SimCol3D - 3D reconstruction during colonoscopy challenge

Area of Science:

• Medical Imaging
• Computational Geometry
• Artificial Intelligence

Background:

• Colorectal cancer is a global health concern, with colonoscopy being a key screening method.
• Navigating endoscopes for polyp detection is challenging, and 3D colon surface mapping could improve detection and training.
• Reconstructing colon anatomy from endoscopic video is technically difficult, and learning-based methods require large datasets.

Purpose of the Study:

• To establish a benchmark dataset (SimCol3D) for data-driven depth and pose prediction in colonoscopy.
• To facilitate research into robust 3D colon reconstruction from endoscopic video.
• To evaluate learning-based approaches for colonoscopy image analysis.

Main Methods:

• The 2022 EndoVis sub-challenge SimCol3D was organized to address colonoscopy data challenges.
• Six international teams participated in three sub-challenges: synthetic depth prediction, synthetic pose prediction, and real pose prediction.
• Methods submitted by participants were evaluated on their performance in depth and pose estimation tasks.

Main Results:

• Depth prediction from synthetic colonoscopy images was found to be a robustly solvable problem.
• Pose estimation during colonoscopy, using both synthetic and real data, remains a significant open research challenge.
• The challenge provided valuable insights into the capabilities and limitations of current learning-based methods.

Conclusions:

• The SimCol3D challenge successfully benchmarked colonoscopy image analysis techniques.
• Depth prediction shows strong potential for improving colonoscopy navigation and mapping.
• Further research is needed to advance pose estimation accuracy for real-world colonoscopy applications.