Estimation of foveal avascular zone area from a B-scan OCT image using machine learning algorithms

Abstract

Purpose The objective of this study is to estimate the area of the Foveal Avascular Zone (FAZ) from B-scan OCT images using machine learning algorithms. Methods We developed machine learning models to predict the FAZ area from OCT B-scan images of eyes without retinal vascular diseases. The study involved three models: Model 1 predicted the FAZ length from B-scan images; Model 2 estimated the FAZ area from the predicted length using 1, 3, or 5 horizontal measurements; and Model 3 converted the FAZ area from pixels to mm2. The models’ performance was evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), and the Coefficient of Determination (R2). The FAZ area was subsequently estimated by sequentially applying Models 1→2→3 on a new dataset. Results Model 1 achieved a MAE of 2.86, MSE of 17.56, and R2of 0.87. Model 2’s performance improved with the number of horizontal measurements, with the best results obtained using 5 lines (MAE: 40.36, MSE: 3129.65, R2: 0.95). Model 3 achieved a MAE of 1.52e-3, MSE of 4.0e-6, and R2 of 1.0. The accuracy of FAZ area estimation increased with the number of B-scan images used, with the correlation coefficient rising from 0.475 (1 line) to 0.596 (5 lines). Bland–Altman analysis showed improved agreement between predicted and actual FAZ areas with increasing B-scan images, evidenced by decreasing biases and narrower limits of agreement. Conclusions This study successfully developed machine learning models capable of predicting FAZ area from OCT B-scan images. These findings demonstrate the potential for using OCT images to predict OCTA data, particularly in populations where OCTA imaging is challenging, such as children and the elderly. Future studies could explore the developmental mechanisms of the FAZ and macula, providing new insights into retinal health across different age groups.