. J. Biomed. Imaging\u003c/em\u003e. \u003cb\u003e2006\u003c/b\u003e (1), 089095 (2006).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLei, D. et al. A spatiotemporal fusion method based on multiscale feature extraction and spatial channel attention mechanism. \u003cem\u003eRemote Sens.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e (3), 461 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSi, Y. et al. SCSA: Exploring the synergistic effects between spatial and channel attention. \u003cem\u003eNeurocomputing\u003c/em\u003e \u003cb\u003e634\u003c/b\u003e, 129866 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAbbasi, S. et al. Improved A-Line and B-Line Detection in Lung Ultrasound Using Deep Learning with Boundary-Aware Dice Loss. \u003cem\u003eBioengineering\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e (3), 311 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNgoc, M. \u0026Ocirc;. V. et al. \u003cem\u003eIntroducing the Boundary-Aware loss for deep image segmentation\u003c/em\u003e. in \u003cem\u003eBritish Machine Vision Conference (BMVC) 2021\u003c/em\u003e. (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMahmoodi, N. et al. Automatically weighted focal loss for imbalance learning. \u003cem\u003eNeural Comput. Appl.\u003c/em\u003e \u003cb\u003e37\u003c/b\u003e (5), 4035\u0026ndash;4052 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eQin, R. et al. \u003cem\u003eWeighted focal loss: An effective loss function to overcome unbalance problem of chest X-ray14\u003c/em\u003e. in \u003cem\u003eIOP conference series: Materials science and engineering\u003c/em\u003e. IOP Publishing. (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLee, R. S. et al. A curated mammography data set for use in computer-aided detection and diagnosis research. \u003cem\u003eSci. data\u003c/em\u003e. \u003cb\u003e4\u003c/b\u003e (1), 1\u0026ndash;9 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eClark, K. et al. The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository. \u003cem\u003eJ. Digit. Imaging\u003c/em\u003e. \u003cb\u003e26\u003c/b\u003e, 1045\u0026ndash;1057 (2013).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKim, M. \u0026amp; Lee, B. D. A simple generic method for effective boundary extraction in medical image segmentation. \u003cem\u003eIEEE Access.\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, 103875\u0026ndash;103884 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang, Y. et al. \u003cem\u003eA 3D\u0026thinsp;+\u0026thinsp;2D CNN approach incorporating boundary loss for stroke lesion segmentation\u003c/em\u003e. in \u003cem\u003eMachine Learning in Medical Imaging: 11th International Workshop, MLMI 2020, Held in Conjunction with MICCAI 2020, Lima, Peru, October 4, 2020, Proceedings 11\u003c/em\u003e. Springer. (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShi, P. et al. \u003cem\u003eCenterline boundary dice loss for vascular segmentation\u003c/em\u003e. in \u003cem\u003eInternational Conference on Medical Image Computing and Computer-Assisted Intervention\u003c/em\u003e. Springer. (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSeth, A. \u003cem\u003eThe Application of Focal Loss in various Domains: A Survey\u003c/em\u003e. \u003cem\u003eInt. J. Res. Appl. Sci. Eng. Technol. (IJRASET)\u003c/em\u003e, (2321\u0026ndash;9653 ). (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAliniya, P. et al. Improved loss function for mass segmentation in mammography images using density and mass size. \u003cem\u003eJ. Imaging\u003c/em\u003e. \u003cb\u003e10\u003c/b\u003e (1), 20 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eQiu, Y. et al. \u003cem\u003eSemi-ECNet: Edge-Consistency Based Semi-Supervised Retinal Vessel Segmentation Network\u003c/em\u003e. in \u003cem\u003e2024 IEEE International Symposium on Biomedical Imaging (ISBI)\u003c/em\u003e. IEEE. (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSun, H. et al. AUNet: attention-guided dense-upsampling networks for breast mass segmentation in whole mammograms. \u003cem\u003ePhys. Med. Biol.\u003c/em\u003e \u003cb\u003e65\u003c/b\u003e (5), 055005 (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHou, X. et al. Mass segmentation for whole mammograms via attentive multi-task learning framework. \u003cem\u003ePhys. Med. Biol.\u003c/em\u003e \u003cb\u003e66\u003c/b\u003e (10), 105015 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRajalakshmi, N. R. et al. Deeply supervised u-net for mass segmentation in digital mammograms. \u003cem\u003eInt. J. Imaging Syst. Technol.\u003c/em\u003e \u003cb\u003e31\u003c/b\u003e (1), 59\u0026ndash;71 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKhan, H. N. et al. Multi-view feature fusion based four views model for mammogram classification using convolutional neural network. \u003cem\u003eIEEE Access.\u003c/em\u003e \u003cb\u003e7\u003c/b\u003e, 165724\u0026ndash;165733 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDuggento, A. et al. \u003cem\u003eAn ad hoc random initialization deep neural network architecture for discriminating malignant breast cancer lesions in mammographic images\u003c/em\u003e2019p. 5982834 (Contrast media \u0026amp; molecular imaging, 2019). 1.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiao, L. \u0026amp; Aagaard, E. M. An open codebase for enhancing transparency in deep learning-based breast cancer diagnosis utilizing CBIS-DDSM data. