Explainable Artificial Intelligence in Medical Image based Diagnosis: A Comprehensive Study

Abstract

Artificial intelligence (AI) has made remarkable contributions to medical imaging, enhancing the accuracy of the diagnosis and its efficiency; the challenge to the adoption in clinical settings has been the lack of clarity of the deep learning models. Explainable Artificial Intelligence (XAI) is a solution to this dilemma, as it provides easy-to-understand and understandable information about the way the model arrived at its decisions. The review summarizes the latest advances in XAI in the radiology, ophthalmology, pathology, neurology, and oncology domains as well as the conventional machine learning attribution algorithms; saliency-based deep learning algorithms, including Grad-CAM, SHAP, and LIME; and multimodal models. Its main clinical uses are cancer detection, classification of Alzheimer's disease, COVID-19 diagnosis, and evaluation of retinal disease. There are both quantitative (fidelity, stability, and localization) and qualitative (interpretability and radiologist trust) evaluation strategies. Although there is significant improvement, there exist dataset quality issues, interpretability-accuracy trade-offs, and generalizability issues across clinical settings. Furthermore, a comparative study with the available literature is done in a systematically organized way the evaluation parameters are clearly developed to evaluate objectively the strengths of our approach. This comparative parameter-based parameter shows the advances of our framework with respect to generalizability, explainability quality, and clinical relevance. Our study makes a more effective justification of its usefulness than the previous methods since the comparison is based on quantifiable measures.