Self-supervised disc and cup segmentation via non-local deformable convolution and adaptive transformer.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Ophthalmology

Background:

• Optic disc and cup segmentation is vital for diagnosing ocular conditions.
• Current methods struggle with limited data and capturing global context, leading to suboptimal accuracy.
• Complex anatomical structures and ambiguous boundaries in optic disc/cup images pose significant challenges.

Purpose of the Study:

• To enhance the accuracy of optic disc and cup segmentation.
• To address limitations of data scarcity and insufficient global contextual information.
• To develop a more robust automated diagnostic tool for ocular pathologies.

Main Methods:

• Proposed a novel network architecture incorporating a non-local dual deformable convolutional block for boundary pattern capture.
• Introduced an adaptive K-Nearest Neighbors (KNN) transformation block to extract global semantic context, modifying the vision transformer.
• Implemented a self-supervised training strategy for network initialization to reduce reliance on labeled data.

Main Results:

• The proposed method achieved state-of-the-art performance on the REFUGE dataset.
• Achieved high Intersection over Union (IOU) scores: 0.9577 for the optic disc and 0.8399 for the optic cup.
• Demonstrated superior segmentation accuracy compared to previous state-of-the-art networks.

Conclusions:

• The developed self-supervised strategy and network architecture significantly improve optic disc and cup segmentation accuracy.
• The method effectively handles complex anatomical structures and ambiguous boundaries.
• This advancement holds promise for more precise and efficient automated diagnosis of eye diseases.