Researches On Vessel Image Enhancement Based On Hessian Matrix And Level-Set Segmentation Algorithm

The segmentation of vascular structure is significant for diagnosis assistance,treatment and surgery planning. Vessel segmentation is especially challengingproblem. Many factors make it difficult to accurately extract blood vessels, such asnoise, intensity inhomogeneity, disturbance induced by closely located adjacentobjects, low contrast and high variability of vessel size. As preprocessing step, vesselenhancement is applied prior to the vessel segmentation. This paper aims to segmentvascular struture in the last three cases and our main work is as follows:(1) Since multi-scale integration results in undesirable diffusion when two vesselsare closely located, we replace the multi-scale gradient vector computation in thegaussian scale space by gradient vector flow. It is edge-preserving and anisotropicdiffusion, equal to the gradient of image at the boundary and varies slowly inhomogeneous region. In addition, we propose a vesselness measure to detect vesselwhich gives high and homogeneous output, the response of background is set anegative value to avoid the evolving curve entering into background when wesegment vessel using level set method.(2) Since level set method can flexible deal with the topological changes and givesmooth segmentation results, we implement a level-set based model for vesselsegmentation according to the characteristics of vessel enhancement image bygradient vector flow. A penalty term is added to force the level set function to beclose to signed distance function. So, our model elimilates the re-initializition processand greatly reduced the computational cost.(3) To evaluate our vessel segmention algorithm, area overlap measure, the rate ofwrong pixel measure and the rate of missed pixel measure are applied in the paper.The experimental data is20retinal images in testing images of DRIVE data library.We take the intersection of the two manual delineations as ground truth. Besides, wepresent a simple and effective method to detect the wrong pixels and missed pixelssegmented by our model.Experimental results demonstrate that our approach can successfully separateclosely adjacent vessels and address the problems of low contrast and varying vesselwidth. Besides, our model has better accuracy and stability.