Predicting branch retinal vein occlusion development using multimodal deep learning and pre-onset fundus hemisection...

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Retinal Vascular Disease Research

Background:

• Branch retinal vein occlusion (BRVO) is a significant cause of visual impairment in adults.
• Predicting BRVO solely from retinal vascular patterns is currently challenging.
• Early detection and prediction of BRVO are crucial for timely intervention and vision preservation.

Purpose of the Study:

• To develop and evaluate a deep learning model for predicting branch retinal vein occlusion (BRVO).
• To assess the efficacy of a multimodal approach combining fundus images and blood vessel (BV) data.
• To identify key retinal vascular features predictive of BRVO onset.

Main Methods:

• A retrospective cohort study utilized hemisection fundus images from patients with and without BRVO.
• A U-net deep learning model segmented retinal structures, including optic discs and blood vessels (BVs).
• Unimodal (fundus or BV only) and a BV-enhanced multimodal models were trained and compared for BRVO prediction.

Main Results:

• The BV-enhanced multimodal deep learning model achieved an area under the receiver operating characteristic curve (AUC) of 0.76.
• The multimodal model demonstrated a prediction accuracy of 68.5% for future BRVO.
• Predictions were concentrated in arteriovenous crossing regions within the retinal vascular arcade.

Conclusions:

• A deep learning model integrating retinal fundus images and blood vessel data shows potential for predicting branch retinal vein occlusion (BRVO).
• The BV-enhanced multimodal approach offers improved predictive performance over unimodal models.
• Further validation with larger, multicenter datasets is necessary to enhance clinical utility and accuracy.