Primary Open-Angle Glaucoma Detection with Vision Transformer: Improved Generalization Across Independent Fundus Photograph Datasets

Abstract

Purpose : To compare the accuracy and generalizability of Vision Transformer, a new deep learning (DL) technique, to a traditional convolutional neural network, ResNet-50, for detecting primary open-angle glaucoma (POAG) using Ocular Hypertension Treatment Study (OHTS) fundus photographs and 5 external datasets.

Methods : 66,715 photographs of 1,636 participants in the OHTS were used to compare the best-performing Vision Transformer model (Data-efficient image Transformer, DeiT) to ResNet-50 for detecting the OHTS Endpoint Committee determinations of POAG attributable to optic disc changes (ENPOAGDISC), visual field changes (ENPOAGVF), or both (ENPOAGANY) and OHTS Optic Disc Reading Center (RCPOAGDISC) and Visual Field Reading Center (RCPOAGVF) determinations. The accuracy of Vision Transformer and Resnet-50 DL models also were compared in 5 external international test datasets of fundus images labeled as glaucoma or healthy: 1) Diagnostic Innovations in Glaucoma Study (DIGS) and African Descent and Glaucoma Evaluation Study (ADAGES, United States) datasets, 2) the ACRIMA (Spain) dataset, (3) the Large-Scale Attention-based Glaucoma (LAG, China) dataset, (4) the Retinal IMage database for Optic Nerve Evaluation (RIM-ONE, Spain) dataset, (5) The Online Retinal Fundus Image Dataset for Glaucoma Analysis and Research (ORIGA, Singapore) dataset. Areas under receiver operator characteristic curves (AUROC) were used to measure model accuracy.

Results : The Vision Transformer models demonstrated similar performance to the ResNet-50 on the OHTS test sets for all 5 ground truth POAG labels (Figure: panel a). The diagnostic accuracy of Vision Transformer was consistently higher than Resnet-50 on the independent external fundus photograph datasets (Figure: panels b-f). For example, the DeiT model AUROC for ENPOAGANY was between 0.08 and 0.20 higher than the ResNet-50 model AUROC.

Conclusions : Vision Transformer has the potential to serve as a critical tool to reduce biases and improve the generalizability of DL models for the detection of glaucoma from fundus photograps.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Comparison of model accuracy (AUROC) between ResNet-50 and DeiT for 5 OHTS ground truth labels (ENPOAGDISC, ENPOAGVF, ENPOAGANY, RCPOAGDISC, and RCPOAGVF) on the OHTS test set and on five external test sets: DIGS/ADAGES, ACRIMA, LAG, RIM-ONE, and ORIGA. Error bars represent 95% CI.