Research On Brain Tumor Segmentation Algorithm Based On 3D Convolutional Neural Network

Gliomas are the most common primary malignant tumors of the brain.Accurate segmentation and quantitative analysis of tumor areas are essential for diagnosis and treatment.There are huge differences in shape location and malignancy of brain tumors for different patients.To make full use of the spatial information of multimodal brain tumor magnetic resonance imaging to achieve accurate tumor segmentation,our research about automatic brain tumor segmentation is based on 3D convolutional neural networks.The main research contents are as follows:(1)To address the inconsistency in the shape,location,and size of brain tumors,this paper proposes an automatic MRI segmentation algorithm for brain tumors based on a two-channel threedimensional densely connected network.The algorithm is based on a three-dimensional convolutional neural network.The two channels use different sizes of convolution kernels to extract multi-scale features under different scales of receptive fields,and then we build densely connected blocks of each channel for feature learning and transmission.Finally the concatenation of two channel features was sent to classification layer to classify central region voxels to segment brain tumor automatically.Experimental results show that the proposed algorithm improves the segmentation accuracy of Deep Medic in tumor nuclear region and enhanced tumor region by 0.06 and 0.03,respectively.(2)Aiming at the segmentation error of some tiny tumor regions and tumor boundaries by 3D UNet,a brain tumor segmentation algorithm based on dense connection enhancement is proposed in this paper.The algorithm is based on the 3D U-Net structure,using convolution instead of max pooling to achieve downsampling;a densely connected block structure is added to the context module to enrich the multiscale features of the network;a new upsampling scheme is designed to replace the original upsampling methods such as interpolation,transpose convolution.Experimental results show that the improved segmentation network can retain multi-level features and extract richer scale features for voxel class prediction.Compared with the 3D U-Net segmentation results,the Dice score for the wholel tumor segmentation is increased by 0.009,The Dice score of the tumor core is increased by 0.07,and the Dice score of the enhanced tumor region is increased by 0.96.(3)Aiming at the problem of segmentation error caused by the class imbalance of brain tumors,this paper proposes a brain tumor segmentation algorithm based on VOI(Volume of Interest)optimization.Our algorithm includes two steps.Firstly,the tumor region is roughly segmented,and the whole tumor is extracted as VOI;then the VOI and multi-modality MRI are fused into the sub-region segmentation network.Experimental results show that the optimization algorithm can effectively reduce the under-segmentation and over-segmentation caused by the class imbalance.The accuracy of segmentation has improved the segmentation accuracy of each tumor sub-region,the algorithm's Dice score for segmentation of the tumor core area is increased by 0.031,and the Dice score for the segmentation of enhanced tumor areas is increased by 0.011.