Screening of Moyamoya Disease From Retinal Photographs: Development and Validation of Deep Learning Algorithms

Area of Science:

• Ophthalmology
• Neurology
• Artificial Intelligence

Background:

• Moyamoya disease (MMD) is a rare cerebrovascular disorder causing abnormal blood vessel formation in the brain.
• MMD diagnosis and progression are challenging due to unpredictable factors and can lead to ocular symptoms.
• Retinal photographs offer potential insights into MMD due to the disease's impact on ocular vasculature.

Purpose of the Study:

• To develop and validate a deep learning model for MMD screening using retinal images.
• To assess the model's ability to predict MMD disease stages from fundus photographs.
• To investigate the utility of retinal photographs as non-invasive biomarkers for MMD.

Main Methods:

• A retrospective observational study analyzed 498 MMD and 3835 healthy retinal photographs.
• A ResNeXt50 deep learning model was trained and validated on preprocessed retinal images.
• Model performance was evaluated using ROC curves, accuracy, sensitivity, and specificity; heatmaps validated feature importance.

Main Results:

• The MMD screening model achieved an Area Under the ROC Curve (AUC) of 94.6% with 89.8% sensitivity and 90.4% specificity.
• MMD stage prediction models demonstrated AUCs of 78% or higher, with Stage 3 performing best at 93.6%.
• Heatmap analysis highlighted the vascular region of the fundus as crucial for MMD prediction.

Conclusions:

• Retinal photographs serve as viable biomarkers for MMD screening and staging.
• Deep learning algorithms can effectively classify MMD disease stages using retinal imaging.
• This AI-driven approach facilitates non-invasive diagnosis and monitoring of Moyamoya disease.