Fully automated choroidal neovascularization diagnosis and segmentation using deep learning in projection-resolved OCT angiography

Abstract

Purpose : To develop and test a fully automated choroidal neovascularization (CNV) diagnosis and segmentation algorithm based on projection-resolved OCT angiography (PR-OCTA) using convolutional neural networks (CNNs).

Methods : In this study, eyes with CNV secondary to age-related macular degeneration and control eyes without CNV underwent 3×3-mm macular OCTA scans (AngioVue, Optovue Inc.). Control eyes included healthy eyes as well as eyes diagnosed with other retinal disease that excluded the presence of CNV including diabetic retinopathy and retinal vein occlusion. Scans were exported and processed with PR-OCTA software. 50 CNV eyes and 60 non-CNV eyes were randomly selected from whole dataset for testing, and the others were used for training.

A certified grader manually delineated the ground truth CNV membrane area and vasculature from projection-resolved (PR) outer retinal angiograms. In the proposed algorithm both the CNV membrane and vasculature were segmented using separately trained CNNs (Fig. 1). First, both the outer retinal OCT structural volume and angiographic en face images were fed into the first CNN in order to segment the membrane and diagnose CNV. Since small residual artifacts could be mis-detected as CNV membranes, a size cutoff (0.004mm², or 49 pixels) was applied. The second CNN then segments vessels within the membrane area, if a membrane was detected.

Results : A total of 1676 scans including 814 CNV scans collected from 117 eyes with CNV and 862 non-CNV scans from 490 eyes without CNV were used. Only 3 out of 60 non-CNV scans were mis-diagnosed as CNV, and the sensitivity and specificity were as high as 100% and 95%. The overall CNV membrane segmentation accuracy (intersection over union) was as high as 0.88. Using the proposed method, we were able to segment both CNV membranes and vasculatures in the scans with relative low scan quality and prevent misdiagnosis on non-CNV scans with challenging pathologies and even motion artifacts in outer retina (Fig. 2).

Conclusions : The proposed deep learning method for CNV detection and segmentation was able to accurately diagnose CNV from normal eyes and eyes with different retinal disease other than CNV commonly encountered in clinical practice. This automated algorithm reliably segmented CNV vasculatures.

This is a 2020 ARVO Annual Meeting abstract.

 

Fig. 1

 

Fig. 2