Development and performance of a novel ‘offline’ deep learning (DL)-based glaucoma screening tool integrated on a portable smartphone-based fundus camera

Abstract

Purpose : There exists no ideal screening tool for glaucoma. We developed and assessed the performance of a novel automated screening tool using DL on monoscopic fundus images. It is deployed on-the-edge on a portable smartphone-based fundus camera, and assesses presence of referable glaucoma.

Methods : We trained a binary classifier for referable glaucoma (glaucoma suspect and likely glaucoma) using DL with a diverse dataset of 5716 images from Asian and Caucasian populations. This included 58% images of referable glaucoma. Five glaucoma specialists were involved in grading. The ground truth came from both the images themselves as well as from a comprehensive glaucoma evaluation. The resulting algorithm was validated on two datasets. Validation set A comprised of 626 images (63% referable glaucoma) captured on the target device on Asian eyes. The reference standard was optic disc assessment on these images. An independent Test set B comprised of 389 images (62% referable glaucoma) captured on Caucasian eyes. The reference standard was the diagnosis of glaucoma made following a complete glaucoma evaluation. The ground truth for both datasets was the interpretation made by glaucoma specialists. The activation maps generated by the algorithm highlight the regions of the input images which contributed most to the automated diagnosis.

Results : The DL algorithm had a sensitivity of 97% (CI 96%-99%), a specificity of 92% (CI 88%-95%) and an AUC of 0.97 in detecting referable glaucoma on validation set A. Sensitivity in detecting likely glaucoma was 0.97. On test set B, the sensitivity for detecting referable glaucoma was 96% (CI 93%-98%) the specificity was 82% (CI 76%-88%) and the AUC 0.93. Sensitivity in detecting likely glaucoma was 0.98. Activation maps showed that the AI relied on key glaucoma features like vertical cup-to-disc ratio, neuroretinal rim abnormalities, disc hemorrhages and retinal nerve fibre layer (RNFL) defects to make a diagnosis.

Conclusions : The DL algorithm based on monoscopic fundus images deployed ‘offline’ on a portable fundus camera has high sensitivity and specificity in screening for referable glaucoma. Further work includes validation in prospective clinical trials.

This is a 2021 ARVO Annual Meeting abstract.

 

Activation maps highlighting areas of abnormality that triggered diagnosis by the AI