Methodology Research On Segmentation Of Brain Tumor Image Assisted By Deep Learning

Clinical practice shows that MRI is very valuable for the evaluation of glioma.They will provide accurate brain tumor imaging and facilitate doctors in obtaining exact brain tumor contour.Usually,imaging experts manually label and segment tumor areas,but manual labeling by experts will bring problems such as labeling time-consuming and labeling differences.While retaining the high segmentation accuracy of tumor images,it brings a fast and reproducible image segmentation scheme for the delineation of tumor targets,which has become an urgent demand in image-guided radiotherapy technology.In recent years,with the continuous development of deep learning technology,the automatic image segmentation method based on deep learning has also made rapid development,which brings an opportunity for automatic and high-precision segmentation of tumor targets.This paper takes the MRI multimodal image of the Bra TS data set as the segmentation object.It uses a convolutional neural network to realize and optimize brain tumors’ automatic and accurate segmentation.The target region is automatically segmented from the MRI multimodal image by establishing the convolution neural network segmentation model.This paper studies a two-dimensional image segmentation algorithm based on UNet++.Firstly,the improved residual module in Resnet is integrated into U-Net and UNet++.The enhanced residual module is used to retain the feature that the minimum feature map contains rich semantic information to improve the feature extraction ability of the network encoder.The improved U-Net is used to train the two-dimensional slice data of Bra TS 2018.Because the enhanced U-Net and the improved UNet++ have part of the same network structure,the migration learning idea is introduced to migrate the trained improved U-Net network weights to the same corresponding position in the UNet++network to realize the initialization operation of the consequences,speed up the convergence speed of the network and improve the segmentation accuracy.The two-dimensional slice data are trained and verified in Bra TS 2018.The simulation results show that this method can achieve effective brain tumor segmentation.The 3D segmented data of Bra TS 2018 is introduced and verified,and the generalization of the improved method is proved by simulation experiments,which still has a good effect on 3D UNet++.While maintaining the target fidelity,a more lightweight 3D UNet++ network is designed.Firstly,MRI images are preprocessed to perform standardization,cross blocking,and data enhancement to eliminate the contrast difference between various modes and improve the network’s generalization ability.Then a lightweight residual structure is designed to replace the double-layer convolution structure in UNet++,which significantly reduces the amount of calculation and parameters compared with the direct use of the double-layer convolution structure network.This paper constructs a lightweight similar residual structure to lighten the similar residual structure.Finally,the CBAM attention mechanism is used to screen the information of the feature map after a long connection.The network can automatically pay attention to practical details and further improve the accuracy of network segmentation.By testing the 3D block data of Bra TS 2018,compared with the direct use of the improved 3D UNet++ network,the parameter of the lightweight convolutional neural network designed in this study is 1 / 13.4 of the improved 3D UNet++network,and the amount of calculation is 1 / 3.7 of the improved 3D UNet++ network.The parameter is reduced by 31425331,and the count is reduced by 1817.96 G so that the actual medical deployment has lower requirements for equipment and faster reasoning speed.At the same time,experiments show that this method effectively improves segmentation accuracy.