Research And Application Of Brain Tumor Image Segmentation Method Based On MDC VNet

Brain tumors are common malignant tumors that form in the brain and cause serious damage to brain function.The resulting mortality rate is increasing year by year,and more than 2 million people die from brain tumors every year.Therefore,early diagnosis and treatment are very important to improve the survival rate and quality of life of patients.However,at present,brain tumor image segmentation still needs to rely on manual segmentation,which is time-consuming,laborious,and has a high rate of misjudgment.In recent years,deep learning technology has been widely used in the field of medical image segmentation.Convolutional neural network and its combined brain tumor segmentation method can significantly improve the quality of tumor image segmentation.However,due to the characteristics of fuzzy contours,complex details and irregular structure of brain tumor images,and the small sample size of medical images,it is difficult to improve the accuracy of brain tumor image segmentation.Therefore,aiming at the difficulties and problems of brain tumor image segmentation,this thesis proposes a new method,that is,MDC VNet segmentation method using multi-depth fusion,deep supervision and CBAM attention mechanism to improve the accuracy of brain tumor image segmentation.The research content of this thesis is as follows:(1)An improved brain tumor segmentation algorithm MDC VNet based on VNet is proposed.Based on VNet,the model introduces multi-depth fusion to enrich the detailed information acquisition of each layer.Deep supervision at multiple stages of upsampling allows shallow networks to be adequately trained.In addition,the CBAM attention mechanism is introduced after upsampling,which can obtain more detailed information of tumor images.In order to further optimize the model,a smoothing optimization is performed based on the Relu activation function,and a new smoothing exponential function SEF is designed.In addition,BN is improved and a new batch normalization method DBN is proposed.In the choice of loss function,two combined methods are used to verify the best combination strategy.One is to combine BCE Loss(Binary Cross Entropy Loss)as a new loss function on the basis of Dice Loss,and the other is to combine Focal Loss as a new loss function on the basis of Dice Loss.(2)We conducted a brain tumor dataset experiment and analyzed the results.A set of data augmentation rules was formulated and the hyperparameters were fine-tuned using the combined strategy of Nadam and Sgd optimizers,which improved the stability and robustness of the brain tumor model training process.In terms of evaluation indicators,we selected three indicators: Dice Similarity Coefficient,Sensitivity and Specificity.By comparing the effects of different loss functions,it is found that the combination of Dice Loss and Focal Loss is the best combination of loss functions in this experiment.(3)We implemented a brain tumor image segmentation system.A set of data augmentation rules was formulated and the hyperparameters were fine-tuned using the combined strategy of Nadam and SGD optimizers,which improved the stability and robustness of the brain tumor model training process.In terms of evaluation indicators,we selected three indicators: Dice Similarity Coefficient,Sensitivity and Specificity.By comparing the effects of different loss functions,it is found that the combination of Dice Loss and Focal Loss is the best combination of loss functions in this experiment.The MDC VNet model proposed in this thesis achieved 92.15%,90.92% and 91.02% in terms of DSC,Sen and Spf indicators,respectively.And compared with various image segmentation methods,we found that the method proposed in this thesis has achieved the best results in multiple evaluation indicators,and successfully improved the accuracy of brain tumor segmentation.