A Dual-Channel Deep Learning Method for Automated Glaucoma Identification using Stereoscopic Disc Photos

Purpose:
We propose a deep learning based method for automated identification of glaucoma patients using stereoscopic disc photos.

Methods:
Stereoscopic images were acquired from a cohort of 669 subjects, including 377 glaucoma patients and 292 normal controls. A novel deep algorithm was developed for differentiating the two groups, which was based on the 50-layer ResNet neural network and pretrained on the ImageNet database. The proposed model features a dual-channel encoding method, which can automatically encode the spatial and volumetric information in a stereoscopic image into multiple feature maps and further combine them for image classification.

Results:
We used 851 stereoscopic images from 569 subjects (192 with images of both eyes) for training the model, and the remaining 145 images from 100 subjects (45 with images of both eyes) for performance validation. A mono-channel ResNet model was also implemented as the baseline method for comparison, which was trained and tested using the monoscopic images from the same cohort. The proposed method achieved superior classification performance (accuracy: 97.9%, sensitivity: 95.4%, specificity: 99.1%), as compared to the baseline mono-channel method (accuracy: 96.6%, sensitivity： 94.1%, specificity: 97.7%.)

Conclusions:
In this work, we developed a dual-channel ResNet model for glaucoma classification using stereoscopic disc images. The results showed that this novel AI model outperformed the classic mono-channel based deep learning method, and thus may have a high potential in computer-aided optic disc assessment.