MICA: Towards Explainable Skin Lesion Diagnosis via Multi-Level Image-Concept Alignment

Multi-level explanations improve both accuracy and interpretability: MICA aligns medical images with clinical concepts at multiple levels (image, token, and concept), allowing the model to achieve high diagnostic performance while remaining interpretable—addressing a key challenge where explainability often comes at the cost of accuracy.

Doctors can intervene in AI decisions: Unlike black-box systems, MICA enables clinicians to modify the model's reasoning by adjusting concept-level inputs, giving medical professionals direct control over how the AI reaches its conclusions rather than accepting opaque recommendations.

AI explanations use human-understandable concepts: The system provides both visual and textual explanations grounded in clinically-relevant concepts rather than abstract patterns, making it easier for doctors to understand why the AI flagged a particular skin lesion and verify if the reasoning is medically sound.

Requires fewer labeled examples to train effectively: The method demonstrates improved "label efficiency," meaning it can learn to diagnose skin lesions accurately with less annotated training data—a practical advantage given the time and cost of obtaining medical expert annotations.

Tested on real clinical data: Validation across three skin image datasets suggests the approach has been evaluated on realistic medical imaging scenarios, supporting its potential applicability in actual clinical settings where trustworthiness is essential.