In-hoc Concept Representations to Regularise Deep Learning in Medical Imaging

Medical AI models often fail in real-world conditions: The paper demonstrates that deep learning models trained on medical images perform well in controlled settings but struggle when encountering new data or conditions they haven't seen before, frequently relying on misleading patterns rather than clinically meaningful features.

A new method improves reliability without extensive labeling: LCRReg uses a small auxiliary dataset to teach models about clinically relevant concepts, requiring no concept labels in the main training data—making it practical for medical institutions that lack extensive annotated datasets.

The approach significantly reduces reliance on spurious correlations: Testing showed that LCRReg substantially improved model robustness when irrelevant visual artifacts were introduced, and enhanced performance on diabetic retinopathy classification tasks under both synthetic and real-world distribution shifts.

The solution is practical and adaptable: Unlike some competing approaches, LCRReg works with existing CNN architectures without requiring architectural changes, making it easier for medical imaging teams to implement across different systems and applications.

Interpretability and safety go hand-in-hand: By guiding models to use semantically meaningful medical concepts rather than hidden spurious patterns, this approach makes AI decisions more understandable to clinicians while simultaneously improving their reliability—addressing both transparency and safety concerns in medical AI deployment.