Analyzing Retinal Vessel Morphology in MS Using Interpretable AI on Deep Learning-Segmented IR-SLO Images

Area of Science:

• Ophthalmology
• Neuroscience
• Medical Imaging

Background:

• Multiple sclerosis (MS) causes central nervous system changes, including retinal structural and vascular alterations.
• Current imaging techniques like optical coherence tomography (OCT) and fundus photography lack MS specificity.
• There is a need for non-invasive, MS-specific biomarkers.

Purpose of the Study:

• To investigate Infrared Scanning-Laser-Ophthalmoscopy (IR-SLO) for identifying MS-specific retinal vascular biomarkers.
• To adapt deep learning models for retinal vessel segmentation in IR-SLO images.
• To evaluate the diagnostic potential of extracted vascular features for MS detection.

Main Methods:

• A deep learning model was adapted for segmenting optic disc, optic cup, and retinal vessels in IR-SLO images.
• Pre- and post-processing steps were optimized for segmentation accuracy (Dice coefficients: 91%-97%).
• Clinically relevant features were extracted, and statistical analyses with SHapley Additive exPlanations (SHAP) were performed.

Main Results:

• Vessel fractal dimension, vessel density in zones B and C, vessel intensity, and width were identified as key differentiators.
• These features distinguished multiple sclerosis patients from healthy controls.
• The adapted deep learning model demonstrated high segmentation accuracy.

Conclusions:

• IR-SLO imaging can detect unique retinal vascular biomarkers in MS patients.
• These biomarkers show potential as non-invasive diagnostic markers for multiple sclerosis.
• IR-SLO offers a promising complementary approach to existing diagnostic methods.