Medical Image Segmentation Of 3D Brain Tumor Based On Convolutional Neural Network

Brain tumors,as a major disease affecting the health of patients,cause about100,000 people suffering from illness every year.Because Magnetic Resonance Imaging(MRI)technology is harmless to the human body and can realize three-dimensional imaging and multi-mode imaging,it is often used in clinical diagnostic processes.However,for different patients,the shape and distribution of brain tumors are different,and doctors inevitably have subjective factors when conducting diagnosis.So the focus of researchers has always been how to use medical imaging to achieve a fine segmentation of brain tumors.Based on the above background,this thesis studies and experiments on the fine segmentation of medical images of three-dimensional brain tumors and mainly carries out the following work: First,because the traditional fine segmentation method can not fully utilize the tumor data with positional inclusion relationship,and can not combine the multimodal image information well,I divide the complex multi-class segmentation task into three simple two-class segmentation tasks,which makes false positive points other than tumors are excluded.Second,for the shortcomings of single-scale feature prediction that image detail information cannot be fully preserved,I adopt multi-scale feature fusion method to supplement image detail information to improve segmentation accuracy.Third,for the problem of small amount of 3D medical image data,I perform two-dimensional cropping on the data,and the input adopt multi-frame two-dimensional data,which greatly expands the data amount and can suppress over-fitting very well.In addition,intraframe information and interframe auxiliary information are extracted by combining interframe convolution and intraframe convolution.Fourth,in order to solve the problem of segmentation inaccuracy of small cell tumors in the experimental results,I use residual structure,dilated convolution and skip layer to combine shallow features and deep features,and I use different downsampling times for different tumor regions to further improve the segmentation performance.In this thesis,Bra TS 2018 is the data set used in the experiment.I design several sets of contrast experiments and prove the advantages of the cascaded network,camparing different scale features and different network structures.The final Dice coefficients of the tumor whole,tumor core and enhancing tumor were 0.9029,0.8317,and 0.7865,respectively.The algorithm proposed in this thesis can achieve high precision segmentation and achieve high performance.