AMEDDE-Net: A Deeply Supervised Multi-Scale Attention Network for Breast Cancer Lesion Segmentation in Digital Mammography

Abstrak

This study proposes the Attention, Multi-scale, Enhanced, Dilated, Dual-head, Ensemble Network (AMEDDE-Net), a deeply supervised segmentation framework that integrates multi-scale feature extraction, enhanced attention mechanisms, dilated convolutions, dual-head prediction, and adaptive ensemble fusion to improve lesion-focused learning and boundary delineation. The model was evaluated using a publicly available Contrast-Enhanced Spectral Mammography (CESM) dataset and five-fold cross-validation. Extensive hyperparameter tuning, ablation studies, and comparisons with state-of-the-art convolutional and transformer-based models were conducted. The experimental results indicated that AMEDDE-Net achieved superior performance, with 97.25% accuracy, a Dice coefficient of 65.10%, an F1-score of 65.10%, and a recall of 71.83%, outperforming conventional U-Net variants and the transformer-based TransU-Net. Ablation analyses confirmed the importance of multi-scale attention and deep supervision, whereas qualitative results demonstrated improved boundary accuracy and reduced False Negatives. These findings highlight the potential of AMEDDE-Net as a reliable tool for automated breast cancer screening and clinical decision support.

Identifier

• DOI: https://doi.org/10.48084/etasr.18480
• OpenAlex: https://openalex.org/W7163739239