Uncertainty estimation of breast cancer classification using deep learning methods with explainable artificial intelligence

Abstract

Globally, breast cancer-related mortality among women requires early and precise diagnostic support to improve patient outcomes. The existing research work discusses using Machine learning and deep learning models for precise diagnosis with a combination of classified outcomes using various datasets. The proposed work focuses on uncertainty estimation of breast cancer using a deep learning framework with Explainable Artificial Intelligence (XAI) with the Curated Breast Imaging Subset of Digital Database for Screening Mammography Dataset. The novelty of the work is that it classifies breast cancer as benign, benign-without-callback, and malignant using a convolutional neural network (CNN) and a vision transformer (ViT), uncertainty estimation using Monte Carlo dropout, and gradient-weighted class activation mapping for visual localization of discriminative regions and SHapley Additive exPlanations for pixel-level feature importance to increase transparency as well as reliability. The consistency of the XAI maps produced by CNN and ViT has been examined quantitatively using the structural similarity index measure. The results show that ViT has more dispersed attention patterns, while CNN focuses more on isolated lesion sites. As a result, the proposed work supports clinical decision-making for precise breast cancer diagnosis, improves model interpretability, and supports uncertainty forecasting.