Deep learning-based segmentation and quantification of retinal capillary non-perfusion on ultra-wide-field retinal fluorescein angiography

Abstract

Reliable outcome measures are required for clinical trials investigating novel agents for preventing progression of capillary non-perfusion (CNP) in retinal vascular diseases. Currently, accurate quantification of topographical distribution of CNP on ultrawide field fluorescein angiography (UWF-FA) by retinal experts is subjective and lack standardisation. A U-net style network was trained to extract a dense segmentation of CNP from a newly created dataset of 75 UWF-FA images. A subset of 20 images was also segmented by a second expert grader for inter-grader reliability evaluation. Further, a circular grid centred on the FAZ was used to provide standardised CNP distribution analysis. The model for dense segmentation was five-fold cross-validated achieving area under the receiving operating characteristic of 0.82 (0.03) and area under precision-recall curve 0.73 (0.05). Inter-grader assessment on the 20 image subset achieves: precision 59.34 (10.92), recall 76.99 (12.5), and dice similarity coefficient (DSC) 65.51 (4.91), and the centred operating point of the automated model reached: precision 64.41 (13.66), recall 70.02 (16.2), and DSC 66.09 (13.32). Agreement of CNP grid assessment reached: Kappa 0.55 (0.03), perfused intraclass correlation (ICC) 0.89 (0.77, 0.93), non-perfused ICC 0.86 (0.73, 0.92), inter-grader agreement of CNP grid assessment values are Kappa 0.43 (0.03), perfused ICC 0.70 (0.48, 0.83), non-perfused ICC 0.71 (0.48, 0.83). Automated dense segmentation of CNP in UWF-FA images achieves performance levels comparable to inter-grader agreement values. A grid placed on the deep learning-based automatic segmentation of CNP generates a reliable and quantifiable method of measurement of CNP, to overcome the subjectivity of human graders.