Precision healthcare in action: A robust MRI framework for breast cancer diagnosis with enhanced feature analysis and boundary refinement

Abstract

Several advanced deep-learning techniques have been developed for breast cancer classification and segmentation. However, challenges such as tumor size variations, class imbalance, and boundary detection often limit their performance. A key limitation lies in the inability to retain location information during feature extraction, coupled with increased model complexity, which escalates computational demands and reduces batch size, further impacting performance. To address these issues, our proposed framework introduces advanced MRI pre-processing and an optimized encoder–decoder architecture. The encoder extracts features, while the decoder reconstructs the image. Efficient feature extraction is achieved through a Semantic Segmentation Module (SSM) with separable Conv2D, reducing computational load. The Essential Feature Attention Module (EFAM) and Deep Residual Module (DRM) further reduce parameters and enhance restoration accuracy by addressing the limitations of traditional U-Net models. The framework employs EFAM in place of concatenation for improved restoration and iterative correction via DRM to prevent input vanishing. Upsampling restores images to their original size, producing a three-channel output for normal, benign, or malignant classes using a sigmoid function. A tailored objective function, combining Dice Score and Weighted Cross Entropy (WCE) which mitigates class imbalance and enhances boundary detection. Evaluations on benchmark datasets demonstrated superior Dice scores: 93.164%, 94.175%, and 89.772% for benign classes and 95.086%, 97.902%, and 93.078% for malignant classes on BreastDM, Duke, and RIDER datasets, respectively. The proposed framework ensures accurate segmentation and classification across varying tumor sizes.

More by Galiya Ybytayeva

• Emotionally expressive facial animation driven by EmotionBERT embeddingsScientific Reports · 2026
• Physics-informed neural network framework for fast and accurate nanofluid heat-transfer prediction in microchannel heat sinksFrontiers in Physics · 2026 · Citations: 2
• Thermal mode modeling using neural network technologies and the finite element methodBulletin of Electrical Engineering and Informatics · 2026
• Explainable multi agent reinforcement learning framework for secure and adaptive communication in UAV swarm based fanetsScientific Reports · 2026 · Citations: 6