Research And Application Of Glioma Image Segmentation Algorithm

Gliomas are the most common primary cranial tumours that pose a serious threat to human health.Magnetic resonance imaging(MRI)technology provides intracranial images of brain tumours,which giving doctors tremendous support in diagnosis and treatment.Achieving accurate segmentation of gliomas has positive implications for doctors’ diagnosis and treatment.Due to the diversity in size,shape and location of brain tumours,and the complexity of their structure,as well as the great variation between patients,it is still difficult to identify and segment gliomas based on MRI images.Traditional methods are time-consuming and unstable,and single-modality MRI images do not provide complete information about gliomas.Therefore,multi-modal MRI images are often integrated to identify and segment gliomas.Based on the above background,this paper uses multi-modal MRI images as the basis for research,and combines deep learning algorithms to achieve automatic segmentation of gliomas,this paper completed the following works:(1)A 3D U-Net model based on dilated dense block is proposed to achieve automatic segmentation of multi-modal MRI glioma images.U-Net has good segmentation performance and is often used in the field of medical image segmentation.However,the special characteristics of glioma make the U-Net model unable to obtain more detailed information in segmentation task.To this end,this paper incorporates the dilated convolution and dense connection blocks on the basis of the U-Net model,and uses the sum of the Generalized Dice Loss function and the Cross-Entropy Loss function as the loss function of the network to alleviate the category imbalance problem in glioma image segmentation.The algorithm was validated on the Bra TS2018 dataset,and the Dice Similarity Coefficients(Dice)achieved on the whole tumor region,tumor core region and enhancing tumor region were 88.59%,82.43% and 78.67% respectively,which were demonstrated experimentally to have better segmentation performance.(2)A model based on recurrent multi-fiber network is proposed to achieve automatic segmentation of multi-modal MRI brain glioma images.Three-dimensional convolutional neural networks have good segmentation performance,but consume a large amount of computational cost and require high hardware resources.Based on the above problems,a lightweight segmentation network is designed.The network uses fiber units to reduce the computation and the storage footprint,solving the high computation cost caused by using 3D convolutional neural network;it uses recurrent multi-fiber modules to improve the network’s ability to extract contextual information;and it uses multiplexer modules to enhance the information exchange between modules.The recurrent multi-fiber network combines the advantages of fiber units,recurrent multi-fiber modules and multiplexer modules to achieve accurate segmentation of gliomas.The algorithm was validated on the Bra TS2018 dataset,with Dice values of89.62%,83.65% and 78.72% for the whole tumor region,tumor core region and enhancing tumor region,respectively.The amount of calculation and parameters were37.24 G and 2.65 M,respectively.Experimental results show that the algorithm has better segmentation performance and significantly reduces the memory consumption.