A Research On Multimodal MRI Image Segmentation Of Brain Glioma

Brain Glioma is the most common malignant tumor of the skull caused by cancer of glial cells in the brain and spinal cord,accounting for 35.2% to 61% of all intracranial tumors.It is characterized by high morbidity,high recurrence rate,high mortality and low cure rate.Magnetic resonance imaging provides clinicians with a wealth of critical information about brain tissue structures and is commonly used in the diagnosis of glioma.However,manual segmentation of gliomas is still the main option under the clinic,but it has the disadvantages of being time-consuming and labor-intensive.Meanwhile,malignant brain tumors have characteristics such as infiltrative nature and obscure borders,which make manual segmentation more difficult.In recent years,deep learning methods based on multimodal MRI brain glioma segmentation have achieved significant performance improvement compared with traditional machine learning methods,and have become the mainstream method in this field and continue to show good development.In this paper,we investigate multimodal MRI glioma segmentation algorithms based on deep learning,and the results are as follows.(1)A multilayer multitask cascade brain tumor segmentation model based on a priori knowledge is proposed to address the problems of uneven pixel class granularity,easy destruction of contextual information during convolution,and difficult segmentation of tumor core region and enhanced tumor core region in multimodal MRI brain tumor image segmentation.The model splits the segmentation task into three steps and performs the binary segmentation of Whole Tumor,Tumor Core,Enhancing Tumor Core simultaneously.Each subtask network uses an optimized U-Net as the base network,and the optimization methods include introducing residual structure,embedding BN layers,using Leaky Relu activation function,using weighted Dice loss function,and introducing Poly learning rate decay.The key points are: cascade operation of multi-layer multi-task at the upsampling of the latter two task networks,transferring the same layer prior knowledge of the previous task to the subsequent network in the first place,completing the guidance of prior knowledge,and suppressing the interference of useless category information.The cascade operation and jump connection can complete the fusion of high and low dimensional information for the model to perform more realistic feature learning.At the same time,the convolution process of the network avoids clipping and uses "SAME" convolution to guarantee the contextual information integrity.Finally,the effectiveness of the proposed model and each improvement is verified through various comparison experiments on the Bra TS dataset.(2)A brain tumor segmentation model that fuses attention mechanism and dilated convolution is proposed to address the problems of unbalanced and multi-scale lesion categories,small convolutional receptive fields,and slow training speed in the segmentation process.The model uses the optimized U-Net as the base network,and introduces the Atrous Spatial Pyramid Pooling module(ASPP)on it that can extract multi-scale image features and fuse multi-scale receptive-field to enhance the learning ability of the model for lesions of different sizes.And the Improved Bottleneck hybrid Attention Module(IBAM)is introduced to realize the attention calibration of the original feature map by the optimized Channel attention module and the spatial attention module which conforms to the HDC design guidelines together.Meanwhile,the Focal loss function is introduced to avoid the category imbalance problem.Finally,by conducting comparison experiments with other methods on the Bra TS dataset,the Dice metrics in the Whole Tumor,Tumor Core,Enhancing Tumor Core reach 0.9106,0.8625 and 0.8368,respectively.it is proved that the two improved modules introduced can improve the segmentation ability of the model respectively,and the segmentation ability of the model reaches the mainstream level.