Medical imaging models improved with concept-based debugging tools
Beyond Explanation: Debugging Medical Imaging Models via Concept Intervention
Computer Vision and Pattern RecognitionMachine Learning
Summary
Medical imaging models can be hard to understand because they work like 'black boxes,' making it difficult to find mistakes. The authors created a tool that helps doctors and engineers look inside these models by focusing on meaningful concepts like specific medical features. This tool lets users test if certain concepts really cause the model's decisions or if they just appear connected by accident. They tested this approach on ultrasound and chest X-ray images, showing that it helps find and fix problems, sometimes even making predictions more accurate.
What this means in practice
- •For clinical ai developers: Diagnose and fix errors in medical imaging AI models by testing key clinical concepts and improving model accuracy with targeted retraining.
- •For radiology software teams: Create more interpretable medical imaging tools that clinicians can trust by integrating concept-level explanations and corrections into AI workflows.
Authors
Samrajya Thapa, Daniel J. Quest, Timothy L. Kline, Carrie L. Langstraat, Emanuel C. Trabuco, Wei Le
Abstract
Medical imaging models often operate as black boxes, limiting interpretability and systematic debugging. We introduce an easy-to-use, plug-and-play framework for concept-based interpretation and model refinement. By aligning a single-modality encoder to BioMedCLIP, we construct a Concept Bottleneck Model (CBM) that enables concept-level interventions. These interventions allow us to isolate causal versus spuriously correlated concepts, validate insights with domain experts, and generate counterfactual samples for targeted fine-tuning. We evaluate our framework on a Mayo Clinic ultrasound dataset and the CheXpert 5x200 chest X-ray dataset. Results demonstrate that concept intervention enables reliable model diagnosis while maintaining, and occasionally improving predictive performance via guided fine-tuning. Our findings highlight the practical value of this framework for controlled, interpretable refinement of clinical deep learning models.