Source-linked AI summary
CholecSeg8k: A Semantic Segmentation Dataset for Laparoscopic Cholecystectomy Based on Cholec80
W. -Y. Hong, C. -L. Kao, Y. -H. Kuo, J. -R. Wang, W. -L. Chang, C. -S. Shih
TL;DR
Limited annotated endoscopic data restricts development of computer-assisted surgery algorithms, including image-based localization. The paper constructs CholecSeg8K from Cholec80, annotating selected laparoscopic cholecystectomy frames at pixel level, and reports class-distribution characteristics of the resulting dataset.
Problem
Researchers lack broadly available, well-labeled endoscopic images for training semantic-segmentation and related prediction models needed in computer-assisted surgery.
Method
The authors extract and annotate selected Cholec80 laparoscopic cholecystectomy frames, defining thirteen classes focused on surgical anatomy with broader coverage for some tissue categories.
Results
8,080 frames from 17 Cholec80 video clips form CholecSeg8K, with pixel annotations across thirteen classes and reported class-pixel ratios.
Takeaways & Limitations
CholecSeg8K provides an open annotated endoscopic dataset for developing semantic-segmentation models and supporting computer-assisted surgery research.
Takeaways & Limitations
Class distributions are imbalanced, with liver and abdominal wall each comprising about 30% of annotated pixels while some tool classes occupy less than 1%.
Abstract
from arXiv · showhide
Computer-assisted surgery has been developed to enhance surgery correctness and safety. However, researchers and engineers suffer from limited annotated data to develop and train better algorithms. Consequently, the development of fundamental algorithms such as Simultaneous Localization and Mapping (SLAM) is limited. This article elaborates on the efforts of preparing the dataset for semantic segmentation, which is the foundation of many computer-assisted surgery mechanisms. Based on the Cholec80 dataset [3], we extracted 8,080 laparoscopic cholecystectomy image frames from 17 video clips in Cholec80 and annotated the images. The dataset is named CholecSeg8K and its total size is 3GB. Each of these images is annotated at pixel-level for thirteen classes, which are commonly founded in laparoscopic cholecystectomy surgery. CholecSeg8k is released under the license CC BY- NC-SA 4.0.
I. INTRODUCTION
Endoscopic computer-assisted systems need reliable image-based localization, but existing localization approaches can require external infrastructure and well-labeled endoscopic images are scarce. The paper introduces an open semantic-segmentation dataset of annotated Cholec80 frames to support such research.
- Endoscopic procedures depend heavily on operator proficiency, while unsteady hand control and organ motion can affect image analysis and operations.
- Many computer-assisted systems use external magnetic or radio-based sensors, but these are prone to errors and impractical for sub-centimeter localization accuracy.
- SLAM localizes cameras from images without infrastructure, but endoscopic navigation requires accurate semantic segmentation to assess the relevant image content.
- Semantic segmentation and object detection require well-labeled endoscopic images, which are not widely available.
- The authors construct an open semantic-segmentation dataset intended to support medical and computer-vision communities developing prediction models.
- 8,080 frames from 17 Cholec80 video clips are annotated with thirteen pixel-level classes, with Figure 1 illustrating a raw image and its color mask.
II. DESCRIPTION OF CHOLECSEG8K DATASET
CholecSeg8K organizes pixel-level annotations from selected Cholec80 videos into a structured dataset focused on cholecystectomy-relevant classes. Its class distribution is imbalanced, and each frame includes multiple mask representations.
- Dataset organization: The dataset uses selected Cholec80 cholecystectomy videos as its source and organizes extracted data into a two-level directory tree.First-level directories correspond to source video clips; secondary directories store groups of 80 images.
- Annotation classes: Each image is represented by 13 classes, including organs, tissues, surgical instruments, and background.The classes target liver and gallbladder recognition while also providing broader coverage through gastrointestinal tract and liver ligament categories.
- Class distribution: The dataset is class-imbalanced: liver and abdominal wall each comprise almost 30% of annotated pixels, while gallbladder comprises 10%.Not all classes appear in every frame, and the authors caution that imbalance may impair training without care.
- Class distribution: L-hook electrocautery and grasper occupy less than 1% of annotated pixels because their physical sizes are relatively small.These tools are frequently used despite their low pixel proportions.
- Mask formats: Each frame has three masks: a color mask for visualization, an annotation mask, and a watershed mask for programming.The watershed mask assigns each annotated pixel the same class ID across its three color channels.
- Mask formats: The dataset stores raw images and all three masks as four PNG files per item.The supplied passages describe raw image, color mask, annotation mask, and watershed mask as the four files.
III. EXAMPLES OF ANNOTATED DATA
The paper illustrates CholecSeg8K annotations through paired raw endoscopic images and color masks, ranging from simple scenes to images containing many anatomical and instrument classes. The examples also show ambiguity at poorly illuminated image boundaries.
- Example presentation: The examples pair raw Cholec80 images on the left with CholecSeg8K color masks on the right.Figures 5–7 present representative extracted images and their class-ID annotations.
- Simple example: Figure 5 shows a simple scene containing gallbladder, liver ligament, fat, and part of the liver.The liver ligament is represented as one class rather than separate classes for individual ligament types.
- Complex examples: Figure 6 shows a more complicated image containing 11 classes, including liver, gallbladder, grasper, and L-hook electrocautery.The example demonstrates simultaneous annotation of anatomical structures and instruments.
- Annotation ambiguity: Low brightness makes annotations near the field-of-view edge difficult for human experts, illustrated by the ambiguous top edge of a grasper.Such ambiguous cases remain in the dataset because its stated aim is segmenting human organs.
IV. ACCESS TO THE DATASET
CholecSeg8K is freely available through Kaggle under the CC BY-NC-SA 4.0 license, with usage instructions provided on the dataset page.
- Availability: CholecSeg8K is freely available on Kaggle under the CC BY-NC-SA 4.0 license.The dataset page also provides instructions for using the data.