Study On Segmentation Algorithms Of CT Images Of Cerebral Hemorrhage Based On Curve Evolution

With the improvement of people’s life,the phenomenon of population aging has become more and more serious in recent years.According to the analysis,the incidence of acute cerebrovascular disease is increasing year by year.Cerebral hemorrhage is a kind of acute cerebrovascular disease.It has the characteristics of high incidence,high disability rate,high mortality rate and high recurrence rate so that it causes heavy burden to families and society.To solve this problem,governments of various countries have invested a lot of money to study it.Quantitative,rapid,accurate and repeatable volume estimation of suspected hematoma can not only help patients with cerebral hemorrhage to diagnose and treat,but also determine whether patients need surgery.At present,the prevention of early hematoma enlargement in patients with intracerebral hemorrhage has become the focus of research.The research of robust and fast computed tomography(CT)image segmentation algorithm for intracerebral hemorrhage has become an important foundation for the prevention of early hematoma enlargement in patients with intracerebral hemorrhage,which provides important data support for the treatment of patients with intracerebral hemorrhage.However,the small amount of data in CT images of cerebral hemorrhage and the lack of publicly labeled data sets lead to the lack of data support for semantics segmentation based on deep learning.Meanwhile,there are often blend signs,black hole signs and Liqi island signs about the data.Moreover,the level set methods have the problem of artificial initialization,which makes it unable to solve the above problems at the same time.To solve problems of CT image segmentation of cerebral hemorrhage,this paper firstly analyzed the CT image of cerebral hemorrhage and transformed its features into the following problems: intensity inhomogeneity,blurred edge,multiple objective and holes.To deal with these problems,an adaptive morphological segmentation of suspected hematoma region is proposed in this paper.The whole process includes three stages: the preprocessing of CT images of cerebral hemorrhage,suspected hematoma region detection based on faster region convolutional neural networks(Faster R-CNN)and morphological segmentation of hematoma based on level set method of Legendre polynomial signed pressure function(LPSPF).After pre-processing CT images of cerebral hemorrhage,the final binary hematoma image is obtained by combining suspected hematoma region detection based on Faster R-CNN and morphological segmentation of hematoma with LPSPF level set method,and then integrating the segmented binary images according to the location information of hematoma detection.Experiments on CT image data of cerebral hemorrhage provided by Neuroscience Center of Chongqing Medical University(NCCMU)show that compared with other typical level set methods,the proposed algorithm can simultaneously solve CT images of cerebral hemorrhage with intensity inhomogeneity,blurred edge,multiple objective and holes.Combining with the idea of hematoma detection,it saves the time cost of artificial initialization of contour and has better robustness and practicality.The main contributions and research contents of this paper are summarized as follows:At first,this paper proposed an adaptive intracranial extraction algorithm for CT images of cerebral hemorrhage.In the process of preprocessing,the intracranial part of ones needs to be extracted separately because of the interference of extracranial part on the segmentation of hematoma after denoising.The paper combines maximum interclass variance method,median filtering and morphology operations to extract intracranial components and lay a foundation for the later hematoma morphological segmentation.Secondly,a method of suspected hematoma region detection based on Faster R-CNN is proposed.After preprocessing,the data is expanded and the hematoma in the image is labeled.Then,Faster R-CNN is used to train and detect the hematoma in the test image.The detected bounding box is used as the initial contour of the suspected hematoma in the level set method.Locating the initial contour of the hematoma automatically can save manpower and time cost of segmentation.Eventually,the detected hematoma images are extracted region of interest(ROI)to shorten the time of level set segmentation and improve the efficiency of segmentation.Finally,a hematoma morphological segmentation algorithm based on LPSPF is proposed.Because of the complexity of the shape and density of CT images of cerebral hemorrhage,the traditional level set method can’t solve problems with intensity inhomogeneity,blurred edge,multiple objective and holes simultaneously.The proposed algorithm combines Legendre polynomial with signed pressure function can solve the problems of intensity inhomogeneity and blurred edge.In addition,the paper improved the proposed algorithm by generalizing LPSPF and combining with edge stopping function to effectively solve the problems of multiple objective and holes.Compared with other classical level set methods,the proposed algorithm has better results in the number of iterations,running time and similarity coefficients,which qualitatively and quantitatively proves the robustness of our model.