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e (1), 27318 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOza, P. R., Sharma, P. \u0026amp; Patel, S. A transfer representation learning approach for breast cancer diagnosis from mammograms using efficientnet models. \u003cem\u003eScalable Computing: Pract. Experience\u003c/em\u003e. \u003cb\u003e23\u003c/b\u003e (2), 51\u0026ndash;58 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eIqbal, A. \u0026amp; Sharif, M. BTS-ST: Swin transformer network for segmentation and classification of multimodality breast cancer images. \u003cem\u003eKnowl. Based Syst.\u003c/em\u003e \u003cb\u003e267\u003c/b\u003e, 110393 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang, B. et al. \u003cem\u003eA method for breast mass segmentation using image augmentation with SAM and receptive field expansion\u003c/em\u003e. in \u003cem\u003eProceedings of the 12th International Conference on Computing and Pattern Recognition\u003c/em\u003e. 2023. 2023. (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen, J. et al. A novel multi-scale adversarial networks for precise segmentation of x-ray breast mass. \u003cem\u003eIEEE access.\u003c/em\u003e \u003cb\u003e8\u003c/b\u003e, 103772\u0026ndash;103781 (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFazilov, S. et al. \u003cem\u003eSegmentation in mammographic images based on the U-NET model\u003c/em\u003e. in \u003cem\u003eAIP Conference Proceedings\u003c/em\u003e. AIP Publishing. (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJin, S. et al. \u003cem\u003eA Breast Mass Image Segmentation Method Based on Improved UNet 3\u0026thinsp;+\u0026thinsp;Network\u003c/em\u003e. in \u003cem\u003eIECON -49th Annual Conference of the IEEE Industrial Electronics Society\u003c/em\u003e. 2023. IEEE. 2023. IEEE. (2023).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Mammographic lesion localization, Lesion segmentation, U-Net architecture, Attention mechanisms, squeeze-and-excitation blocks, Class imbalance","lastPublishedDoi":"10.21203/rs.3.rs-6800663/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6800663/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eBreast cancer remains a leading cause of mortality among women worldwide, with early detection via mammography significantly improving patient outcomes. Automated segmentation of mammographic lesions using deep learning can enhance diagnostic efficiency; however, existing methods face critical challenges: (1) severe class imbalance (\u0026lt;\u0026thinsp;2% foreground pixels), (2) small lesion sizes (3\u0026ndash;15 mm), and (3) limited annotated datasets, which hinder clinical applicability. To overcome these limitations, we propose SE-Attention U-Net, a hybrid framework featuring squeeze-and-excitation blocks for adaptive feature refinement, attention gates to focus on salient regions, and a novel loss function explicitly designed to address extreme class imbalance.\u003c/p\u003e\u003cp\u003eWe evaluated our approach on the publicly available CBIS-DDSM dataset, a widely recognized benchmark in mammography research. Our model achieved state-of-the-art performance with a Dice coefficient of 91.00%, Jaccard coefficient of 86.01%, accuracy of 99.54%, precision of 97.97%, sensitivity of 97.08%, and an F1-score of 97.53%. These results demonstrate robust lesion localization and minimized false positives, outperforming existing methods. The proposed framework shows significant potential for clinical integration, providing radiologists with a reliable tool for early and accurate segmentation of small breast lesions.\u003c/p\u003e","manuscriptTitle":"SE-Attention U-Net: A Hybrid Loss-Optimized Model for Small Breast Lesion Segmentation in Mammography","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-20 08:38:19","doi":"10.21203/rs.3.rs-6800663/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-01-08T06:19:10+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-05T00:01:07+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-10-17T20:02:09+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"182435562971169288877317466798606220527","date":"2025-10-09T13:01:23+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"270472869871166475673765467549605359986","date":"2025-10-07T14:18:44+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-10-07T11:53:52+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-09-15T09:35:51+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-06-13T13:30:15+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-06-12T09:06:02+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-06-02T09:18:34+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"c014a339-6e47-4225-b4a5-d1185af48323","owner":[],"postedDate":"October 20th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[{"id":56496841,"name":"Biological sciences/Cancer"},{"id":56496842,"name":"Health sciences/Medical research"},{"id":56496843,"name":"Health sciences/Oncology"}],"tags":[],"updatedAt":"2026-01-08T06:24:35+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-20 08:38:19","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6800663","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6800663","identity":"rs-6800663","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}