Effects of Study Population, Labeling and Training on Glaucoma Detection Using Deep Learning Algorithms.

To compare performance of independently developed deep learning algorithms for detecting glaucoma from fundus photographs and to evaluate strategies for incorporating new data into models. Two fundus photograph datasets from the Diagnostic Innovations in Glaucoma Study/African Descent and Glaucoma Evaluation Study and Matsue Red Cross Hospital were used to independently develop deep learning algorithms for detection of glaucoma at the University of California, San Diego, and the University of Tokyo. We compared three versions of the University of California, San Diego, and University of Tokyo models: original (no retraining), sequential (retraining only on new data), and combined (training on combined data). Independent datasets were used to test the algorithms. The original University of California, San Diego and University of Tokyo models performed similarly (area under the receiver operating characteristic curve = 0.96 and 0.97, respectively) for detection of glaucoma in the Matsue Red Cross Hospital dataset, but not the Diagnostic Innovations in Glaucoma Study/African Descent and Glaucoma Evaluation Study data (0.79 and 0.92; Deep learning glaucoma detection can achieve high accuracy across diverse datasets with appropriate training strategies. Because model performance was influenced by the severity of disease, labeling, training strategies, and population characteristics, reporting accuracy stratified by relevant covariates is important for cross study comparisons. High sensitivity and specificity of deep learning algorithms for moderate-to-severe glaucoma across diverse populations suggest a role for artificial intelligence in the detection of glaucoma in primary care.

Informations de copyright

Copyright 2020 The Authors.

Déclaration de conflit d'intérêts

Disclosure: M. Christopher, None; K. Nakahara, None; C. Bowd, None; J.A. Proudfoot, None; A. Belghith, None; M.H. Goldbaum, None; J. Rezapour, German Research Foundation (DFG, research fellowship grant RE 4155/1-1) (F), German Ophthalmological Society (DOG) (F); R.N. Weinreb, Aerie Pharmaceuticals (C), Allergan (C), Eyenovia (C), Implantdata (C), Unity (C), Heidelberg Engineering (F), Carl Zeiss Meditec (F), Centervue (F), Bausch & Lomb (F), Genentech (F), Konan Medical (F), National Eye Institute (F), Optos (F), Optovue Research to Prevent Blindness (F); M.A. Fazio, National Eye Institute (F), EyeSight Foundation of Alabama (F), Research to Prevent Blindness (F), Heidelberg Engineering (F); C.A. Girkin, National Eye Institute (F), EyeSight Foundation of Alabama (F), Research to Prevent Blindness (F), Heidelberg Engineering (F); J.M. Liebmann, Aerie Pharmaceuticals, Alcon, Allergan, Bausch & Lomb, Carl Zeiss Meditec, Eyenovia, Galimedix, Heidelberg Engineering, Zeiss Meditec, National Eye Institute, Novartis, Research to Prevent Blindness, Inc.; G. De Moraes, Novartis (C), Galimedix (C), Belite (C), Reichert (C), Carl Zeiss Meditec (C), Heidelberg Engineering (F), Topcon (F); H. Murata, The Ministry of Education, Culture, Sports, Science and Technology of Japan (Grant number 25861618) (F); K. Tokumo, None; N. Shibata, None; Y. Fujino, The Ministry of Education, Culture, Sports, Science and Technology of Japan (Grant number 20768254) (F); M. Matsuura, The Ministry of Education, Culture, Sports, Science and Technology of Japan (Grant number 00768351) (F); Y. Kiuchi, None; M. Tanito, None; R. Asaoka, The Ministry of Education, Culture, Sports, Science and Technology of Japan (Grant number 18KK0253, 19H01114 and 17K11418) (F), Daiichi Sankyo Foundation of Life Science, Suzuken Memorial Foundation, The Translational Research program, Japan Agency for Medical Research and Development (AMED, Strategic Promotion for Practical Application of Innovative Medical Technology [TR-SPRINT]); L.M. Zangwill, Carl Zeiss Meditec (F), Heidelberg Engineering (F), National Eye Institute (F), BrightFocus Foundation (F), Optovue (F), Topcon Medical System Inc. (F)