Predicting High Coronary Artery Calcium Score From Retinal Fundus Images With Deep Learning Algorithms.
Level 3 - non-randomized controlled study
Retrospective cross-sectional diagnostic algorithm validation study using concurrent CT reference standards.
PubMed 33184590 · doi:10.1167/tvst.9.2.28
What was done
Researchers developed and evaluated an Inception-v3 deep learning algorithm to predict high coronary artery calcium scores (CACS) from retinal fundus photographs. The dataset consisted of individuals who had same-day bilateral fundus imaging and coronary computed tomography scans. The algorithm was evaluated using 5-fold cross-validation across different CACS thresholds comparing high CACS against a CACS of 0, using both unilateral and bilateral images. Ablation studies with vessel-inpainted and fovea-inpainted images were performed to determine which anatomical features drove algorithm performance.
What was found
In a dataset of 44,184 images from 20,130 individuals, the algorithm discriminating between CACS of 0 and CACS > 100 achieved an AUROC of 82.3% (95% CI, 79.5%–85.0%) with unilateral images and 83.2% (95% CI, 80.2%–86.3%) with bilateral images. Diagnostic performance improved with higher CACS cutoffs up to a plateau at CACS > 100. Occlusion of the fovea reduced the AUROC, and occlusion of retinal vessels reduced AUROC further.
Why it matters
This demonstrates that retinal fundus photography combined with deep learning can serve as a radiation-free, accessible screening tool to identify individuals with clinically significant coronary calcification.
Limits
Performance was evaluated only via internal 5-fold cross-validation without external cohort validation. The classification task was framed as extreme categorization (CACS > 100 vs. CACS = 0) rather than continuous risk prediction or classification across intermediate calcium scores (1–100). The abstract does not report participant demographics, clinical sensitivity/specificity, or the clinical utility of the model compared to traditional cardiovascular risk calculators.
Cited by
- supports AI analysis of retinal images can predict coronary artery calcium scores, stroke risk, and heart disease.