Abstract
Purpose :
To identify biometric parameters that explain misclassifications by a deep learning classifier to detect gonioscopic angle closure based on anterior segment OCT (AS-OCT) images.
Methods :
Subjects of the Chinese American Eye Study underwent a complete ocular examination with gonioscopy and AS-OCT imaging of each quadrant of the anterior chamber angle. A previously described deep learning algorithm was applied to one AS-OCT per quadrant to predict a gonioscopy grade based on the modified Shaffer grading scale (grades 0/1 = closed; grades 2/3/4 = open). Median biometric measurements were calculated and compared between prediction classes using the Kruskal-Wallis test due to non-normal distributions. Pairwise comparisons of median biometric measurements between prediction classes were performed using the post-hoc Dunn’s test, adjusted for multiple comparisons.
Results :
Among 584 total images, 271 images with angle closure (true positive, TP) and 224 images with open angle (true negative, TN) were correctly predicted by the classifier. 77 images with open angle were overcalled as closed (false positive, FP). 12 images with angle closure were undercalled as open (false negative, FN).
There were significant differences (p < 0.001) in median parameter measurements between prediction classes among anterior chamber angle (ACA) parameters (angle opening distance, trabecular iris space area, scleral spur angle). The order of median ACA parameter measurements from smallest to largest was: TP, FN, FP, TN.
There were significant differences (p < 0.036) in median parameter measurements between prediction classes among anterior segment (AS) parameters. For some AS parameters (iris area, anterior chamber depth), the order of median parameter measurements from smallest to largest differed from the ACA parameters: TP, FP, FN, TN. This order was the same for other AS parameters (iris curvature, lens vault) when median parameter measurements were ordered from largest to smallest.
Conclusions :
Overall, there were more FP than FN by the deep learning classifier. Our findings suggest that when there is disagreement between ACA and AS parameters, the classifier preferentially makes predictions based on AS over ACA parameters. A second layer of quantitative analysis may help identify initially misclassified images and improve classifier accuracy in detecting gonioscopic angle closure.
This is a 2021 ARVO Annual Meeting abstract.