Prediction of best-corrected visual acuity (BCVA) from color fundus photography (CFP) using deep learning (DL)

Abstract

Purpose : To predict BCVA from CFP images acquired from patients with neovascular age-related macular degeneration (nAMD).

Methods : We performed a retrospective analysis of CFP images and associated BCVA measurements from the phase 3 MARINA (NCT00056836) and ANCHOR (NCT00061594) trials for nAMD. Using the Inception-ResNet-v2 convolutional neural network, a DL regression model was developed to predict BCVA at the concurrent visit from CFP images. A binary classification model was also developed to predict BCVA of < 69 letters (Snellen equivalent of < 20/40) at the concurrent visit from CFP images. Models were trained and tuned via 5-fold cross-validation on 36,541 images from 707 patients in MARINA and tested on an external validation test set of 33,591 images from 413 patients in ANCHOR. Images used for training and testing were acquired from study and fellow eyes at all available visits (screening through month 24). Internal capture fields of F1M, F2, and F3M from left and right stereo views were included, whereas external views of the eye were excluded from the analysis. To remove extraneous information, RGB images were cropped to fit the circular field of view of the CFP and resized to 299 × 299 pixels. To evaluate model performance, the coefficient of determination (R²) was used for regression, and the area under the receiver operating characteristic curve (AUROC) was used for classification. Predictions for each image were averaged to the patient-eye-visit level. To account for the distance at which BCVA was measured, metrics were calculated for the corresponding chart distances of 2 and 4 m. Models were evaluated based on the in-sample cross-validation tuning set (MARINA) and out-of-sample test set (ANCHOR).

Results : In the regression model, R² for predicting BCVA at a chart distance of 2 and 4 m was 0.58 (95% CI, 0.57, 0.60) and 0.60 (95% CI, 0.56, 0.63), respectively, in the test set (Table 1). In the classification model, AUROC for predicting BCVA of < 69 letters at a chart distance of 2 and 4 m was 0.86 (95% CI, 0.85, 0.87) and 0.88 (95% CI, 0.86, 0.89), respectively, in the test set.

Conclusions : Out-of-sample, the identified DL model demonstrates a relatively strong association between the information extracted from the CFP images and visual acuity as measured by BCVA.

This is a 2021 ARVO Annual Meeting abstract.

 

Model performance on tuning (MARINA) and test (ANCHOR) datasets