Research On MRI Brain Tumor Segmentation Based On Convolution Neural Network

Glioma is the most common intracranial tumor in medical clinical research,which poses great harm to patients and even poses a life-threatening threat.Multimodal magnetic resonance imaging contains complementary and comprehensive feature information,which is used as a preoperative diagnostic basis for glioblastoma diseases.Segmentation of brain tumor based on multimodal MRI is difficult and time-consuming.Compared to traditional brain glioma segmentation methods,deep learning based brain glioma segmentation methods do not rely on human experience and can quickly and accurately segment the location of the lesion area,improving clinical diagnosis and treatment efficiency.This paper proposes an improved deep residual attention algorithm based on the UNet network model to address the issues of insufficient feature extraction ability of multimodal magnetic resonance sequences,fuzzy tumor boundaries,and uneven image strength in glioma segmentation tasks.By increasing the depth of the network model,introducing attention mechanism,and residual mechanism,the segmentation performance of the algorithm model is improved,thereby achieving accurate segmentation of glioma.The specific research content is as follows:(1)In response to the problem of insufficient feature extraction ability in multimodal sequences,the feature layer in the encoder is replaced by an efficient convolutional neural network Efficient Net-B7,which uniformly reduces the depth,width,and input image resolution of the network convolutional layer.By adding feature layers,the model’s feature extraction ability is improved;(2)To address the issue of low segmentation accuracy caused by network deepening,a residual attention module is designed in the encoder and decoder connection section of the network model,which improves model performance through residual structure and attention mechanism;(3)Aiming at the problem of blurry boundary and uneven intensity of glioma image,an efficient channel attention mechanism is introduced in the decoder part,which can enable cross channel information interaction between feature layers,make the model focus on important feature information,and improve the learning ability of the model for micro Semantic information.In order to validate the model constructed in this article,the model was tested and evaluated on two datasets of the Brain Tumor Segmentation Challenge in 2019 and 2021,and compared with six classic segmentation networks.On the test set of the Brain Tumor Segmentation Challenge in 2019,the Dice coefficients of the whole tumor area,the tumor core area and the enhanced tumor area reached 0.8546,0.8773 and 0.7938 respectively,and the Hausdorff distance reached 2.5636,1.5392 and 2.7066 respectively,which were better than the control group;On the test set of the Brain Tumor Segmentation Challenge in 2021,the Dice coefficients of the overall tumor area and the tumor core area reached0.8394 and 0.7539 respectively,and the Hausdorff distance of the tumor core area and the enhanced tumor area reached 1.9759 and 3.0162 respectively,which were better than the control group.The experimental results indicate that the deep residual attention network model constructed in this paper can effectively improve the training efficiency and accuracy of brain glioma segmentation.