Spatial-frequency fusion for retinal vessel segmentation

Abstract

The segmentation of retinal blood vessels is clinically significant for diagnosing many ocular disorders and can assist in identifying multiple medical conditions, including diabetes, atherosclerosis, and cardiovascular disease. Therefore, accurate identification of the retinal blood vessels in the fundus can significantly aid physicians in diagnosing and treating their patients’ conditions. In this paper, we propose a retinal blood vessel segmentation method that combines the spatial and frequency domains. Existing CNN methods obtain local features by using convolutional operations in the spatial domain, and are not capable enough in obtaining global spatial feature information. Therefore, we introduce a Fourier transform to obtain global information and learn the long-distance distribution of blood vessels. In the frequency domain, we designed a multiscale Gaussian high-pass filter to adaptively enhance the edge features of blood vessels of different scales. Since frequency domain information is more concerned with global dependencies and spatial information is more capable of capturing local detailed features, the fusion of frequency and spatial domains can effectively capture the general trends and complex details within the hidden layer space. In order to assess the model’s efficacy, we conducted tests using the pre-existing DRIVE and CHASE_DB1 datasets. Our accuracy achieved 96.90 and 97.81 respectively, and a sensitivity of 83.80 was obtained for the DRIVE dataset. By observing the segmented image, our segmentation is more accurate, clearer, and noise-free than the results of other proposed methods.