Iterative multi-scale deep learning framework for reliable breast cancer diagnosis

Abstract

Accurate diagnosis of breast cancer from histopathological images is challenging due to variable tissue morphology and the subjectivity of manual interpretations. While CAD systems offer automated diagnosis, they often lack robust feature representation, contextual understanding, and integration of expert knowledge, limiting their effectiveness, especially in distinguishing invasive from pre-invasive carcinoma. This study presents a comprehensive deep learning-based diagnostic framework that integrates five novel modules to improve interpretability, feature robustness, and decision reliability. The Multiscale Attention Integrated Self-Supervised Representation (MAISSR) learns pathology-aware embeddings via co-optimized multiscale attention and contrastive learning. The Morphological-Geodesic Graph Convolutional Network (MG-GCN) combines geodesic topology with glandular morphology in a spatial graph model to capture epithelial transitions. The Hyper-Resolution Fusion Network with Cellular-Density Priors (HRF-CDPNet) enhances resolution in critical regions using cellular density maps. The Contextual Relational Transformer with Progression-Encoding (CRT-PE) models disease progression using spatial-contextual tokens to improve invasion mapping. Finally, Adaptive Cross-Modality Decision Calibration (ACM-DC) uses a reinforcement-learning-based agent to align machine predictions with expert annotations, especially in ambiguous cases. This integrated approach yields marked improvements in diagnostic metrics: F1 score increased from 82.5% to 89.3%, AUC from 0.88 to 0.94, and diagnostic agreement with experts from 85.2% to 94.8%. Overall, this work demonstrates the potential of a multi-factorial, multi-perspective framework to advance breast cancer diagnosis through optimized feature learning, spatial reasoning, and expert-machine synergy.