Core Concepts
Including optical flow improves surgical instrument segmentation in nnU-Net.
Abstract
The study explores integrating optical flow (OF) into the nnU-Net framework to enhance surgical instrument segmentation. OF provides temporal information, benefiting classes with high movement. Different OF representations were tested, showing improvements in detection accuracy. Limitations include the inability of OF to differentiate instruments from tissues and constraints on augmentations due to framework restrictions. The study highlights the potential of OF for improving semantic segmentation results.
Stats
The Cholec80 dataset contains 80 videos captured at 25 fps.
CholecSeg8k dataset includes 8080 frames of laparoscopic cholecystectomy cases.
RGB baseline achieved a Mean DC of 53.97%.
Quotes
"Results showed that the use of OF maps improves the detection of classes with high movement."
"With this new input, the temporal component would be indirectly added without modifying the architecture."
"The major contribution is to demonstrate the ease of including OF in nnU-Net architecture."