Retinal Vessel Segmentation Based on Multiscale Fuzzy C-Means Clustering with Non-Local Spatial Information Constraints

Abstract

The morphological features of retinal vessels are of profound significance in the screening and clinical diagnosis of the diseases related to fundus. This paper proposes a retinal vessel segmentation method based on multiscale fuzzy C-means clustering with non-local spatial information constraints (MsFCM_NLS). Firstly, the response image of B-COSFIRE (Bar-selective Combination of Shifted Filter Responses) filter and the response image of phase congruency are extracted from the preprocessed color fundus image. These two images are fused using non-subsampled contourlet transform to obtain a feature image. Subsequently, Gaussian filtering is applied to this feature image to construct a multiscale feature image sequence. Finally, the MsFCM_NLS clustering is performed on the multiscale feature image sequence to achieve binary classification of the pixels, followed by a post-processing step and yield the ultimate segmentation result. Experimental results on the DRIVE dataset indicate that the proposed method obtains an average accuracy, sensitivity, and specificity of 0.9481, 0.7724, and 0.9774, respectively. In comparison with the FCM clustering and the MsFCM clustering methods, the proposed approach yields an improvement in both detection accuracy and sensitivity for retinal vessels. Furthermore, the segmented vessel networks exhibit better continuity and completeness.