Comparative validation of multi-instance instrument segmentation in endoscopy: Results of the ROBUST-MIS 2019...

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Robotics

Background:

• Accurate intraoperative tracking of laparoscopic instruments is crucial for computer and robotic-assisted surgeries.
• Existing methods struggle with robustness (e.g., blood, smoke) and generalization across different surgical procedures and institutions.

Purpose of the Study:

• To benchmark the robustness and generalization of algorithms for medical instrument segmentation and detection in endoscopic videos.
• To address limitations in current automated surgical instrument tracking methods.

Main Methods:

• Organized the Robust Medical Instrument Segmentation (ROBUST-MIS) challenge, an international benchmarking competition.
• Utilized a dataset of 10,040 annotated images from 30 surgical procedures across three surgery types.
• Evaluated algorithms on binary segmentation, multi-instance detection, and multi-instance segmentation tasks with increasing domain gaps.

Main Results:

• Algorithm performance significantly degraded as the domain gap between training and testing data increased.
• The best-performing algorithms achieved high average detection and segmentation quality.
• The challenge confirmed that generalization remains a significant hurdle for current methods.

Conclusions:

• Robustness and generalization are critical challenges for medical instrument tracking in endoscopic surgery.
• Future research should focus on improving the detection and segmentation of challenging instrument features (small, crossing, moving, transparent).
• Benchmarking efforts like ROBUST-MIS are vital for advancing surgical computer vision.