Interpretable polyp classification via end-to-end Concept Bottleneck Models with vision-language concept alignment

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Most AI models for colonoscopy lack transparency, hindering clinical adoption, particularly for sessile serrated lesions (SSLs).
• Concept Bottleneck Models (CBMs) offer a solution by making AI predictions interpretable through human-understandable concepts.

Purpose of the Study:

• To develop and evaluate an end-to-end CBM for colonoscopic polyp classification.
• To provide concept-level explanations aligned with clinical standards for AI-driven polyp detection.

Main Methods:

• Defined 62 clinical concepts from established classification systems (NICE, JNET, Paris).
• Utilized BiomedCLIP for pseudo concept labeling and integrated a concept projection layer with an EfficientNet-B3 backbone.
• Trained the model using cross-entropy and mean squared error concept alignment, evaluated on 508 colonoscopic images.

Main Results:

• The CBM achieved a macro F1 score of 95.20%, comparable to a black-box baseline (94.97%).
• No significant performance difference was observed between the CBM and the baseline.
• Concept ablation indicated distributed concept utilization, with no single concept significantly impacting performance.

Conclusions:

• An end-to-end CBM successfully provided interpretable polyp classification with performance on par with black-box models.
• Concept alignment may function as a regularization technique, warranting further investigation.
• This framework empowers clinicians with concept-level insights for AI-assisted colonoscopy.