Research Of Brain Glioblastoma Segmentation Method Based On Deep Learning

Glioma is the most common primary brain tumor,which poses a huge threat to the health of patients.The segmentation of brain glioma based on MRI is an important means to assist doctors to observe,analyze and diagnose the external morphology of brain glioma.In the current methods of brain glioma segmentation,the traditional image processing and machine learning methods are not ideal,so the deep learning based methods are widely used in brain glioma segmentation.And in the methods of glioma segmentation based on deeplearning,the full convolution network model has a good segmentation effect of glioma segmentation.In order to solve the problems of small receptive field,shallow model depth and large information loss in encoding and decoding process of full convolution network model in glioma segmentation method based on deep learning,the 2DResUnet(2D Residual Block Unet)model is proposed.The model adds ResBlock mechanism on the basis of Unet model,adds Gauss noise layer to the input layer of the model for data enhancement,and uses twodimensional convolution layer to replace the pool layer,so it has stronger feature extraction ability than the original Unet model.At the same time,the sum of Generalised Dice Loss and Weighted Cross Entropy is used as the loss function in the training process,which can alleviate the imbalance of categories of glioma data and effectively segment glioma.Aiming at the deficiency of acquiring three-dimensional spatial information in twodimensional full-convolution network and the problem of memory consumption in threedimensional full-convolution network,the DM-DA-Unet(Dual Multidimensional Dense Attention Unet)model for glioma segmentation is proposed.The model uses full convolution network of different dimensions to enhance the segmentation of glioma in different stages.The mechanisms,such as DenseBlock,Attention,and multi-scale fusion,are used to optimize the model structure,and fixed area sampling is used to reduce the memory consumption of three-dimensional convolution network.Therefore,multisequence information of glioma image can be fully utilized to further improve the segmentation accuracy of glioma.The model segmentation effect is evaluated with the BraTS dataset.Local 5-fold crossvalidation is performed using the BraTS18 dataset and the official online assessment is performed using the BraTS17 dataset.The results of the evaluation show that the average Dice Score of edema,core and enhanced regions of DM-DA-Unet on the BraTS17 validation set reaches 0.90,0.80,0.74,and the average sensitivity reaches 0.89,0.77,0.75.This result is close to the best model segmentation result on the current BraTS17 validation set and the DM-DA-Unet can accurately segment gliomas.