Retinal vascular segmentation combined with PCNN and morphological matching enhancement
Xu Guangzhu1,2, Wang Yawen1, Hu Song1, Chen Peng1,2, Zhou Jun3, Lei Bangjun1,2     
1. College of Computer and Information Technology, China Three Gorges University, Yichang, Hubei 443002, China;
2. Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering, China Three Gorges University, Yichang, Hubei 443002, China;
3. Department of Diagnostic Ultrasound, the First College of Clinical Medical Science, China Three Gorges University, Yichang, Hubei 443002, China

Overview: Studies indicate that retinal blood vessels are the only deep micro-vessels in a human body that can be observed directly in a non-invasive way. The variation of color or the morphological structure of vascular networks can reflect the effects on human health of various eye diseases and cardiovascular and cerebrovascular diseases. Therefore, the extraction and analysis of retinal vascular is of great significance for medical personnel to diagnose and treat these diseases as early as possible. Due to the limitation of image acquisition equipment and the complex structure of retinal blood vessels, manual extraction of retinal blood vessels has problems of heavy workload and strong subjectivity. Aiming at the problem, this paper proposes a novel automatic retinal vessel image segmentation algorithm based on matched filter enhancement and region growth pulse coupled neural network. Firstly, the original fundus image is pre-processed with a 2D Gaussian filter bank and a 2D Gabor matched filter bank to achieve the contrast enhancement and denoising. By combining these two kinds of filters, the final fused retina image can present more details and less artifact noisy micro-vessels. Secondly, a modified regional growing pulse coupled neural network with fast linking mechanism is adopted. The pixel with the highest brightness is selected as the seed, and adaptive connection coefficients and specially designed terminating conditions are employed to control the growth of the candidate blood vessel area. Operating in this way can overcome the shortcomings of the regular region-growing technique requiring fixed preselected seeds and the traditional PCNN not being able to terminate automatically. In order to evaluate the performance of the proposed algorithm, the DRIVE image dataset, which has been widely used for retina image processing, is adopted. The dataset was acquired using a Canon CR5 non-mydriatic 3CCD camera and each image was captured using 8 bits per color plane at 768 pixelsx584 pixels. The dataset of 40 images has been divided into a training set and a test set, both containing 20 images. The experimental results demonstrate that the algorithm can maintain the consistency of the segmented results and meanwhile achieve the multi-value segmentation of fundus vascular images. The whole algorithm performs well in the DRIVE fundus database. The average accuracy, sensitivity and specificity of the algorithm respectively are 93.96%, 78.64% and 95.64% in DRIVE fundus database. There are fewer vascular breakpoints in the segmentation results, and the micro-vessels are clear. We believe that this work has good application prospects.

Supported by National Natural Science Foundation of China (61402259, 61272236, U1401252) and Yichang Applied Basic Research Project (A19-302-13)