Automated segmentation of retinal nonperfusion area in fluorescein angiography in retinal vein occlusion using convolutional neural networks

• DOI: 10.1002/mp.14640
• Publication Date: 2020-12-10T22:53:00Z
• Authors: Ziqi Tang; Ximei Zhang; Guangqian Yang; Guanghua Zhang; Yubin Gong; Ke Zhao; Juan Xie; Junjun Hou; Jia Hou; Bin Sun; Zhao Wang
• Keywords: Retinal Vein

Purpose Retinal vein occlusion (RVO) is the second most common cause of vision loss after diabetic retinopathy due to retinal vascular disease. nonperfusion (RNP), identified on fluorescein angiograms (FA) and appearing as hypofluorescence regions, one significant characteristics RVO. Quantification RNP crucial for assessing severity progression However, in current clinical practice, it mostly conducted manually, which time‐consuming, subjective, error‐prone. The purpose this study develop fully automated methods segmentation using convolutional neural networks (CNNs). Methods FA images from 161 patients were analyzed, areas annotated by three independent physicians. optimal method use multi‐physicians’ labeled data train CNNs was evaluated. An adaptive histogram‐based augmentation utilized boost CNN performance. based context encoder module developed compared with existing state‐of‐the‐art methods. Results proposed achieved excellent agreements physicians images. performance can be improved significantly method. Using averaged labels best consensus all physicians, a mean accuracy 0.883±0.166 fivefold cross‐validation. Conclusions We reported segment RVO Our work help improve workflow, useful further investigating association between disease progression, well evaluating treatments management