Detecting Glaucoma Using 3D Convolutional Neural Network of Raw SD-OCT Optic Nerve Scans

Author/Creator ORCID

Date

2019-10-14

Department

Program

Citation of Original Publication

Noury, Erfan; Mannil, Suria S.; Chang, Robert T.; Ran, An Ran; Cheung, Carol Y.; Thapa, Suman S.; Rao, Harsha L.; Dasari, Srilakshmi; Riyazuddin, Mohammed; Nagaraj, Sriharsha; Zadeh, Reza; Detecting Glaucoma Using 3D Convolutional Neural Network of Raw SD-OCT Optic Nerve Scans; Image and Video Processing (2019); https://arxiv.org/abs/1910.06302

Rights

This item is likely protected under Title 17 of the U.S. Copyright Law. Unless on a Creative Commons license, for uses protected by Copyright Law, contact the copyright holder or the author.

Abstract

Background. Glaucoma is a chronic progressive optic neuropathy with characteristic visual field defects and corresponding structural changes, including nerve fiber layer thinning and optic nerve neuroretinal rim loss. These changes are traditionally monitored by SD-OCT (Spectral Domain Optical Coherence Tomography), which contains a large amount of 3D voxel information in a 6mm × 6mm × 2mm cube of data. However, only a fixed 3.4 mm diameter circle (2D slice) centered over the optic nerve is currently extracted using automated segmentation of the retinal nerve fiber layer thickness (RNFL). This RNFL thickness is reported relative to a normative database to help detect thinning and neuroretinal rim loss, which does not use the additional information in the optic nerve head cube. Clinicians rarely scroll through the entire cube. Therefore we propose developing and validating a three-dimensional (3D) deep learning system using the entire unprocessed OCT optic nerve volumes to distinguish true glaucoma from normals in order to discover any additional imaging biomarkers within the cube through saliency mapping. The algorithm has been validated against 4 additional distinct datasets from different countries using multimodal test results to define glaucoma rather than just the OCT alone. We hypothesize that the output from this 3D model, alongside a map of the regions where the model attends to make a prediction, can help identify novel diagnostic information in the cube. Methods. 2076 OCT (Cirrus SD-OCT, Carl Zeiss Meditec, Dublin, CA) 6 mm cubes centered over the optic nerve, 200 × 200 × 1024 volumes of 879 eyes (390 healthy and 489 glaucoma) from 487 patients, age 18-84 years, were exported from the Glaucoma Clinic Imaging Database at the Byers Eye Institute, Stanford University, from March 2010 to December 2017. This included bilateral eyes of 391 patients and unilateral eyes of 97 patients with a right eye to left eye ratio of 1.05:1. A 3D deep neural network was trained and tested on this unique OCT optic nerve head dataset from Stanford. 570 randomly selected optic nerve head cube scans of eyes with a diagnosis of glaucoma (True Glaucoma) and 342 scans of eyes with a normal diagnosis (True Normal) were used for training. A total of 81 scans of eyes with True Glaucoma and 32 scans of eyes with True Normal annotations were included in the primary validation set. 58 scans of eyes with True Glaucoma annotation and 50 scans of eyes with a True Normal annotation were included in the test set. A total of 3620 scans (all obtained using the Cirrus SD-OCT device) from 1458 eyes obtained from 4 different institutions, from United States (943 scans), Hong Kong (1625 scans), India (672 scans), and Nepal (380 scans) were used for external evaluation. True Glaucoma for the training data was defined as glaucomatous disc changes along with defects on SD-OCT RNFL and/or GCIPL (thickness and/or deviation) maps with corresponding visual field defects as well as intraocular pressure lowering treatment upon chart review. The range of glaucoma patients included mild to severe without excluding high myopes. True normal was defined as cases with non-glaucomatous optic disc with no structural defects on OCT RNFL/GCIPL deviation or sector maps and normal visual fields upon chart review. Results. The 3D deep learning system achieved an area under the receiver operation characteristics curve (AUROC) of 0.8883 in the primary Stanford test set identifying true normal from true glaucoma. The system obtained AUROCs of 0.8571, 0.7695, 0.8706, and 0.7965 on OCT cubes from United States, Hong Kong, India, and Nepal, respectively. We also analyzed the performance of the model separately for each myopia severity level as defined by spherical equivalent and the model was able to achieve F1 scores of 0.9673, 0.9491, and 0.8528 on severe, moderate, and mild myopia cases, respectively. Saliency map visualizations highlighted a significant association between the optic nerve lamina cribrosa region in the glaucoma group. Conclusions. A 3D convolutional neural network using SD-OCT optic nerve head cubes can distinguish true glaucoma from normal with good accuracy and this generalized to multiple diverse external SD-OCT datasets. Highlighted areas from saliency mapping revealed new areas within the deep lamina cribrosa. This deserves further investigation, as there is potential to monitor laminar changes even after RNFL has thinned.