Polyp retrieval drops the hand-drawn boxes: two examples beat most methods given ten
Synopsis
Comparing polyps in colonoscopy depends on complementary views taken from different angles and imaging conditions, yet recent contrastive methods still build polyp tracklet representations from manually annotated bounding boxes. An annotation-free framework constructs tracklet views directly from raw colonoscopy videos and runs temporal self-supervised learning on full-procedure videos, including frames with no visible polyp. On REAL-Colon it beats both self-supervised and supervised baselines on tracklet retrieval and re-identification, and in few-shot polyp retrieval, a task closer to clinical use, two examples surpass most competing methods given ten.
Interpretation
Learning polyp tracklet representations no longer requires manually annotated bounding boxes, and training moves from selected clips to full-procedure videos, including frames without visible polyps. Earlier contrastive methods learned tracklet representations from temporal cues but still started from manually annotated bounding boxes. On REAL-Colon, the annotation-free method outperforms both self-supervised and supervised baselines on tracklet retrieval and re-identification.
The method cuts tracklet views directly from raw colonoscopy video and applies temporal self-supervised learning to those views, removing the annotation step entirely. This is annotation-free representation learning: the video's own temporal structure supplies the training signal instead of hand-drawn boxes. Training and evaluation run on full-procedure videos, which include frames with no visible polyp.
Few-shot polyp retrieval is introduced as a task closer to clinical use: a few observations of a target polyp are used to retrieve its remaining tracklets. The number of examples required drops to single digits compared with the usual setting. With only two examples, the method surpasses most competing methods that are given ten.
Perspective
The demonstrated scope is tracklet retrieval, re-identification and few-shot polyp retrieval on REAL-Colon, with training and evaluation on full colonoscopy videos; the material also states that a modest number of unlabeled procedures suffices for competitive performance, making on-site model development more practical.
The material describes the effect of a modest number of unlabeled procedures as competitive performance, without stating a threshold for how many procedures are needed or giving specific figures for each task.