Glioma Genotyping Based On Fusion Of Multiple Sequence MR Images

Correctly judging whether the genotype of brain glioma is a mutation or a wild type will help doctors make the correct prognostic treatment.In view of the fact that the biopsy will cause certain harm to the patient,and the artificial observation of the MRI image is low in accuracy,this article uses computer-assisted methods to judge gliomas.In this paper,for the purpose of glioma typing,multi-sequence MRI images as the data basis,and deep learning as the method,different deep learning network structures are proposed from preprocessing to tumor typing.The innovative work of this article includes the following:(1)For the preprocessing,3D Slicer and other software are time-consuming,laborious and low accuracy on MRI skull stripping operation.This paper proposes AD2MNet,an MRI skull stripping method based on anisotropic convolution and feature fusion.This network constructs a skull stripping neural network by combining anisotropic convolution,multi-view and multi-scale fusion methods.It has been trained and tested on the local data provided by the radiology department of Jiangsu Province Hospital.Dice value,Jaccard value,PPV value and Sensitivity value reached 98.92%,93.75%,95.35% and 97.11%,respectively.This method is compared with BSE,BET and other algorithms,and the generalization ability experiment is done,which proves that it has a high precision of skull stripping and a strong generalization ability.(2)In view of the fact that the biopsy will cause harm to the patient and the accuracy of the glioma genotype is low manually through MR images,a genotyping method D-ResNet based on the pyramid dilated convolution ResNet network is proposed.This method uses the dilated convolution to construct the the pyramid dilated convolution module to improve the shallow receptive field of ResNet.Through training and testing on the local data set and the public data set from TCIA,its accuracy rate is improved by 1.52% compared with the original ResNet network.In addition,this method has been compared with DenseNet,MLP and other algorithms,and generalization ability experiments have been done,which proves that the accuracy of the network has been improved and has a strong generalization ability.(3)In view of the fact that the annotation of 2DMR image data for genotyping is laborintensive and the sequence feature extraction is incomplete,the method of using 3D convolution,multi-sequence image fusion input and CA attention mechanism is proposed to improve the network.The improved network MA3D-ResNet is trained and tested on the local data set and the public data set from TCIA.Compared with the D-ResNet network,the accuracy rate is increased by 10.74%.In addition,through comparison experiments with C3D,DensNet3D and other algorithms and generalization ability experiments for predicting tumor grade,it is concluded that the network has high accuracy and a strong generalization ability.