Beyond Explanation: Debugging Medical Imaging Models via Concept Intervention

2026-10-06

AI Summary

This study addresses the limited interpretability and systematic debuggability of medical imaging models arising from their black-box nature. We propose a plug-and-play concept intervention framework that constructs a Concept Bottleneck Model (CBM) leveraging BioMedCLIP-based multimodal alignment, enabling concept-level interventions to disentangle causal from spurious associations. Furthermore, expert-validated counterfactual samples are generated to guide targeted fine-tuning. Experiments on ultrasound and chest X-ray datasets demonstrate that this approach achieves reliable model diagnostics while significantly enhancing controllability without compromising predictive performance. By maintaining or improving accuracy alongside improved transparency, this work establishes a novel paradigm for the safe optimization of artificial intelligence in medical imaging.

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.

Problem

• Medical Imaging
• Interpretability
• Model Debugging
• Black Box
• Concept Intervention

Innovation

• Concept Bottleneck Model
• Concept Intervention
• BioMedCLIP
• Counterfactual Samples
• Model Debugging