Research On Deep Learning-based Brain Tumor MR Image Registration

The brain tumor is one of the diseases that seriously endanger the lives and health of patients.However the complexity of brain structure and the heterogeneity of brain tumor tissues pose significant challenges for doctors to diagnosis,treatment and prognosis estimation.Magnetic resonance imaging is widely used by doctors in clinical examinations because of its advantages of non-radiation,tomographic imaging in any orientation,and good imaging effects on structural tissues.MRI can assist doctors in formulating brain tumor surgery plans and estimating prognosis,thereby reducing functional damage and postoperative complications caused by surgery,improving the quality of life of patients and ensuring further neurological recovery after surgery.Brain atlas has always been an important means to study brain structure and function.Using brain atlas to register brain images can help understand the brain specific state of each brain area,thereby guiding surgery and radiotherapy.However the current registration of brain images is mostly based on normal brain images.The registration of brain tumor images and brain atlases has always been a challenge due to the heterogeneity of tumors.In order to analyze the tumor brain image with reference to the brain region image of the brain template,this thesis takes the registration of the brain tumor MR image and the normal brain atlas as the research object,and uses the deep learning to construct a brain tumor tissue recovering method based on convolutional neural network,and improves the registration accuracy of brain tumor images and brain atlases.The main research contents of this thesis are as follows:(1).Realize tissue repair and registration of brain tumors MR images.To study how to ensure the accuracy of registration between the brain tumor image and the normal brain,this thesis proposes to use Partial Convolution Network(PConv-Net)to complete the tissue repair of the tumor area on the brain tumor MR image,and then register it with the brain atlas.This method first uses a segmentation network to segment the tumor,generating different sizes and positions tumor segmentation masks of the brain tumor MR image,and then uses PConv-Net to simulate the normal tissue of the tumor segmentation mask area,finally registers the recovered tumor image with the normal brain image.The experiment results based on Brats 2018 glioma data show that the method in this thesis can recover the tumor area well and obtain accurate registration results.(2).Using the region of interest(ROI)of the atlas to quantitative analysis the brain tumor images and determine the infiltrated area by the tumor.In this study,the patient’s original tumor image and the image based on tissue recovered are registered with the template image,respectively.The deformation fields obtained from the registration are used to map the ROI of the template image to the brain tumor image and the recovered tumor image.By comparing their ROI we can complete the quantitative analysis of the specific brain areas state,assisting in selection later surgical plans and treatment methods.The experimental results show that for brain tumor images with different shapes,locations well and sizes,the method proposes in this thesis can analyze the brain regions of tumor lesions by comparing the brain regions of brain tumor images before and after recovering.