DEEP LEARNING METHODS FOR BREAST CANCER DIAGNOSIS: A STRUCTURED CRITICAL REVIEW OF ARCHITECTURES

Abstract

This structured critical review evaluates the application of deep learning techniques in breast cancer diagnosis, emphasizing a rigorous qualitative and methodological appraisal of validation practices. While Convolutional Neural Networks (CNNs), Attention Mechanisms, and Vision Transformers have demonstrated strong performance in retrospective studies processing complex imaging modalities (mammography, ultrasound, MRI, and histopathology), their transition to clinical deployment remains severely hindered. This review critically discusses recurring methodological vulnerabilities prevalent in contemporary literature, most notably data leakage via image-level dataset splitting, the absence of independent external validation, and the consequent inflation of diagnostic accuracy. By highlighting these systemic bottlenecks, we propose a paradigm shift toward inherent model interpretability, federated learning, and strict patient-level dataset partitioning to ensure the safe clinical translation of artificial intelligence in oncology.