A Systematic Review of Machine Learning and Explainable AI in Breast Cancer Detection and Diagnosis: From Black-Box Models to Interpretable Clinical Decision Support Systems

Abstract

Breast cancer is one of the major causes of cancer-related deaths in women across the world. Timely and proper diagnosis is paramount to the higher survival. The recent developments in machine learning (ML) and deep learning (DL) have demonstrated impressive potential in breast cancer detection yet the black-box character restricts the clinical application. Explainable Artificial Intelligence (XAI) has become an essential area that can eliminate this gap by ensuring that AI decisions can be clearly read and trusted by clinicians. This systematic review is based on the results of recent research (20192025) on the use of ML/XAI in breast cancer detection that covers imaging and genomic and clinical data. We also review methodological trends, measure of performance, interpretability methods and clinical integration issues. We have seen in our review that there is a distinct transition to human-centered interpretable systems with ensemble methods, hybrid AI approaches, and visual analytics in the lead no longer performance-driven models. It is also through these studies that we find significant gaps in research, and what should be pursued in the future is to use a combination of data modalities, like mammography and patient history and/or genomic markers, to ensemble the best approaches towards accuracy and contextual explanation. Moreover, as noted in the review, post-hoc explanation methods such as SHAP and LIME are still predominant, but the focus on constructing naturally understandable models of clinical safety is also increasing. The main challenges remain the standardization of evaluation measures to explain, the computational complexity of the complex XAI models, and the necessity to develop and test XAI in clinical contexts to promote further uptake and patient outcomes through essential clinician-in-the-loop validation and evaluation.This review concludes that in the future, it is necessary to focus on making models inherently interpretable, develop XAI in a clinician-in-the-loop and conduct robust trials to help transition XAI to reliable clinical use and promote more adoption and better patient outcomes.