Evaluating robustness and interpretability of AI models for disease detection