Research On Brain Tumor Segmentation Technology Based On Multi-modal Fusion Res U-Net Neural Network

Medical imaging is the key technology of medical research,diagnosis of diseases,treatment planning,evaluation of treatment effect and other applications.Medical image acquisition technology has been widely used,such as Computed Tomography(CT),Magnetic Resonance Imaging(MRI),Ultrasound Scanning and other medical imaging technology has been widely used in medical diagnosis.Medical image processing is one of the important fields of medical image research.Medical image segmentation is one of the important research directions of medical image processing.The region of interest segmented from the medical image can be used as an important reference for doctors’ diagnosis and treatment.At present,there are many researches in the field of medical image segmentation based on traditional methods has implemented segmentation of medical image.In recent years,with the improvement of GPU computing power and the development of deep learning,image segmentation method based on convolutional neural network has been realized.With its end-to-end processing method and good segmentation effect,this kind of method has become the main medical image segmentation method.Compared with the implementation based on traditional algorithm,the segmentation method based on convolution neural network can realize the automatic segmentation of medical image,which greatly improves the speed and accuracy of image segmentation.This thesis mainly studies the application of convolution neural network in brain tumor segmentation.At present,there are still severe challenges in brain tumor segmentation.Firstly,the shape of glioma is changeable and invasive,and the boundary is very fuzzy,so it is not easy to segment.Secondly,the intensity of MRI image is not balanced,and the image intensity varied from different equipment and different imaging time.Finally,brain tumor has the problem of unbalanced category,and most areas of the image are non-tumor tissue.Based on the above problems,this thesis proposes a multi-modal Res U-Net image segmentation model for brain tumor MRI images segmentation.The main research contents of this thesis are as follows:(1)A Res U-Net brain tumor segmentation model based on multi-modal fusion is proposed.This model not only combines the idea of U-Net network structure and residual module,but also proposes a weighted multi-modal information fusion method according to the multi-modal image information of MRI image.The model is evaluated with other main methods on the Bra TS2017 dataset,and the results show that the proposed method is better than other methods in segmentation accuracy.(2)MRI brain tumor image segmentation is extended to three-dimensional scale,and 3D medical image segmentation is realized by a multi-modal Res U-Net based on three-dimensional convolution kernel.At the same time,by improving the network structure and image preprocessing,the influence of a large number of additional parameters brought by the dimension extension of the model on network training is reduced.Experimental results show that the proposed method is competitive in brain tumor MRI image segmentation.