Objective:To explore the value of MRI radiomics model based on deep learning features in the differential diagnosis for skull-based meningioma.Methods:Cohort 1 comprised MRI data of 361 patients,including 208 patients with skull-based meningiomas and 153 patients with other skull-based tumors from one hospital.Cohort 2 comprised MRI data of 129 patients,including 48 patients with skull-based meningiomas and 51 patients with other skull-based tumors from another hospital and 30 patients with meningioma from the Meningioma-SEG-CLASS dataset in The Cancer Imaging Archive(TCIA).All tumors were classified into 5 categories,solitary fibrous tumors/hemanyiopericytomas(label_0),meningiomas(label_1),lymphomas(label_2),metastatic tumors(label_3),cartilage-derived and other similar tumors(label_4).Cohort 1 was divided into the training set(299 cases)and the test set(62 cases)at a ratio of 8:2,cohort 2 served as the external validation set.T2WI and enhanced T1WI images were cropped at the maximum level of tumors and input into convolution neural networks(ResNet50,DenseNet121,Inception v3)for training,the deep learning features extracted from networks were combined for developing MRI radiomics model.A multilayer perceptron(MLP)was selected as the classifier to evaluate the model's performance,and then proceeded with external validation.Results:Compared with convolution neural networks,MRI radiomics model based on deep learning features significantly improved the classification prediction performance of skull-based tumors.The predictive accuracy in the external validation set was 0.829.Multiclass ROC curve showed AUCs of label_0,label_1,label_2,label_3 and label_4 were 0.94,0.97,0.91,0.93 and 0.86,respectively.Conclusion:MRI radiomics model based on deep learning features has good performance and robust generalization capacity in distinguishing skull-based meningioma.
万方数据
维普
