Deriving Visual Cues from Deep Learning to Achieve Subpixel Cell Segmentation in Adaptive Optics Retinal Images

Area of Science:

• Ophthalmology
• Biomedical Imaging
• Computational Neuroscience

Background:

• Direct visualization of photoreceptor cells is crucial for understanding retinal diseases.
• Current segmentation methods struggle with ambiguous cell boundaries common in adaptive optics (AO) retinal imaging.
• Accurate cell segmentation is essential for monitoring disease progression and potential blindness.

Purpose of the Study:

• To develop an advanced method for segmenting photoreceptor cells in challenging AO retinal images.
• To overcome limitations of existing methods in handling ambiguous cell boundaries and touching cells.
• To enable more precise evaluation of photoreceptor cell morphology for disease monitoring.

Main Methods:

• Developed AOSeg-Net, a multi-channel U-Net model for predicting cell boundary probabilities.
• Incorporated cell centroid and region distribution information to aid segmentation.
• Utilized a region-based level set algorithm with a five-color theorem guarantee for subpixel segmentation of touching cells.

Main Results:

• AOSeg-Net demonstrated superior performance compared to existing approaches on 428 high-resolution retinal images.
• Achieved a Dice coefficient of 84.7% and an average symmetric contour distance of 0.59 micrometers.
• Significantly outperformed the comparative method (Dice: 78.4%, Contour Distance: 0.80 micrometers).

Conclusions:

• AOSeg-Net provides a robust and accurate solution for photoreceptor cell segmentation in AO retinal imaging.
• The method effectively addresses challenges posed by low contrast and densely packed cells.
• This advancement holds significant potential for improving the diagnosis and management of retinal diseases.