Automated Grading of Diabetic Retinopathy with Ultra-Widefield Fluorescein Angiography and Deep Learning

Figure 2

Detection of biomarkers. The low brightness biomarkers, i.e., nonperfusion areas on UWFA images, were obtained from the real image and the fake image by an automatic detection algorithm. A minimum filter extracts the local low brightness distribution of an image, which reveals the darker tissue or the nonperfusion area. The bright biomarkers, i.e., leakage and microaneurysms on UWFA images, are obtained from the results of localization and classification. The fake image is close to the normal image domain, where the abnormally bright areas are replaced by a lower brightness appearance. The difference image of the real image and fake image reveals the anomaly detection of leakage. A thresholding operation leverages the anomaly map to the segmentation of leakage. The leakage index is the ratio of the leakage area to .