Explainable Artificial Intelligence for Medical Image Diagnosis: A Review

Incorporating learning into diagnostic radiology has significantly enhanced precision in X-ray, CT, MRI and ultrasound imaging; however, the "black-box" characteristic of these models continues to be the main barrier to broad clinical implementation. This structured review offers a clinically focused classification of Explainable Artificial Intelligence (XAI) methods tailored for the four leading imaging modalities accounting for over ۸۵% of worldwide diagnostic procedures. A new three-dimensional framework is proposed that categorizes XAI approaches based on (۱) the degree of interpretability (pixel, feature, concept, decision) (۲) reliance on the model (specific versus agnostic) and (۳) timing of explanation (post-hoc versus intrinsic). Key methods (Grad-CAM, SHAP, attention mechanisms, prototype-based networks and Bayesian counterfactuals) are positioned within this classification using modality-examples drawn from CheXpert, BraTS, LIDC-IDRI and BUSI datasets. Practical case examples in brain tumor segmentation (MRI), lung disease identification (X-ray/CT) and breast cancer classification (ultrasound/mammography) show that explanations at both the concept and decision levels reliably produce diagnostic odds ratios above ۱۴۰, decrease procedures by ۱۹-۳۱% and enhance radiologist confidence by as much as ۳۴%. A three-layer clinical evaluation framework, along with metrics, is suggested to unify the evaluation of explanation quality. Current challenges and five actionable future directions (federated XAI benchmarks, edge-deployable real-time systems, multimodal integration, counterfactual simulation engines, and legally auditable human-in-the-loop workflows) are outlined to accelerate trustworthy AI deployment in radiology.

کلیدواژه ها:

XAI ، Medical Imaging ، X-ray ، CT ، MRI