| Gliomas are the most common malignant tumor in the brain,accounting for 80%of all malignant brain tumors.Gliomas can be divided into glioblastoma(GBM)and lower-grade glioma(LGG).Accurate grading of GBM and LGG is crucial for treatment decision and prognosis for patients.At present,histopathological diagnosis after biopsy is the gold standard for glioma grading in clinical practice.However,there are some limitations due to its invasiveness,lake of timeless,and time-consuming.Magnetic resonance imaging(MRI)technology,which is a non-invasive and auxiliary method,can provide structural and functional information in clinical diagnosis,and plays an important role for glioma grading.Currently,many researches have studied glioma grading based on MRI,but they are limited to using single methodology.In this study,we mainly investigated the feasibility of a hybrid structure integrating radiomics features and deep features based on multi-modal MRI for glioma grading.In this paper,two hierarchical frameworks integrating radiomics features and deep features based on multi-modal MRI for glioma grading were proposed.For method one,the 2.5D orthogonal planes of multi-modal MR images,namely,the transverse,the coronal,and the sagittal plane,were taken as inputs.Then,radiomics and deep learning methods,were used to extract radiomics features and deep features from the three orthogonal plane images,respectively.The feature-of-bag algorithm was used to encode the two features of each orthogonal plane,which made features more compact and produced high-dimension sparse representation.Then,kernel fusion strategy was used to fuse the three orthogonal plane images,and the united representation of three orthogonal plane images could be obtained.Finally,a SVM classifier based on the fusion of the two features was constructed for glioma grading.For method two,3D VOIs were used as input.Then,radiomics model were used to extract global features of tumors and deep learning model were used to obtain deep local features.Relief algorithm was used to select global and local features respectively,so as to alleviate overfitting.Then,the kernel fusion method was used to combine the global radiomics features and local deep features from different sequences.Finally,a SVM classifier based on different features could be established for glioma grading.A total of 567 patients from two data sets were included in this study,one of which was the TCGA data set(GBM=106,LGG=127),which was classified into training cohort(GBM=85,LGG=101)and internal validation set(GBM=21,LGG=26)to train the model and select parameters.The other data set was from a local hospital(GBM=105,LGG=229),which was used as an independent external testing set to evaluate the performance of the models.All patients underwent T1-weighted post-contrast enhanced(T1ce)and T2 fluid-attenuated inversion recovery(T2 FLAIR)MRI scanning.The performance of the models were assessed using the area under receiver operating curve,sensitivity,specificity,and t-test.The experimental results indicated that the two methods show significant generalization ability at the internal validation cohort and the external testing cohort.The performance of method 2 was slightly better than that method 1,and both methods were better than those based only on radiomics features or deep features.The predictive results of this research were also comparable to the diagnostic performance of radiologists.This study demonstrated the feasibility of integrating radiomics features and deep features based on multi-modal MRI to develop a noninvasive model for glioma grading. |
PDF Full Text Request