Segmenting Surgical Tools Robustly On Non-adversarial Generated Corruptions, the EndoVis'24 challenge at MICCAI 2024, organized by the ARCADE Lab at Johns Hopkins University.

Challenge website · Paper (arXiv:2407.11906) · Code and baselines (CaRTS)

Overview

SegSTRONG-C asks how robust surgical tool segmentation is to realistic, non-adversarial corruptions. The task is binary segmentation of robot-assisted surgical tools against the tissue background, in stereo endoscopic images that can also be treated as video.

The data is paired by robotic replay: the same tool motion is recorded clean and again under smoke, bleeding, low brightness and an alternative background. Every condition of a subsequence therefore shows the same instant at each frame index and shares one ground-truth mask. Models train on clean data only and are evaluated on the corrupted versions.

Split	Sequence / subsequences	Conditions	Stereo frames	Images (incl. masks)	Size
train	3/0, 3/2, 4/0, 4/1, 4/2, 5/0, 5/2, 7/0, 7/1, 8/1, 8/2	regular	3,300	13,200	22.5 GB
val	1/0, 1/1, 1/2	regular, smoke, blood, low_brightness, bg_change	900 per condition	10,800	30.1 GB
test	9/0, 9/1, 9/2	regular, smoke, blood, low_brightness, bg_change	900 per condition	10,800	29.9 GB

Every subsequence has 300 frames, each with a left and a right view.

The challenge used part of the val and test folders:

Validation: regular and bg_change. That's 3,600 frames: 3 × 300 × 2 views × 2 conditions.
Ranked test: smoke, blood and low_brightness. That's 5,400 frames: 3 × 300 × 2 views × 3 conditions.

The release includes every condition for both splits.

Layout

SegSTRONGC_train/<seq>/<subseq>/regular/{left,right}/<frame>.png
SegSTRONGC_train/<seq>/<subseq>/ground_truth/{left,right}/<frame>.png
SegSTRONGC_val/val/<seq>/<subseq>/{regular,smoke,blood,low_brightness,bg_change}/{left,right}/<frame>.png
SegSTRONGC_val/val/<seq>/<subseq>/ground_truth/{left,right}/<frame>.png
SegSTRONGC_test/test/<seq>/<subseq>/{regular,smoke,blood,low_brightness,bg_change}/{left,right}/<frame>.png
SegSTRONGC_test/test/<seq>/<subseq>/ground_truth/{left,right}/<frame>.png

Images: RGB PNG, 1920×1080, named 0.png … 299.png.
Masks (ground_truth): 8-bit grayscale PNG, 1920×1080, 0 = background and 255 = tool, so divide by 255 to get labels. Every mask contains some tool; tools average 13–20% of pixels, depending on the split.
Pairing: <frame>.png in any condition folder of a subsequence pairs with ground_truth/<side>/<frame>.png of the same subsequence.

The files are the original release, unzipped from SegSTRONGC_{train,val,test}.zip with the folder structure unchanged.

Quick start

from huggingface_hub import snapshot_download
import numpy as np
from PIL import Image

# Download one split only (the full dataset is about 82 GB).
root = snapshot_download("hding2455/SegSTRONGC", repo_type="dataset",
                         allow_patterns=["SegSTRONGC_train/*"])

frame = f"{root}/SegSTRONGC_train/3/0"
image = np.array(Image.open(f"{frame}/regular/left/0.png"))           # (1080, 1920, 3) uint8
mask = np.array(Image.open(f"{frame}/ground_truth/left/0.png")) > 0   # (1080, 1920) bool, True = tool

For corrupted evaluation data, use for example allow_patterns=["SegSTRONGC_test/test/9/*/smoke/*", "SegSTRONGC_test/test/9/*/ground_truth/*"].

Evaluation

The challenge ranked entries by the Dice similarity coefficient (DSC) and normalized surface distance (NSD) at several tolerances, aggregated over the three corruption domains: smoke, bleeding and low brightness. Baseline models and evaluation code are in CaRTS.

License

The SegSTRONG-C data is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license (CC BY-NC-SA 4.0):

use it with attribution (cite the paper below);
for non-commercial purposes only;
share derivatives under the same license.

The CaRTS code is licensed separately.

Citation

@misc{ding2024segstrongcsegmentingsurgicaltools,
      title={SegSTRONG-C: Segmenting Surgical Tools Robustly On Non-adversarial Generated Corruptions -- An EndoVis'24 Challenge},
      author={Ding, Hao and Lu, Tuxun and Zhang, Yuqian and others},
      year={2024},
      eprint={2407.11906},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}