Research On Mutil-modal MRI Brain Tumor Segmentation

Brain tumor is one of the highest mortality diseases in the world,and the survival time of patients with malignant brain tumor is not more than two years.Accurate localization and judgment of tumor regions and types in patients are important for diagnosis and the future treatment planning.MRI technique is widely applied in the clinical diagnosis of brain tumor as it provides high contrast and high resolution of the soft tissue,as well as large-field and multi-directional observations.Automatic and reliable brain tumor segmentation methods are urgently required since the traditional artificial analysis of MRI images can not meet the requirements of modern medical diagnosis.With the rapid development of deep learning in image processing,multi-modal MRI brain tumor segmentation based on deep learning has become a hot topic.This thesis aims to explore the multi-modal MRI brain tumor segmentation algorithms based on deep learning.The current research and some common approaches of multi-modal MRI brain tumor segmentation are presented.And two multi-modal MRI brain tumor segmentation methods based on deep learning are proposed to address the existing difficulties and challenges in this task.In clinical practice,physicians usually first segment the most distinct tumor area,and different modal data in multi-modal MRI brain tumor images are of different importance to specific brain tumor regions.Inspired by this,this thesis proposes a novel multi-task MRI brain tumor segmentation network based on multi-modal feature fusion.The multi-category MRI brain tumor segmentation,which is considered as a common semantic segmentation in most existing works,is treated as a multi-task segmentation.Thus,a multi-task cascaded segmentation network structure is designed to make full use of the relationship of each tumor area,but also alleviate the class imbalance between the tumor areas.Then the multi-modal aware feature embedding(MAFE)is introduced to fuse the multi-modal information by weighting them according to the importance to specific tumor area.The proposed model is implemented on the Tensor Flow framework in a Ubuntu 18 computer with a GTX1080 Ti GPU.Experiments on multi-modal MRI brain tumor data set show that the proposed method achieves outperforming performance in segmenting the desired brain tumor areas and demonstrate the components of the proposed method.Multi-modal MRI brain tumor data sets are usually small in scale.But the multi-modal data sets contain abundant modal feature information.To this end,this thesis proposes a novel brain tumor segmentation method based on multi-modal feature learning.The proposed multi-modality feature learning framework consists of two learning processes: the multimodality feature transition(MMFT)process and the multi-modality feature fusion(MMFF)process.the multi-modality feature transition process aims at learning rich feature representations by transiting knowledge across different modality data with a generative adversarial network.And the multi-modality feature fusion process aims at fusing knowledge from different modality data to predict the tumor areas.The proposed model is implemented on the Pytorch framework in a Ubuntu 18 computer with a GTX 1080 Ti GPU.Comprehensive experiments are conducted on two multi-modal MRI brain tumor data sets,which show that the proposed deep multi-modality feature learning framework can effectively improve the brain tumor segmentation performance when compared with baseline methods and state-of-the-art methods.