Fovea localization by blood vessel vector in abnormal fundus images

In human eyes, macula is responsible for sharp central vision with fovea in the center. The location of fovea becomes an important landmark in diagnosing the retinal diseases. As macula doesn't have the clear boundary and obvious shape, deep learning techniques to locate the fovea often fail in complicated lesions and insufficient training samples, and the unsupervised method is incapable for illumination variations. In this paper, a new unsupervised fovea localization method using the retinal raphe and region searching is proposed, and the blood vessel vector (BVV) model is developed. After detecting blood vessels and OD by U-net and probability bubbles, the BVVs are conceived and the retinal raphe is obtained by summating all the BVVs, then the fovea is estimated through the local region searching. Compared with the parabola model, the BVV model does not involve the coordinate transformation and reduces the complexity to the linear time cost O(N). Two other unsupervised techniques the parabola model and intensity searching and five supervised techniques cGAN, U-net, DRNet, MedTnet and EANet are included and compared. The global feature of retinal vessels is utilized which makes the proposed method more robust to the lesions than the other localization methods. The experiments on public datasets Kaggle, MESSIDOR and IDRiD validate the effectiveness of the proposed method by the student's t-test, and our method obtains the least average Euclidean distance to the groundtruth on Kaggle and almost least on Base 33 of MESSIDOR. (c) 2022 Elsevier Ltd. All rights reserved.