BC-SSL-X: A HYBRID SELF-SUPERVISED VISION TRANSFORMER FRAMEWORK WITH GRADIENT-GUIDED EXPLAINABILITY FOR INTERPRETABLE BREAST CANCER IMAGE CLASSIFICATION

Abstract

Breast cancer remains one of the most prevalent and diagnostically challenging malignancies worldwide, demanding computational methods that are simultaneously accurate, interpretable, and data-efficient for clinical adoption. This paper presents BC-SSL-X, a unified framework that integrates (1) a dual-objective self-supervised pre-training strategy combining contrastive learning and masked image modeling, (2) a hierarchical vision transformer (HViT) backbone with cross-scale attention fusion adapted for histopathology, mammography, and ultrasound imagery, and (3) a gradient-guided explainability engine that fuses Grad-CAM++ and Integrated Gradients attributions. Together, these components address the dual clinical requirement of state-of-the-art diagnostic accuracy and visually trustworthy predictions. BC-SSL-X is evaluated on five public benchmark datasets spanning three imaging modalities. On BreakHis, the framework attains 97.84% classification accuracy and an AUC of 0.9931, surpassing the strongest supervised and self-supervised baselines considered in this study. Ablation experiments confirm the complementary contribution of each architectural component, and qualitative and quantitative explainability analyses show strong spatial agreement between the fused attribution maps and expert-annotated regions of interest.