Abstract
Purpose :
Our previous studies demonstrated high accuracy of the deep learning algorithm in classifying visual field into glaucomatous and non-glaucomatous. The aim of this study was to evaluate iGlaucoma 1.0, a smartphone app on classification of glaucomatous versus non-glaucomatous VFs.
Methods :
We included 642 VFs of 432 patients from 3 eye centers across mainland China from March 2019 to September 2019. Medical records, VF reports and optical coherence tomography (OCT) reports of the patients who visited the glaucoma clinics were collected. Ground truth labelling was made according to the VF and OCT reports. Pattern deviation probability plots (PDPs) in printed VF reports were captured with phone camera, then uploaded to the server and analyzed by the app. Diagnostic performance and time cost of VF analysis by the DDS were compared with three ophthalmologists. Recognition accuracy of the app on patterns in the PDP region was recorded and analyzed. Area under curve (AUC), sensitivity and specificity of the app in glaucoma diagnosis were compared with the ophthalmologists.
Results :
The app showed excellent performance in recognition of different patterns in PDPs. The recognition accuracies of blank space, >5%, <5%, <2%, <1% and <0.5% patterns were 0.999, 0.999, 0.996, 0.996, 0.995 and 1.000 respectively. The AUC of the app in VF classification was 0.949 (0.933-0.965) with a sensitivity of 0.907 and a specificity of 0.854, while the ophthalmologists achieved an AUC of 0.850 (0.819-0.882) with a sensitivity of 0.858 and a specificity of 0.843. Total time spent was 556 seconds for the app versus 3033 seconds for the ophthalmologists.
Conclusions :
iGlaucoma had excellent recognition of PDP pictures captured by phone camera and demonstrated better diagnostic performance than ophthalmologists. The app would be useful for assisting clinical diagnosis of glaucoma based on VF.
This is a 2020 ARVO Annual Meeting abstract.