Author: Fethallah Benmansour (Switzerland)
Co-authors: Qi Yang, Dimitrios Damopoulos, Neha Anegondi, Ales Neubert, Jelena Novosel, Beatriz Garcia Armendariz
Purpose
Artificial intelligence models could optimise screening of patients with diabetic retinopathy at risk of disease progression for participation in clinical trials as wells as support clinical decision making to improve patient outcomes. Here we assess the performance of deep learning (DL) employing 7-field colour fundus photographs (7F-CFP) in automated identification of eyes with moderately severe and severe non-proliferative diabetic retinopathy (NPDR) among patients with diabetes.
Setting/Venue
Primary care centres in the United States.
Methods
The eyes of 37,358 patients with diabetes were analysed using data (including images) collected between 1999 and 2016 (Source: Inoveon Corporation, Oklahoma City, OK). Professional graders at a centralised reading centre assessed DR severity and the presence of clinically significant macular edema from 7F-CFP. DR severity was graded using the Early Treatment Diabetic Retinopathy Study Diabetic Retinopathy Severity Scale (DRSS). Prevalence of moderately severe or severe NPDR (DRSS 47–53), considering the worst DRSS score at the patient level, was 2.2% in this cohort. The dataset was split into 80% for model training, 10% for tuning and 10% for testing, for a total of 29,890, 3732 and 3736 patients with 1,430,046, 180,534 and 180,135 images, respectively. A DL Inception-v3 model with transfer learning was trained at the image level on all 7 fields of view (including stereoscopy) for being either DRSS 47–53 or not. Predictions were averaged over all fields of view to provide a prediction at the eye level. Model performance metrics in terms of area under the receiver operating characteristic (AUROC) curve, specificity, sensitivity and positive predictive value are reported.
Results
The best model was selected based on performance on the tuning set, as well as the optimal cutoff for specificity and sensitivity maximising the Youden index. The model performed well on the testing set, as shown by an AUROC of 0.988 (95% CI, 0.9872, 0.9879), sensitivity of 96.39% (95% CI, 96.28%, 96.55%), specificity of 96.24% (95% CI, 96.21%, 96.25%) and positive predictive value of 0.368 (95% CI, 0.366, 0.370).
Conlusions
Our findings demonstrate that DL can support automated identification of eyes with DRSS 47–53. The model can optimise screening of patients at risk of disease progression for participation in clinical trials before vision threatening presentations occur, as well in clinical practice. Future research will further refine this proof-of-concept algorithm, with validation conducted on other independent diverse datasets and in a real-world setting.
Financial Disclosure
Fethallah Benmansour, Dimitrios Damopoulos, Ales Neubert, Jelena Novosel and Beatriz Garcia Armendariz are employees of F. Hoffmann-La Roche Ltd., Basel, Switzerland. Qi Yang, Neha Anegondi and Daniela Ferrara are employees of Genentech Inc., South San Francisco, CA
Comments
Please add below extra co-author: Co-author 7 first name: Daniela Co-author 7 last name: Ferrara Co-author 7 affiliation: Genentech Inc., South San Francisco, CA, USA
