Self-Supervised Sequence Recovery for Semi-Supervised Retinal Layer Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Ophthalmology

Background:

• Automated layer segmentation in optical coherence tomography (OCT) is crucial for diagnosing retinal diseases.
• Severe retinal diseases degrade the performance of existing automated segmentation methods.

Purpose of the Study:

• To develop a robust semi-supervised layer segmentation network to address model failures in abnormal retinas.
• To enhance the accuracy of retinal layer segmentation in diseased OCT images.

Main Methods:

• A semi-supervised network utilizing cross-consistency training between different decoders to improve encoder representation.
• A self-supervised sequence prediction branch with a layer spatial pyramid pooling (LSPP) module for multi-scale feature extraction.
• Integration of optical coherence tomography angiography (OCTA) to compensate for disease-induced information loss.

Main Results:

• The proposed method demonstrates more robust segmentation results compared to current supervised methods.
• Achieves advanced segmentation performance exceeding state-of-the-art semi-supervised methods.
• Effectively segments retinal layers even in the presence of severe retinal diseases.

Conclusions:

• The developed semi-supervised network significantly improves the robustness and accuracy of retinal layer segmentation in diseased OCT images.
• The integration of cross-consistency training, self-supervised sequence prediction, and OCTA data offers a promising approach for automated retinal disease diagnosis.