Abstract
Purpose :
Primary angle-closure glaucoma (PACG) is a major cause of irreversible blindness in Asia. This study aims to develop code free deep learning (CFDL) models for angle-closure detection from AS-OCT images, while comparing the usability of different CFDL platforms for model development, evaluated by model performance and available features. A bespoke custom deep learning (CDL) model was also developed for comparison with CFDL models to evaluate their performances.
Methods :
1900 anterior chamber angle (ACA) images (1130 open-angle and 770 angle-closure) were collected from 272 eyes of 185 subjects using AS-OCT. 1515 and 385 ACA images were used for model training and testing, respectively. CFDL models were trained on 4 platforms: Amazon, Apple, Baidu, and Google. In addition, a CDL model was developed with Tensorflow and Python to facilitate quantitative analysis between performance of CFDL models and CDL model. The sequential convolutional neural network layers of the CDL model were implemented by Keras with Tensorflow. Both models were trained and tested by the same dataset. Model performance was evaluated by F1 score. 9 platform features such as confusion matrix generation were highlighted as significant for model development. Features available on the four selected platforms were compared.
Results :
Google offered the highest number of significant features among platforms, with 7 out of 9 significant features available. Amazon achieved 1.000 F1 score, 100% sensitivity, specificity and 1.000 AUPRC. Google achieved 0.995 F1 score, 99.5% sensitivity, specificity and 1.000 AUPRC. Baidu achieved 0.945 F1 score, 94.5% sensitivity and specificity. Apple achieved 0.769 F1 score, 76.9% sensitivity and specificity. Bespoke CDL model achieved 0.995 F1 score, 99.5% sensitivity, specificity and 0.997 AUPRC. Apple performed significantly poorer than other platforms and CDL model, while there were no significant performance differences between Amazon/Baidu/Google/CDL model.
Conclusions :
The CFDL models could detect angle-closure with high accuracy and reliability, demonstrating comparable performance to bespoke CDL models. These models could be deployed to support angle-closure screening to facilitate early detection, triage for specialist evaluation and prompt treatment, which reduces glaucoma progression and vision loss risk, while being easily accessible without coding or hardware requirement.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.