Predicting High Coronary Artery Calcium Score From Retinal Fundus Images With Deep Learning Algorithms

Area of Science:

• Ophthalmology
• Cardiology
• Artificial Intelligence

Background:

• Coronary artery calcium (CAC) scoring is crucial for cardiovascular risk assessment but typically requires computed tomography (CT) scans.
• Developing non-invasive, radiation-free screening methods for high CAC is a significant clinical need.

Purpose of the Study:

• To evaluate the efficacy of deep learning algorithms in predicting high CAC scores from retinal fundus images.
• To establish retinal fundus imaging as a potential inexpensive and radiation-free alternative for CAC screening.

Main Methods:

• Deep learning models (inception-v3) were trained and validated using a dataset of 44,184 retinal fundus images from 20,130 individuals.
• Performance was assessed by distinguishing high CAC scores (e.g., >100) from no CAC (0) using both unilateral and bilateral fundus images.
• Ablation studies involving vessel and fovea inpainting were conducted to identify key image features.

Main Results:

• The deep learning algorithm achieved an area under the receiver operating curve (AUROC) of 83.2% for predicting high CAC using bilateral fundus images.
• Discrimination performance improved with bilateral imaging and showed a plateau for CAC thresholds above 100.
• Inpainting retinal vasculature or fovea reduced AUROC, indicating these areas are important for prediction.

Conclusions:

• Deep learning algorithms can effectively recognize visual patterns in retinal fundus images associated with high CAC.
• Utilizing bilateral retinal images enhances the prediction of coronary artery calcium scores.
• Retinal fundus imaging holds promise as a tool for non-invasive cardiovascular risk screening.