Multidisease Deep Learning Neural Network for the Diagnosis of Corneal Diseases

• Published in: American journal of ophthalmology Vol. 226; pp. 252 - 261
• Main Authors: Elsawy, Amr, Eleiwa, Taher, Chase, Collin, Ozcan, Eyup, Tolba, Mohamed, Feuer, William, Abdel-Mottaleb, Mohamed, Abou Shousha, Mohamed
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 01.06.2021; Elsevier Limited
• Subjects: Algorithms; Area Under Curve; Artificial intelligence; Automation; Classification; Cornea; Corneal Diseases - diagnosis; Deep Learning; Diabetes; Diabetic retinopathy; Diagnosis, Computer-Assisted; Disease; Dry Eye Syndromes - diagnosis; FDA approval; Female; Fuchs' Endothelial Dystrophy - diagnosis; Health care; Humans; Keratoconus - diagnosis; Male; Medical diagnosis; Middle Aged; Neural networks; Neural Networks, Computer; Patients; Prospective Studies; Public health; Quality of life; ROC Curve; Tomography; Tomography, Optical Coherence; Transplants & implants; United States > US
• ISSN: 0002-9394; 1879-1891; 1879-1891
• DOI: 10.1016/j.ajo.2021.01.018

To report a multidisease deep learning diagnostic network (MDDN) of common corneal diseases: dry eye syndrome (DES), Fuchs endothelial dystrophy (FED), and keratoconus (KCN) using anterior segment optical coherence tomography (AS-OCT) images. Development of a deep learning neural network diagnosis algorithm. A total of 158,220 AS-OCT images from 879 eyes of 478 subjects were used to develop and validate a classification deep network. After a quality check, the network was trained and validated using 134,460 images. We tested the network using a test set of consecutive patients involving 23,760 AS-OCT images of 132 eyes of 69 patients. The area under receiver operating characteristic curve (AUROC), area under precision-recall curve (AUPRC), and F1 score and 95% confidence intervals (CIs) were computed. The MDDN achieved eye-level AUROCs >0.99 (95% CI: 0.90, 1.0), AUPRCs > 0.96 (95% CI: 0.90, 1.0), and F1 scores > 0.90 (95% CI: 0.81, 1.0) for DES, FED, and KCN, respectively. MDDN is a novel diagnostic tool for corneal diseases that can be used to automatically diagnose KCN, FED, and DES using only AS-OCT images.