Contrastive Learning under Noisy Temporal Self-Supervision for Colonoscopy Videos

Researchers introduced a noise-aware temporal contrastive framework for colonoscopy videos that learns polyp representations without expensive manual labels.

The method is designed to tolerate noisy pairings in real clinical workflows, making self-supervised learning more practical for endoscopy data.

It outperformed prior self-supervised and supervised baselines on multiple downstream tasks, including retrieval, re-identification, size estimation, and classification.

The approach could reduce reliance on expert annotation while improving AI support for colonoscopy analysis.