Prediction of activity in eyes with macular neovascularization due to age-related macular degeneration using deep learning

Abstract Background: To evaluate several deep learning algorithms to detect activity of macular neovascularization (MNV) using en face optical coherence tomography angiography (OCTA) images. Methods: Choriocapillaris en face OCTA 6x6 mm images from eyes with neovascular AMD imaged with the RTvue-XR Avanti SD-OCTA (Optovue) device were included in this retrospective analysis. Multiple machine learning models were trained to classify the presence of MNV activity by OCTA imaging, using the presence of fluid on the structural OCT as the ground truth evidence for activity. Specifically, a five-fold cross-validation was applied to assess the different models’ performance. The performance of the various models was evaluated by using the ROC and its area under the curve (AUC). A power analysis was used to assess the effect of sample size on models’ performance. Results: 637 en face OCTA images from 97 patients were included in this analysis. We observed that en face OCTA appearance of the MNV lesion was a poor predictor of disease activity. The algorithms used did not demonstrate good performance: Resnet (0.51 [0.36,0.65]), simple CNN (0.54[0.39,0.69]), LR+PCA (0.53[0.41,0.64]), Resnet-Scratch (0.48[0.34,0.62]). We performed a power analysis to examine changes in performance as the sample size increased and saw no positive trend, suggesting that a substantial improvement in performance would not be expected with a larger sample. Conclusions: We observed that en face OCTA images alone are poor predictors of MNV lesion activity. This suggests that strong biomarkers of disease activity may not be encoded within the en face OCTA image.