Segmentation for Early Tumor Detection in Mammograms Via Temporal Discrepancy Analysis and Dynamic Loss Weighting

Detecting discrepancies between prior and current (P&C) mammograms is crucial for identifying subtle temporal changes, such as architectural distortions (AD) and calcifications, which indicate cancer. While radiologists effectively compare P&C images in clinical practice, deep learning models, particularly CNNs, often struggle to capture these changes over time, especially in limited-data scenarios. To address this, we propose a new hybrid deep learning framework specifically designed to analyze P&C discrepancies. The architecture identifies and processes multi-scale tissue changes to enhance segmentation accuracy by including specialized modules, referred to as feature discrepancy blocks (FDBs). Additionally, we develop a composite loss with sigmoid-based dynamic weighting to adaptively balance spatial consistency and boundary precision, ensuring stable learning. In this study, we show that by leveraging temporal context and regularization, our approach outperforms baseline methods, achieving precise binary segmentation masks and superior performance in dice score, mean IoU, sensitivity, and specificity. The code is available at: https://github.com/NabaviLab/MGSeg