Self-relabeling for noise-tolerant retina vessel segmentation through label reliability estimation.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Deep learning significantly enhances retinal vessel segmentation.
• Accurate ground-truth segmentation maps are crucial for training deep learning models.
• Manual annotations often contain errors, particularly for thin retinal vessels, hindering model performance.

Purpose of the Study:

• To develop a method for automatic and iterative correction of noisy segmentation labels.
• To improve the accuracy of retinal vessel segmentation by addressing label noise during training.

Main Methods:

• A novel approach uses historical predicted label maps from different training epochs.
• The method employs self-supervision to refine predicted labels.
• Noisy supervised labels are dynamically corrected during network training.

Main Results:

• The method demonstrated significant improvements in label map accuracy on DRIVE, STARE, and CHASE-DB1 datasets.
• Achieved 4.0- and 10.7-point improvements on [Formula: see text] and PR metrics for synthetic noise.
• Effectively improved label map quality for pseudo-labeled and manually labeled noise.

Conclusions:

• The proposed method enhances retinal image segmentation performance compared to existing techniques.
• Simultaneous noise correction in label maps leads to superior segmentation outcomes.
• Validates the effectiveness of the automated label correction approach for deep learning in medical imaging.