Federated Learning for Microvasculature Segmentation and Diabetic Retinopathy Classification of OCT Data

• DOI: 10.1016/j.xops.2021.100069
• Publication Date: 2021-10-13T00:04:39Z
• Authors: Julian Lo; Timothy T. Yu; Da Ma; Pengxiao Zang; Julia P. Owen; Qinqin Zhang; Ruikang K. Wang; Mirza Faisal Beg; Aaron Y. Lee; Yali Jia; Marinko V. Sarunic
• Keywords: F1 score

To evaluate the performance of a federated learning framework for deep neural network-based retinal microvasculature segmentation and referable diabetic retinopathy (RDR) classification using OCT angiography (OCTA).Retrospective analysis clinical OCTA scans control participants patients with diabetes.The 153 en face images used were acquired from 4 instruments fields view ranging 2 × 2-mm to 6 6-mm. The 700 eyes RDR consisted structural projections commercial systems.OCT delineated manually verified by retina experts. Diabetic (DR) severity was evaluated specialists condensed into classes: non-RDR RDR. configuration demonstrated via simulation clients compared other collaborative training methods. Subsequently, applied over multiple institutions models trained tested on data same institution (internal models) different (external models).For segmentation, we measured accuracy Dice similarity coefficient (DSC). For classification, accuracy, area under receiver operating characteristic curve (AUROC), precision-recall curve, balanced F1 score, sensitivity, specificity.For both applications, achieved similar as internal models. Specifically, model (mean DSC across all test sets, 0.793) fully centralized dataset DSC, 0.807). mean AUROC 0.954 0.960; attained 0.956 0.973. Similar results are reflected in calculated evaluation metrics.Federated showed traditional applications while maintaining privacy. Evaluation metrics highlight potential increasing domain diversity generalizability data.