WSPolypNet: Weakly Supervised Polyp Localization in Colonoscopy Videos

WSPolypNet is a weakly supervised framework that localizes polyps in colonoscopy videos using only video-level labels, avoiding costly frame-level annotations. It employs a 3D CNN to generate class activation maps, enhances them with a multi-view strategy, and refines the results with MedSAM2 segmentation. The method achieves higher CorLoc scores—up to 47.80% at IoU 0.3—and a recall of 94.51%, especially improving detection of small polyps.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.

• computer-vision