Robust region encoding and layer attribute protection for the segmentation of retina with multifarious abnormalities

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Retinal layer segmentation is crucial for diagnosing and monitoring various eye diseases.
• Existing segmentation models struggle with the complex abnormalities present in conditions like AMD, DR, and CSC.
• Optical coherence tomography angiography (OCTA) offers complementary vascular information to structural SD-OCT data.

Purpose of the Study:

• To develop a robust retinal layer segmentation model for optical coherence tomography angiography (OCTA) en face projection generation.
• To enhance model performance despite multifarious retinal abnormalities.
• To improve the accuracy and reliability of retinal layer segmentation in diseased eyes.

Main Methods:

• Proposed a robust retinal layer segmentation model incorporating OCTA vascular distribution to supplement SD-OCT structural information.
• Introduced a multitask layer-wise refinement (MLR) module to refine segmentation results layer-by-layer, reducing sensitivity to abnormalities.
• Designed a region-to-surface transformation (RtST) module to convert segmented layer regions into layer surfaces for improved accuracy.

Main Results:

• The model was evaluated on 273 eyes (95 normal, 178 diseased) including AMD, DR, CSC, and CNV.
• Achieved high Dice Similarity Coefficients (DSC) on normal eyes (superficial: 98.92%, deep: 97.48%, outer: 98.87%) and abnormal eyes (superficial: 98.35%, deep: 95.33%, outer: 98.17%).
• Demonstrated state-of-the-art layer segmentation performance compared to existing models.

Conclusions:

• The proposed model exhibits outstanding performance and robustness against retinal abnormalities.
• OCTA modality significantly aids in achieving accurate retinal layer segmentation.
• The developed method is effective for generating en face projections from segmented retinal layers in various retinal diseases.