Different BI-RADS breast cancer diagnosis using MobileNetV1 and vision transformer based on explainable artificial intelligence (XAI)

Comprehensive Summary

This paper proposes a hybrid deep learning model for multi-class BI-RADS breast cancer image classification from mammograms using a lightweight convolutional neural network (MobileNetV1) and Vision Transformers (ViT). Using the King Abdulaziz University BC Mammogram dataset (KAUBC), feature-level fusion, and a bagging logistic regression classifier, the proposed model demonstrated high performance using stratified five-fold cross-validation against other convolutional neural network and transformer-based architectures. The proposed model demonstrated high performance with accuracy, sensitivity, and specificity all above 99% for important BI-RADS categories such as normal, probably benign, suspicious, and highly suggestive of malignancy. Explainable AI techniques such as Grad-CAM and Grad-CAM++ were used to visualize important regions that contributed towards model predictions using heatmaps, which can be useful as an efficient and explainable CAD system for breast imaging.

Outcomes and Implications

The medical implications of these results are that an explainable AI-based augmented BI-RADS system could help improve the performance and consistency of radiologists' interpretations of mammography images by standardizing category assignment and highlighting suspicious regions, which are known to be problematic, especially for dense breast tissue where subtle signs are easily overlooked. This system, if externally validated on diverse populations, vendors, and imaging protocols, could be a valuable second reader system that could help reduce missed cancers and unnecessary recalls, although further studies are necessary to validate its effect on patient management and screening outcomes.