Predicting the generalization of computer aided detection (CADe) models for colonoscopy

Abstract

Abstract Generalizability of AI colonoscopy algorithms is important for wider adoption in clinical practice. However, current techniques for evaluating performance on unseen data require expensive and time-intensive labels. We show that a "Masked Siamese Network" (MSN), trained to predict masked out regions of polyp images without labels, can predict the performance of Computer Aided Detection (CADe) of polyps on colonoscopies, without labels. This holds on Japanese colonoscopies even when MSN is only trained on Israeli colonoscopies, which differ in scoping hardware, endoscope software, screening guidelines, bowel preparation, patient demographics, and the use of techniques such as narrow-band imaging (NBI) and chromoendoscopy (CE). Since our technique uses neither colonoscopy-specific information nor labels, it has the potential to apply to more medical imaging domains.

Keywords

• Image foundational model
• Self-supervision
• Endoscopy
• Computer-aided detection (CADe) system
• Artificial intelligence