Key points are not available for this paper at this time.
Artificial intelligence (AI) is increasingly being integrated into radiological workflows. Despite strong diagnostic performance, most deep learning systems operate as "black boxes," limiting transparency, trust, and regulatory acceptance. Explainable artificial intelligence (XAI) aims to address this gap by providing insight into model behavior and aligning predictions with radiological reasoning. This review provides a structured synthesis of the current landscape of XAI in radiology across major subspecialties, including chest imaging, breast imaging, neuroimaging, musculoskeletal imaging, abdominal and cardiovascular imaging, and deep learning–based image reconstruction. Across these domains, a broad range of explainability approaches has been explored, including gradient-based saliency and attribution methods (e.g., Grad-CAM, LRP, Integrated Gradients), perturbation-based techniques (e.g., occlusion, SHAP), concept-based reasoning (e.g., TCAV), example- and prototype-based explanations, and uncertainty quantification. Persistent limitations include methodological unreliability, lack of quantitative validation, dataset bias, hidden stratification, and poor workflow integration. Prospective multicenter evaluations and standardized reporting remain rare. Looking ahead, the field is moving from visually appealing overlays toward robust, task-specific, and workflow-embedded explanations. Promising directions include concept-based and counterfactual reasoning, integration of uncertainty into explanations, radiology-specific vision–language models, and structured reporting standards. By coupling methodological rigor with clinical relevance, XAI can enhance transparency, mitigate bias, and support trustworthy implementation of AI systems in radiology.
Haupt et al. (Wed,) studied this question.