Clinical AI begins with a practical constraint: reliable ground truth is scarce and expensive. Every segmentation mask, diagnostic label, or expert judgment requires limited specialist time. This makes it essential to build models that can learn reliably from limited, weak, and imperfect supervision. But predictive performance alone is not enough in high-stakes clinical settings. Clinicians also need to understand what evidence supports a model's decision and where that decision may fail.

Explanation, however, is only the beginning. A clinician should be able to question, challenge, and correct an AI system, rather than simply accept or reject a one-way prediction. This turns explainability into contestability, where human expertise can directly intervene in the decision process and the AI can revise, defer, or acknowledge competing evidence.

The final step is to make those interactions persist. Clinical knowledge changes, patient populations shift, and new failure cases continually emerge. Expert corrections should therefore do more than resolve a single decision: they should become feedback that improves future behavior. My research follows this progression, from data-efficient learning, to explainable and contestable decision-making, and ultimately to self-evolving AI systems that improve through structured human feedback.