Points-Supervised Fundus Vessel Segmentation via Shape Priors and Contrastive Learning.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Fully supervised fundus vessel segmentation requires extensive pixel-wise annotations, which are costly and time-consuming.
• Weak annotations simplify labeling but limit comprehensive information learning.
• Pseudo-labeling methods can be hindered by false positive predictions, negatively impacting training.

Purpose of the Study:

• To introduce point annotations as a cost-effective alternative for fundus vessel segmentation.
• To propose the Points-based Vessel segmentation Network (PVN) for enhanced segmentation accuracy.
• To balance annotation cost with supervision information quality.

Main Methods:

• Developed the Points-based Vessel segmentation Network (PVN) utilizing point annotations.
• Incorporated Point Activation Maps to learn vessel shape priors as soft supervision, mitigating pseudo-label noise.
• Designed a novel contrastive learning method (pixels-and-regions-mixed) to learn discriminative features.

Main Results:

• PVN demonstrated superior performance on multiple fundus image datasets compared to other point-supervised methods.
• Achieved excellent segmentation performance even with as little as 1% annotated pixels.
• The method proved flexible and easily integrated with other frameworks.

Conclusions:

• Point annotations are effective for fundus vessel segmentation, significantly reducing annotation effort.
• PVN offers a novel and efficient approach to fundus vessel segmentation with minimal supervision.
• This work pioneers the use of point annotations in this domain.