Research On Medical Image Segmentation And Application For Cerebral Hematoma And Brain Tumor

Cerebral hemorrhage and brain tumors can cause a squeezing effect on normal intracranial tissues,severely damage the central nervous system,and endanger the life of the patient.The treatment of brain lesions is generally neurosurgery.Doctors need to diagnose and analyze the medical images of the lesions in the preoperative stage,and often use medical image segmentation techniques to assist in processing.However,the segmentation algorithm still has many challenges and difficulties.In brain hematoma segmentation,the boundary between the lesion and the normal tissue is blurred,and the gray information is similar,resulting in the segmentation result including normal tissue,which is prone to misdiagnosis;in brain tumor segmentation,the tumor includes infiltrating edema,tumor nucleus,and For the gangrene part,the segmentation algorithm is difficult to accurately determine the contours of each part of the brain tumor,and the huge amount of multi-modal data will increase the segmentation time and affect the doctor’s monitoring of the brain tumor.Aiming at the problems and difficulties of the current brain lesion segmentation algorithm,this thesis proposes an improved level set brain hematoma segmentation algorithm framework and a U-Net brain tumor segmentation algorithm combined with wavelet transform.The main innovations and improvements are shown in below:(1)The thesis uses the image pyramid and the minimum bounding box algorithm to perform twice Fast Marching cerebral hematoma segmentation,and the purpose is to improve the algorithm calculation speed and segmentation accuracy without adjusting the parameters often.This stage obtains the coarse segmentation result.In the fine segmentation stage,the thesis uses GMM-EM classification algorithm to retain the cerebral hematoma.Besides,in order to reduce the number of iterations in the GMM-EM calculation,K-Means++ algorithm is used to initialize the classification center,which reduces the calculation time of the GMM-EM.Compared with the same type of algorithms,the framework proposed in this thesis has a significant improvement in segmentation performance.(2)In the brain tumor segmentation algorithm,a discrete wavelet transform is combined into the lightweight design module of the Ghost Net to replace the convolution layer in the U-Net functional layer and purpose is to reduce the amount of parameters and retain the characteristics of brain tumors.In order to avoid the structural information of brain tumors in the the pooling and down-sampling process.An Haar wavelet transform is used to replace the pooling layer,and the high-frequency components are transformed correspondingly to the up-sampling process of the U-Net decoding path,which restore the structural information of the brain tumor.Finally,two neural network models are designed to segment the whole tumor and the tumor nucleus,which are connected by a cascade strategy to realize the end-to-end training process.Compared with the state of art algorithm,the multi-modal brain tumor segmentation algorithm proposed in this thesis is still competitive.(3)In addition,this thesis also proposes a loop calibration algorithm for mixed reality registration,which uses the SVD algorithm to register the virtual world coordinate and the real world coordinate,so that the virtual model and the real model can be successfully matched,which allows the doctor can observe the patient’s anatomy intuitively.In order to implement algorithms mentioned above into application,this thesis also develops a medical optical navigation system to carry out phantom experiments,simulates the neurosurgery process,and verifies the reliability and practicability of the segmentation algorithm.