SVD-based maximum principal curvatures for segmentation of blood vessels from diabetic retinopathy fundus images using level sets

Retinal fundus images provide a large amount of space for blood vessel segmentation. The result of segmentation is not only a vital indicator of diabetic retinopathy (an eye-related disease), but also a good facilitator for ophthalmologists (eye specialists or eye doctors) to carry out essential diagnostic procedures. Diabetic retinopathy, especially in diabetic patients, is manifested when tiny blood vessels get bleed. In this paper, a novel method called singular value decomposition (SVD)-based maximum principal curvatures (SVD-PC) using level set procedure from retinal images and segmentation of blood vessel vascular structure is proposed. Most of the researchers worked on the green RGB retinal image channel. But, because of intensity variations that generate blurriness, the green channel version of RGB has hidden noise, which is a tough challenge to overcome in a processing image. Moreover, the blood vessels in an image have insufficient local contrast, and to tackle this issue for efficient image processing, distinct filters were applied in the past. Here, in this proposed method, SVD, which is a mathematical technique related to vector calculus and matrix algebra, is used to efficiently deal with the complexities aroused working with green channel and filters. SVD, in our proposed method, as a pre-processing technique is effective in: (i) extracting colored feature set of blood vessel pixels from input image and (ii) effectively converting feature set pixel image into gray. Later, maximum principal curvature values of those converted gray image feature set pixels are calculated in the post-processing phase, followed by ISOData Thresholding to segment the tree-shaped vasculature, which morphological operator is applied to remove the unwanted falsely segmented vessels. The algorithm is implemented in MATLAB and has given segmentation accuracy of 97.8%. This proposed algorithm operated on images of STARE as well as DRIVE datasets. It is evident that singular value decomposition is quite effective and outstanding when compared with another proposed method in this paper, i.e., using maximum principal curvatures based on principal component analysis.