Segmentation Of Brain Tumor Based On Deep Learning

Brain tumors are one of the diseases that seriously endanger human life.The irregular shape and volume of tumors may appear anywhere in the brain,posing a serious threat to the life and health of patients.Limited to the current level of medical technology,the most effective solution is to detect and treat brain tumors in an early stage.Magnetic Resonance Imaging MRI,as a non-invasive technology,can provide high-resolution,high-contrast images without skull artifacts,and is widely used in clinical diagnosis and treatment.Doctors usually combine MRI imaging of different modal sequences,complementary information,and manually divide the specific location of the tumor to diagnose and treat the patient.However,manual segmentation relies on the doctor’s professional knowledge and clinical experience,and is time-consuming and labor-intensive.Therefore,the development of accurate and efficient automatic brain tumor segmentation algorithms is of great help to clinical medicine.Based on the convolutional neural network in deep learning,this paper combines the characteristics of multi-modal MRI image data to design an efficient and accurate neural network model for brain tumor segmentation tasks.Combining the current status and trends of brain tumor research in various fields,this article improves the algorithm from different perspectives such as network structure and loss function on the basis of 3D U-Net,explores the high utilization of multi-modal information of brain MRI images,and improves the network For the segmentation performance of details,a brain tumor segmentation algorithm with more accurate segmentation results can be obtained.The specific research content is as follows:1.Aiming at the problem that brain tumors have multiple regions and the boundaries of each region are not obvious,a multi-modal fusion brain tumor segmentation algorithm based on improved 3D U-Net is proposed.Using the differences of MRI images to complement the information,a cascade network architecture suitable for the automatic segmentation of brain tumors is constructed,which gradually completes the segmentation of brain tumors from coarse segmentation to fine segmentation.The network uses 3D U-Net as the basic structure and introduces residual connections to reduce information loss in traditional networks to a certain extent.In terms of loss function,this paper introduces category weights on the basis of Tversky loss function to obtain weighted loss function WTL,which improves the accuracy of segmentation for small targets.The experimental results show that on the Bra TS2018 validation datasets,the Dice coefficients of the proposed network in the entire tumor,tumor core,and enhanced tumor are 0.883,0.834,and 0.746,respectively,which are better than classic 3D U-Net and cascaded 3D U-Net.Aiming at the problem of low accuracy of brain tumor segmentation,a brain tumor segmentation algorithm based on 3D U-Net with dual attention mechanism is proposed.On the basis of 3D U-Net,the position attention mechanism and channel attention mechanism are introduced,and finally the dual attention mechanism 3D U-Net is obtained,which improves the network’s attention to key features and obtains more accurate brain tumor image segmentation results.In the encoder part,group normalization(GN)is used to replace the commonly used batch normalization(BN)operation to reduce the impact of too small batch size on network accuracy and improve network performance.At the same time,in the loss function part,a hybrid loss function sensitive to boundary information is proposed to process the detailed information of brain tumors to obtain a clearer tumor area.Finally,3D U-Net,Res Unet with residual connection and 3D AU-Net with dual attention mechanism are compared and tested on the Bra TS2018 datasets.The proposed brain tumor segmentation algorithm has Dice coefficients of 0.898,0.850 and 0.776 for the entire tumor,tumor core and enhanced tumor,respectively.The overall performance of the four algorithms is the best,which proves that the proposed method can significantly improve the segmentation accuracy of the network.