An AI model to estimate visual acuity based solely on cross-sectional OCT imaging of various diseases

To develop an artificial intelligence (AI) model for estimating best-corrected visual acuity (BCVA) using horizontal and vertical optical coherence tomography (OCT) scans of various retinal diseases and examine factors associated with its accuracy.

OCT images and associated BCVA measurements from 2,700 OCT images (accrued from 2004 to 2018 with an Atlantis, Triton; Topcon, Tokyo, Japan) of 756 eyes of 469 patients and their BCVA were retrospectively analysed. For each eye, one horizontal and one vertical OCT scan in cross-line mode were used. The GoogLeNet architecture was implemented. The coefficient of determination (R²), root mean square error (RMSE) and mean absolute error (MAE) were computed to evaluate the performance of the trained network.

R², RMSE, and MAE were 0.512, 0.350, and 0.321, respectively. R² was higher in phakic eyes than in pseudophakic eyes. Multivariable regression analysis showed that a higher R² was significantly associated with better BCVA (p < 0.001) and a higher standard deviation of BCVA (p < 0.001). However, the performance was worse in an external validation, with R² of 0.19. R² values for retinal vein occlusion and age-related macular degeneration were 0.961 and 0.373 in the internal validation but 0.20 and 0.22 in the external validation.

Although underspecification appears to be a fundamental problem to be addressed in AI models for predicting visual acuity, the present results suggest that AI models might have potential for estimating BCVA from OCT in AMD and RVO. Further research is needed to improve the utility of BCVA estimation for these diseases.