Concept-Based Explainable AI in Retinal Imaging: Core Concepts, Techniques, and Challenges

Abstract

Abstract In high-stakes domains like healthcare, establishing trust in AI-assisted decisions is critical for fostering confidence among experts and researchers. However, standard explanation methods may not fully meet clinical needs, offering raw data, such as saliency maps, rather than human-readable text and auditable, transparent reasoning. To bridge this gap, researchers are increasingly adopting concept-based approaches that express model reasoning through human-readable concepts grounded in expert-backed clinical knowledge. Although this field has advanced rapidly in recent years, retinal C-XAI remains methodologically heterogeneous, with limited agreement on taxonomy, evaluation, and clinical validation. This review identifies recurring reliability gaps in retinal C-XAI and synthesises potential directions for addressing them, including independent concept-level auditing, robustness testing, and clinician-informed validation. As an illustrative example, we discuss how the established SMILE methodology could be adapted for concept-level auditing in retinal C-XAI; this is presented as a candidate research direction rather than a new or clinically validated framework.