oa Cerebral Blood Vessel Segmentation Using Gauss-hermite Quadrature Filtering And Automatic Seed Selection

Background & Objective: Blood vessel segmentation has various applications such as proper diagnosis, surgical planning, and simulation. However, the common challenges faced in blood vessel segmentation are mainly vessels of varying width and contrast changes. In this paper, we propose a segmentation algorithm, where: (1) a histogram-based approach is proposed to determine the initial patch (seeds) and (2) on this patch, a Gauss- Hermite quadrature filter is applied across different scales to handle vessels of varying width with high precision. Subsequently, a level set method is used to perform segmentation on the filter output. Methods: In spatial domain, a Gauss-Hermite quadrature filter is basically a complex filter pair, where the real component is a line filter that can detect linear structures, and the imaginary component is an edge filter that can detect edge structures; the filter pair is used for combined line-edge detection. The local phase is the argument of the complex filter response that determines the type of structure (line/edge), and the magnitude of the response determines the strength of the structure. Since the filter is applied in different directions, all filter responses are then combined to produce an orientation invariant phase map by summing filter responses for all directions. We use 6 filters with center frequency pi/2. To handle vessels of varying width, a multi-scale integration approach is implemented. Vessels of different width appear both as lines and edges across different scales. These scales are combined to produce a global phase map that is used for segmentation. The resulting global phase map contains detailed information about line and edge structures. For blood vessel segmentation, a local phase of 90 degree indicates edge structures. Therefore, it is necessary to consider only the real part of the quadrature filter response. Edges will be found at zero crossing whereas positive and negative values will be obtained for inside and outside of line structures. Therefore, level set (LS) approach is utilized that uses the real part of the phase map as a speed function to drive the deforming contour towards the vessel boundary. In this way, the blood vessel boundary gets extracted. An initial patch on the desired image object is a requirement in this algorithm to start calculating the local phase map. It is obtained by first selecting a few possible partitions using peaks (local maxima) in the intensity histogram. Then, optimal number of seeds is obtained by an iterative clustering of these peaks using their histogram heights and grey scale difference. The seeds around the object form the patch. Results & Conclusion: The proposed method has been tested on 6 subjects of Head MRT Angiography having resolution 416×512×112. We use 6 filters of size 7×7×7 and 4 scales in this experiment. The average time required by MATLAB R14 to perform segmentation is 3 m for one subject by a 2 GB RAM and core2duo processor (without optimization). The resulted segmentation is promising and robust in terms of boundary leakage as can be observed from the Figure.