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

Among human tumors,brain tumors can occur at any age,and the majority of patients with malignant tumors seriously threaten human survival and health.Most clinical medicine uses Magnetic Resonance Imaging(MRI)technology to mark tumor areas.Due to the variety of tumor shapes and sizes,clinicians have limited experience reserves,and marking different tumor areas is time-consuming and laborious.Therefore,it is necessary to use computer-aided diagnosis technology to help clinicians quickly segment brain tumor areas.In deep learning,convolutional neural network provides a new method for computeraided diagnosis.However,there are still some problems in segmenting brain tumor regions using computer-aided diagnosis techniques.Firstly,2D brain tumor segmentation methods severely ignore the spatial dimensionality of brain tumor MRI images.Secondly,the tumor area is small,and the shape and size are different,so it is relatively difficult to segment.Finally,the accuracy of the segmentation will also affect the patient’s life.To address the above issues,this thesis explores two segmentation algorithms for brain tumor MRI images based on 3D convolutional neural networks.(1)A deep supervision and attention-based 3D brain tumor MRI image segmentation network DAUnet is proposed.First,a BA module for extracting features is designed,which consists of a Bottleneck module and an attention module.Add attention to the feature map in the spatial and channel dimensions,and also join the residual connection to fuse the original feature map with the noticed feature map.The attention here is called 3D SC attention.Adding 3D SC attention to the BA module helps DAUnet to extract more useful information.Secondly,the CASP module is designed to increase the receptive field of the feature map without changing its resolution.After using the standard convolution,the Atrous Spatial Pyramid(ASP)module is added to strengthen the association between different layers of the network.Finally,DAUnet uses deep supervision as an auxiliary branch,which combines deep learning and regularization techniques to supervise the model during training and automatically adjust parameters so that the model can better fit the training data.The use of deep supervision improves the robustness and generalization ability of the model.Through experiments on BraTS 2020 and FeTS 2021 datasets and comparison with other advanced methods,it is proved that DAUnet can accurately segment tumor regions in brain MRI images.(2)A multi-branch attention-based segmentation network MBANet for 3D brain tumor MRI images is proposed.First,the optimized shuffling unit is used to form the Basic Unit(BU)module of MBANet.In the BU module,the group convolution operation is used after the input channels are separated,and the channel shuffling is used to interfere with the convolution channels after fusion.Second,MBANet employs a novel multi-branch 3D Shuffle Attention(SA)module as the attention layer in the encoder.The 3D SA module is grouped along the channel dimension to divide the feature map into small features.For each small feature,the 3D SA module constructs channel attention and spatial attention while adopting the BU module.Finally,in order to strengthen important semantic features and suppress unnecessary features,the 3D SA module is also used in the skip connection of MBANet.Experiments on the BraTS 2018 and BraTS 2019 datasets show that the Dice of enhance tumor,whole tumor and tumor core reaches 80.18%,89.80%,85.47% and 78.21%,89.79%,83.04%,respectively.The excellent segmentation performance of MBANet shows that compared with other state-of-the-art segmentation methods,MBANet has significantly improved.