A Unified Deep Learning Framework integrating Multi scale features, Attention mechanism and Contrastive learning for Breast Cancer subtype Classification using Histopathology Images

Abstract

Breast cancer is a most prominent disease among individuals and its subtype classification is necessary and very critical for the personalized treatment. Many deep learning algorithms are predominantly used for the detection and classification of the breast cancer but, still they are struggling with many critical issues such as class imbalance problem, inadequate uncertainty quantification and inter-class similarity. This paper presents a unified deep learning approach that integrates the parallel convolution paths, multi-head attention mechanism, and contrastive learning for the automated categorization of four major subtypes of breast cancer malignancy such as Ductal carcinoma, Lobular carcinoma, Mucinous carcinoma and Papillary carcinoma. Unlike conventional methods, this framework unifies class imbalance handling and multi-scale fusion through information-theoretic regularization with progressive learning and confidence quantification. Experimental results demonstrate the exceptional performance of 98.8% overall accuracy, 98.3% balanced accuracy, 0.98 Matthews Correlation Coefficient, and 0.98 macro-average AUC-ROC, Confidence interval of 0.99, uncertainty estimation of 0.01, with per-class accuracy exceeding 99% for all subtypes. The notable results shows that the proposed framework performs well in comparing with the other methods. The framework also provides clinically interpretable confidence scores with decision thresholds, thereby making it most suitable for real-world clinical decision support systems.