Purpose: To develop a software package for automated measuring of the trabecular-iris angle (TIA) using ultrasound biomicroscopy.
Methods: Ultrasound biomicroscopy images were collected and the TIA was manually measured by specialists. Different models were used as the convolutional neural network for the automatic TIA measurement. The root-mean-squared error, explained variance, and mean absolute percentage error were used to evaluate the performance of these models. The interobserver reproducibility, coefficient of variation, and intraclass correlation coefficient were calculated to evaluate the consistency between the manual measured and the model predicted values.
Results: ResNet-18 had the best performance in root-mean-squared error, explained variance, and mean absolute percentage error among all 5 models. The average difference between the angles measured manually and by the model is -0.46±3.97 degrees for all eyes, -1.67±5.19 degrees for open angles, and 0.75±1.43 degrees for narrow angles. The coefficient of variation, intraclass correlation coefficient, and reproducibility of the total TIA measurements are 6.8%, 0.95, and 6.1 degrees for all angles; 6.4%, 0.99, and 7.7 degrees for open angles; and 8.8%, 0.93, and 4 degrees for narrow angles, respectively.
Conclusions: Preliminary results show that this fully automated anterior chamber angle measurement method can achieve high accuracy and have good consistency with the manual measurement results, this has great significance for future clinical practice.