EndoVis 2024 Challenge · Challenge concluded

SegCol Challenge

Semantic Segmentation for Tools and Fold Edges in Colonoscopy Data

University College London · Surgical Robot Vision · EndoMapper

Overview

SegCol is the first colonoscopy dataset with detailed pixel-level annotations of colon fold edges and three surgical tool classes. Sampled from EndoMapper videos, it captures challenging variation in anatomy, illumination, motion, occlusion and image quality.

The dataset establishes benchmarks for pixel-level segmentation and classification. The challenge has concluded; this page documents the released data, evaluation protocol and published reference baselines.

Dataset

SegCol contains 8,440 RGB frames at 640 × 480 resolution. Frames were sampled in consecutive 40-frame batches at 10–12 FPS. All partitions, including labels hidden during the challenge, are now publicly available.

Partition	Frames	Labels
Training	2,000	Public
Validation	560	Public
Formerly unlabelled training pool	5,320	Public
Test	560	Public
Total	8,440	Released

Annotations

Each single-channel mask encodes background (0), Captivator (127), Forceps (128), Balloon (129) and one-pixel fold edges (255).

Complete SegCol overview showing training, validation, unlabelled and test examples

Benchmark

Tasks

Task 1 — Supervised segmentation. Train on the labelled development set and segment fold edges and three instrument classes on held-out frames.

Task 2 — Active learning. Select a 400-frame annotation subset from the 5,320-frame pool and retrain the reference models under a fixed annotation budget.

Evaluation metrics

The benchmark reports Dice for region overlap, AP across confidence thresholds, ODS and OIS for boundary quality, and CLDice for topology-aware agreement of thin structures.

Published baselines

DexiNed is used for fold edges and DeepLab for the three instrument classes. The Task 1 reference scores are:

Target	Model	ODS	OIS	Dice	AP	CLDice
Folds	DexiNed	0.2461	0.2525	0.2684	0.0894	0.2736
Tool 1	DeepLab	0.0546	0.0721	0.3636	0.1330	0.2440
Tool 2	DeepLab	0.0489	0.0698	0.0930	0.0331	0.0414
Tool 3	DeepLab	0.1214	0.1255	0.8041	0.6515	0.2837

Citation

If you use SegCol in your research, please cite:

@article{ju2024segcol,
  title={SegCol Challenge: Semantic Segmentation for Tools and Fold Edges in Colonoscopy Data},
  author={Ju, Xinwei and Daher, Rema and Caramalau, Razvan and Huang, Baoru and Stoyanov, Danail and Vasconcelos, Francisco},
  journal={arXiv preprint arXiv:2412.16078},
  year={2024}
}

Resources

Complete dataset — images and released annotations in the EndoMapper archive.
Code and evaluation — benchmark implementation, metrics and reference methods.
Challenge archive — the original EndoVis 2024 challenge record.