From black-box to clinical reasoning: CARE-MD a novel framework for Self-eXplainable skin lesions

Deep learning models have achieved remarkable diagnostic accuracy in medical imaging, yet their lack of interpretability limits clinical trust and deployment. This study presents CARE-MD (Clinical Algorithm for Reasoning-Enhanced Medical Diagnosis), a self-explainable framework that mirrors the clinical reasoning process to provide transparent, concept-driven, and case-referable diagnosis of skin lesions.

CARE-MD follows a four-phase reasoning structure: Observation, Interpretation, Reference, and Validation. In Phase 1, attention-guided localization isolates the lesion region using a modified U-Net backbone. Phase 2 employs a Concept Bottleneck Model (CBM) model to predict human-interpretable dermatological concepts. Phase 3 introduces prototype-based reasoning to compare latent features with previously learned prototypes for case-level interpretability. Phase 4 validates explanations through consistency analysis between attention maps, concept activations, and clinician-annotated lesion regions. The framework was evaluated on two publicly available datasets-ISIC 2018 and HAM10000-using metrics such as accuracy, precision, sensitivity, specificity, Dice coefficient, and explanation consistency score.

CARE-MD achieved 87.2% accuracy, 86.8% precision, 85.6% sensitivity, and 88.1% specificity on the ISIC 2018 dataset. Cross-dataset testing on HAM10000 showed stable performance with minimal decline across all metrics, indicating strong generalization capability. The attention-guided module improved Dice score by 3% compared with baseline Attention U-Net, while the prototype referencing module enhanced explanation consistency by 6%. Qualitative results confirmed that CARE-MD produces coherent alignment between salient lesion regions, predicted dermatological concepts, and prototype-based references.

CARE-MD provides a structured approach toward aligning deep learning predictions with clinical reasoning by embedding interpretability into each diagnostic phase. While the framework shows improved interpretability and competitive performance relative to the evaluated baselines, it introduces additional architectural complexity and relies on concept-level supervision. These findings suggest that CARE-MD represents a promising direction for self-explainable medical diagnosis, while also highlighting inherent trade-offs between interpretability, model complexity, and annotation requirements. A notable strength of CARE-MD is its unified integration of multiple interpretability mechanisms within a clinically motivated reasoning structure.