Research On Segmentation Algorithm Of Brain Tumor MRI Images Based On Deep Learning

MRI images of brain tumors are an important basis for doctors to diagnose and treat brain tumors.Segmentation of brain tumors is a highly professional task,which needs experienced doctors.The task of segmenting brain tumors from MRI images is cumbersome and time-consuming.Meanwhile,the number of professional doctors is limited,but the number of patients is increasing.Therefore,researchers have proposed many automatic brain tumor segmentation methods to reduce the pressure of doctors and segment brain tumors more accurately.With the rapid development of deep learning,segmenting brain tumors with deep learning models has become a research hotspot,and many excellent models have emerged.However,most studies have the following problems:(1)Directly apply excellent models in other fields are to brain tumor segmentation task,and don’t improve these models according to the characteristics of brain tumor segmentation task;(2)The segmentation results often have a large number of false negatives;(3)The structure of models is more and more complex.This thesis builds CLCU-Net and LGMSU-Net based on deep learning to cover the shortages in some models for brain tumor MRI image segmentation.Compared with UNet,CLCU-Net adds cross-level connections on the encoding path and between the encoding path and the decoding path to fuse multi-scale features and restore detailed features.Meanwhile,a novel segmentation attention module is placed on the connections of CLCU-Net for selectively aggregating features,specifically,fuse useful information and discard useless information.LGMSU-Net is based on CLCU-Net and uses a novel self-attention module which can extract global information to reasonably and efficiently fuses the local features extracted by convolution kernels,the global features extracted by self-attention modules and the multi-scale features extracted by cross-level connections for brain tumor segmentation.This thesis extracts 176700 axial slices from the BraTS 2018 dataset to train,validate and test models.Subtract the mean and divide by the standard deviation within the brain region on each slice to solve the problem that intensities in different modalities and patients are various.Moreover,during training,employ random rotations,elastic deformation,and gamma augmentation to increase the amount of data.Dice score,precision,recall,the number of parameters of every model and the time spent segmenting a case are selected,and many ablation experiments and comparative experiments are carried out on the proposed two models and nine existing advanced models,which show that:(1)CLCU-Net and LGMSU-Net can segment brain tumors effectively and they are advanced in the existing brain tumor segmentation models;(2)Effective fusion of multi-scale features can improve the performance of brain tumor segmentation models;(3)Effectively fusing local features,global features and multi-scale features can improve the performance of brain tumor segmentation models.The experimental methods and results of this thesis can be used as a reference for other researchers to promote the development of brain tumor segmentation algorithms based on deep learning.