Estimation of the best corrected visual acuity from optical coherence tomography images using deep learning: a validation study

Abstract

Purpose : We previously developed a deep learning algorithm using data from single institution to estimate best-corrected visual acuity (BCVA) from optical coherence tomography (OCT) images. This study aimed to validate the algorithm using data from another institution, either with or without standardization of the OCT images size.

Methods : The algorithm was trained using 2,756 OCT (Atlantis and Triton; TOPCON, Tokyo, Japan) images of 809 eyes of 509 patients, and corresponding logMAR BCVA that were measured on the same day at Jichi Medical University Hospital (Tochigi, Japan) from 2004 to 2018. Atlantis and Triton instruments have the same optical component. For each eye, a maximum of five OCT images including one horizontal and one vertical scans with resolution 992 x 992 were used. Using 10-fold cross validation, models with lowest (0.25) and highest (0.47) standard error were used. Validation study was performed using data from Saitama Medical Center (Saitama, Japan). The resolution was 1143 x 622 due to different scanning mode. The scans were cropped to 992 x 992 and the margin was filled with black for standardization. Pearson’s correlation coefficient and standard error of prediction were calculated. Standard error was compared using paired t-test between the two models.

Results : Pearson’s correlation coefficients were 0.4402 (P < .0001) and 0.4440 (P < .0001) in the best and the worst model, respectively. Standard error of the best model was significantly higher than that of the worst model (0.2985 vs 0.2975, P < .0001). When OCT images without cropping was used, the standard errors were 0.3293 and 0.3143, respectively.

Conclusions : This study demonstrates that accuracy of estimation of BCVA was better when the OCT image size was standardized. The DL-algorithm with high standard error was rather more robust.

This is a 2020 ARVO Annual Meeting abstract.