OCTAve: 2D En Face Optical Coherence Tomography Angiography Vessel Segmentation in Weakly-Supervised Learning With...

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Deep learning for microvasculature annotation in OCTA requires extensive, costly manual annotation.
• Weak annotation, like centerline labeling, offers a cost-effective alternative but challenges segmentation quality.

Purpose of the Study:

• To develop a novel weakly-supervised learning method (OCTAve) for microvasculature segmentation in OCTA using centerline annotations.
• To improve segmentation accuracy and reduce artifacts compared to existing methods.

Main Methods:

• Utilized scribble-based weakly-supervised learning, adapting it for centerline annotations.
• Introduced a novel self-supervised deep supervision method based on Kullback-Liebler divergence to enhance segmentation performance.

Main Results:

• OCTAve demonstrated superior quantitative and qualitative segmentation results on public datasets (ROSE, OCTA-500) compared to baseline and naive methods (p < 0.001 for Dice's coefficient).
• Achieved results comparable to fully supervised methods with an average performance drop of less than 10%.

Conclusions:

• Weakly-supervised learning, augmented with self-supervised deep supervision, effectively addresses the challenge of microvasculature segmentation with limited annotations.
• The proposed OCTAve method offers a practical solution to reduce annotation costs and expert workload in OCTA analysis.