Research On Image Segmentation Algorithm And Detection Algorithm Of Ischemic Stroke In CT

Image segmentation is one of the important branches of image processing,and active contour model is one of the most successful methods of imagesegmentation algorithm. The field of image segmentation includes imagesegmentation algorithm and evaluation algorithm for image segmentation. Wetake the image segmentation as the main line in this dissertation, and mainlywork on three topics: active contour models, evaluation algorithm for imagesegmentation and detection algorithm of ischemic stroke in CT.First, the existing Wasserstein distance-based segmentation models areunable to segment image with intensity inhomogeneity effectively, thus, wedevelop a novel nonparametric Wasserstein distance-based active contourmodel that is able to utilize image histogram information in local region. Toquantify the similarity between two regions, we proposed to compare theirrespective histograms using the Wasserstein distance; Due to a Gaussian kernelintroduced, intensity information in local regions is extracted and embedded inmodel to guide the motion to overcome the difficulties in image segmentationcaused by intensity inhomogeneities. Experiment results prove that our modelsegments texture images with intensity inhomogeneity effectively.Secondly, some classic variational models, such as the active contourmodel without edge (CV) and the region-scalable fitting energy model (RSF),are sensitive to the initial contour. A good guess of initial contour is needed to obtan satisfactory segmentation results. Therefore, we propose an adaptiveaugmentment of edge energy based active contour model. This model is mainlycomposed of a range domain kernel function, space domain kernel function,and edge indicator function. With the range domain kernel function, bigweights were assigned to pixels in detected regions to enhance the energy ofpixels close to the boundaries, such that this model was independent of initialcontour; the space domain kernel function could extract local intensityinformation to overcome the effect of intensity inhomogeneity; the edgeindicator function was used to improve the detection of the boundaries ofobject. Experiment results show this model is insensitive to contourinitialization and robust to intensity inhomogeneity, and is helpful forsegmenting brain in brain CT images.After that, we study the detection algorithm of ischemic stroke in CT,which is a huge challenge for the segmentaion algorithm. This detectionalgotihm has five main steps: brain region extraction, image correction,ventricle segmentation, cerebral fluid (CSF) segmentation and detection ofischemic stroke. We first obtain the brain region by stripping the skull, andthen correct inclination angle of brain by aligning midline of brain with thevertical center line of a slice. In ventricle segmentation step, the biggestdifficult is to prevent the lesion into the segmentation result of ventricle region.To overcome this difficulty, we adopt the three-dimensional connectiveityanalysis, the ring of brain edge analysis and dual-threshold difference method,the adaptive ROI template matching method. Then, the CSF is segmented bythe seed growing threshold and the linear classifier. Finally, we apply acontralateral subtraction technique to extract a set of suspicious ischemic strokeregions. The experiment results show that our detection algorithm could have adesirable performance for the detection of ischemic stroke in CT.Finally, an objective global evaluation algorithm of segmentation results isproposed, which is based on computing the deviation of the segmentation results from reference segmentation. By taking into account imageunderstanding view, the discrepancy between two results is weighted based onspatial contextual information, thereby obtaining the different weightiness.Moreover, relative distance, foreground and background searching distance areintroduced into the proposed metric, and the value of this metric is invariant tothe segmentation images with varying scale. In order to overcome the irrationalevaluation caused by the excessive over-segmentation and under-segmentation,a penalized term is proposed. Experiment results prove that our evaluationaccords with the view of the image understanding and it is more global androbust.