Abstract
Purpose :
The detection of angle closure (AC) is a complex task that requires specialized knowledge and years of experience to accurately diagnose. To address this challenge, we implemented a custom DL model for the identification of AC from circumferential anterior chamber angle (ACA) images captured by automated NGS-1 gonioscope (NIDEK Co, Japan). The use of custom DL models may represent a promising approach for improving AC detection.
Methods :
Phakic subjects >40 years of age, with no relevant ophthalmic history, were consecutively recruited from a glaucoma clinic (Singapore National Eye Hospital). Each subject underwent same-day evaluation with NGS-1 gonioscope and gonioscopy in a dark room prior to pupillary dilation. The presence of AC (primary outcome measure) was defined as the non-visibility of the posterior trabecular meshwork in two or more gonioscopic quadrants by two masked, glaucoma fellowship-trained observers. The images obtained were labeled and categorized using the gonioscopic grading as the reference standard. An independent test set of 40 open and 12 closed-angle images was randomly selected, with the remaining selected as the training set for the deep learning model. A custom DL model based on VGG16 architecture was then trained to detect AC. A series of data sampling experiments were conducted on the training set to mitigate data scarcity and imbalance. The best DL model is selected from the experiments based on the area under the receiver operating characteristic curve (AUC) metrics evaluated on the independent test set.
Results :
We analyzed a total of 256 images from 138 subjects from a hospital-based sample of which 198 images were open-angle and 58 were closed-angle images.The best DL model was able to classify GS-1 circumferential anterior chamber CA images with AUC of 0.82 (95%CI 0.78 to 0.86), sensitivity of 0.83 (95%CI 0.79 to 0.87) and Specificity 0.75 (95% CI 0.71 to 0.79) based on the independent test set.
Conclusions :
The DL algorithm successfully detected gonioscopic AC using circumferential ACA images obtained by the automated gonioscope GS-1 in a hospital-based sample. These findings may validate the automation of AC assessment, including imaging and classification in glaucoma clinics.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.