MRI Brain Tumor Image Segmentation Based On 2D/3D Medley Convolution

The image segmentation task of MRI brain tumor usually contains four targets:background region,necrotic region,whole region,and core region,and the boundary of each target region is blurred,which requires specialized medical knowledge and may lead to recognition errors due to subjective factors.The incidence of brain tumors is increasing,which has irreversible damage to the human nervous system.To meet this challenge,2D and 3D neural networks for brain tumor image recognition have been rapidly developed.According to different convolution methods and different feature extraction ranges,brain tumor segmentation is usually divided into 2D brain tumor segmentation and 3D brain tumor segmentation.In this paper,by analyzing the brain tumor features,we find that its data set has both planar features and residual 3D spectral information.Based on this theory,we propose a two-way medley convolutional network DMC-Net that can extract 2D and 3D features in parallel.The network achieves feature extraction by parallelizing the main path and the branch path,and uses the two branches to extract 2D features and 3D features respectively,and performs feature fusion effectively.The 2D main path network,with the proposed Mini-At Res Net50,extracts planar 2D features and uses the U-Net++ upsampling operation with its unique jump connection to fully utilize the features extracted by MiniAt Res Net50 and fuse them,so that each layer of features is utilized and feature information is not wasted.3D branched network first extracts the features through 3 times3 D separable convolution operations to extract higher dimensional features,after which the output of each layer of convolution is fused with features separately,and finally the output of the 3 separable convolutions is fused in the whole,and the global output result of the 3D branched network is directly self-fused with the output of the 2D branched after one layer of convolution,thus realizing the feature extraction of 2D information and 3D information.Based on Bra Ts dataset,this study compares the performance of the designed DMCNet with four classical and advanced biomedical image segmentation networks from both quantitative evaluation metrics and qualitative segmentation effects,and confirms that the DMC-Net based on 2D and 3D medley has higher feature extraction capability for brain tumors and significantly reduces the number of parameters and computation compared with Trans UNet.Compared with Trans UNet,DMC-Net has significantly reduced the number of parameters and computational energy consumption,and requires less hardware configuration,which makes DMC-Net have better performance when processing brain tumor MRI images.Figure [43] Table [11] Reference [60]...