YoloCurvSeg: You only label one noisy skeleton for vessel-style curvilinear structure segmentation

Area of Science:

• Computer Vision
• Medical Image Analysis
• Machine Learning

Background:

• Weakly-supervised learning (WSL) offers a solution to high annotation costs in image segmentation.
• Existing WSL methods struggle with curvilinear structures and limited data.
• Curvilinear structure segmentation (e.g., vessels, nerves) is crucial in medical imaging.

Purpose of the Study:

• To develop a novel WSL segmentation framework for curvilinear structures.
• To address the challenges of limited supervision and non-star-convex shapes in segmentation.
• To enable accurate segmentation with minimal annotations.

Main Methods:

• Proposed YoloCurvSeg framework utilizing image synthesis.
• Background generation via inpainting dilated skeletons.
• Foreground generation using Space Colonization Algorithm and contrastive learning synthesizer.
• Training a segmenter with synthetic and unlabeled data.

Main Results:

• YoloCurvSeg significantly outperforms state-of-the-art WSL methods on four datasets.
• Achieved >97% of fully-supervised performance with minimal (0.03%-1.40%) annotations.
• Demonstrated effectiveness on OCTA500, CORN, DRIVE, and CHASEDB1 datasets.

Conclusions:

• YoloCurvSeg effectively segments curvilinear structures using sparse annotations.
• The image synthesis approach overcomes limitations of existing WSL methods.
• This framework drastically reduces annotation effort while maintaining high segmentation accuracy.