Research On CT Image Segmentation Of Intracranial Hemorrhage Based On Active Contour And Deep Learning

Intracranial hemorrhage(ICH)is a very serious health problem with a high probability of occurrence,and its diagnosis and treatment are extremely important for improving the average life expectancy of human beings.With the aging of our country’s population becoming more and more serious,as the most common acute cerebrovascular disease,stroke has surpassed cancer and has become the number one killer disease in our country,and ICH is one of the important causes of stroke.Estimation of the volume of the suspected hematoma area is a key step in the treatment of patients.When a patient has severe neurological symptoms such as severe headache or loss of consciousness,the doctor will check the patient’s skull medical imaging images to confirm whether there is intracranial hemorrhage.If it does,the doctor will look for the location and type of the hematoma,and estimate the volume of the hematoma.This process is complicated and usually time-consuming.Therefore,the research of high-efficiency and automated Computed Tomography(CT)image segmentation algorithm for intracranial hemorrhage is an important basis for assisting doctors in estimating the volume of suspected hematoma regions,and provides necessary data support for the treatment of patients with intracranial hemorrhage.Although the Active Contour Model(ACM)has been widely used in the field of image segmentation due to its advantages that it does not require samples for training and can obtain complete closed object boundaries,since the CT image data of intracranial hemorrhage often have blend signs,black hole signs and Liqi island signs and other features that affect the segmentation results,the existing ACM cannot simultaneously achieve good segmentation results under the influence of various features.The thesis summarizes the features and segmentation difficulties of CT images of intracranial hemorrhage as: intensity inhomogeneity,low contrast and blurred boundaries,multiple targets and the existence of holes.Since the gray value of some suspected lesions in the intracranial part is not significantly different from the gray value of the extracranial part,this thesis first extracts the suspected hematoma area as the object of segmentation.In order to overcome the above segmentation difficulties,this thesis proposes an adaptive variable exponent ACM based on Legendre polynomial to segment the suspected hematoma area.However,when the image has strong local relevance,this method is more sensitive to the location of the initial contour.Therefore,we further consider the use of deep learning-based segmentation methods to solve the above problems.However,due to the lack of annotated datasets for intracranial hemorrhage and the small amount of data samples,it is impossible to effectively train an ideal pixel-level semantic segmentation model based on the neural network for segmentation.Therefore,we consider combining the advantages of ACM and CNN,extending the segmentation range to the full image,and propose a hematoma morphology segmentation method combining ACM and CNN.The experimental results prove that the proposed adaptive variable exponent ACM based on Legendre polynomial can simultaneously segment the suspected hematoma area CT images of intensity inhomogeneity,low contrast and blurred boundaries,as well as multiple targets and holes,which is better than the existing ACMs.In addition,the hematoma morphology segmentation method that combines ACM and CNN proposed in this thesis eliminates the intervention of artificial initialization on the basis of the above method,and combines with CNN to improve the performance,and has good practicability and robustness.The research content and contribution of this thesis can be summarized into the following two aspects.First,this thesis proposes an adaptive variable exponent ACM based on Legendre polynomial to segment the suspected hematoma area.First of all,this thesis extracts the Region of Interest(ROI)from CT images of intracranial hemorrhage,which not only avoids the interference of the extracranial part on the hematoma segmentation,but also improves the efficiency of ACM in segmenting the hematoma morphology.The proposed method uses a set of Legendre basis functions to fit the region intensity,so that the hematoma area with intensity inhomogeneity can be segmented.In addition,the proposed method also combines the gradient information of the image with the region information,and designs the variable exponent function to enable the exponent of the fidelity term in the energy functional to be updated adaptively,which overcomes the existing segmentation difficulties of low contrast,blurred boundaries,multiple targets and holes.Finally,the proposed method avoids false contours in the segmentation results by combining the distance regularization term.The experimental results prove that the proposed method can not only obtain relatively high segmentation accuracy,but also can spend relatively little calculation time.Second,this thesis proposes a hematoma morphology segmentation method combining ACM and CNN.First of all,since a certain number of data samples are required to train the deep learning model to output coarse segmentation results,data augmentation is required.Then,the proposed method uses CNN to learn the shape of the hematoma to be segmented,and outputs the coarse segmentation result,which is the probability map.Finally,the proposed method converts the probability map as the initial contour of ACM,and calculates the parameter functions through the probability map to adjust the parameter settings of the ACM,and then outputs the final accurate segmentation result end-to-end.The experimental results prove that the proposed method has superiority in the segmentation of intracranial hemorrhage CT images.It has better segmentation accuracy than the state-of-the-art segmentation methods and avoids the addition of artificial initialization intervention.