Deep Learning for Glaucoma Detection and Identification of Novel Diagnostic Areas in Diverse Real-World Datasets

• DOI: 10.1167/tvst.11.5.11
• Publication Date: 2022-05-12T14:31:38Z
• Authors: Erfan Noury; Suria S. Mannil; Robert T. Chang; An Ran Ran; Carol Y. Cheung; Suman S. Thapa; Harsha L. Rao; Srilakshmi Dasari; Mohammed Riyazuddin; Dolly Chang; Sriharsha Nagaraj; Clement C. Tham; Reza Zadeh
• Keywords: Optic disc; Optic disk

Purpose: To develop a three-dimensional (3D) deep learning algorithm to detect glaucoma using spectral-domain optical coherence tomography (SD-OCT) optic nerve head (ONH) cube scans and validate its performance on ethnically diverse real-world datasets cropped ONH scans. Methods: In total, 2461 Cirrus SD-OCT of 1012 eyes were obtained from the Glaucoma Clinic Imaging Database at Byers Eye Institute, Stanford University, March 2010 December 2017. A 3D neural network was trained tested this unique raw OCT dataset identify multimodal definition excluding other concomitant retinal disease neuropathies. total 1022 363 glaucomatous (207 patients) 542 291 normal (167 included in training, 142 48 (27 61 39 (23 validation set. 3371 (Cirrus SD-OCT) four different countries used for evaluation model: non overlapping test (USA) consisted 694 scans: 241 113 66 patients 453 157 89 patients. The Hong Kong (total 1625 scans; 666 196 99 959 277 155 patients), India 672 211 147 98 461 171 101 Nepal 380 158 143 222 174 109 external evaluation. model then evaluated manually new called DiagFind. region by identifying appropriate zone image expected location relative Bruch's Membrane Opening (BMO) commercially available imaging software. Subgroup analyses performed groups stratified eyes, myopia severity glaucoma, set cases without field defects. Saliency maps generated highlight areas make prediction. model’s compared that specialist all information subset cases. Results: system achieved area under curve (AUC) values 0.91 (95% CI, 0.90–0.92), 0.80 0.78–0.82), 0.94 0.93–0.96), 0.87 0.85–0.90) Stanford, Kong, India, datasets, respectively, perimetric AUC 0.99 0.97–1.00), 0.96 0.93–1.00), 0.92 0.89–0.95) severe, moderate, mild cases, an 0.77 value 0.90–0.93) versus human grader with same (\(P=0.99\)). terms recall defects found be 0.76 (0.68–0.85). highlighted lamina cribrosa superficial retina as regions associated classification. Conclusions: convolutional (CNN) cubes can distinguish countries. additional random cropping data augmentation reasonably scans, indicating importance detection. Translational Relevance: CNN developed identified