The application value of pre-trained deep learning neural network model in differentiating central nervous system tumors
Objective To build a pre-trained deep learning neural network model based on MRI T1 weighted enhanced sequence;to evaluate the diagnostic performance of identifying the primary central nervous system lymphoma(PCNSL),glioblastoma(GBM)and brain metastase(BM).Methods A retrospective analysis were conducted on the patients who underwent surgical treatment and were diagnosed with PCNSL,GBM,and BM by postoperative pathological examination from January 2015 to June 2022 at the Neurosurgery Department of the First Affiliated Hospital of China Medical University.A total of 149 cases were enrolled according to the inclusion standard,including 43 PCNSL,62 GBM,and 44 BM.Those were randomly divided into a training set and a test set.After obtaining imaging data,the Slicer software was used for delineating regions of interest and pre-processing was conducted.A deep learning neural network model based on EfficientNetV2 with pre-trained and fine-tuned on ImageNet datasets was constructed,and then trained on the training set by transfer learning.The model was validated on the testing set,and then the classification performance metrics was evaluated by receiver operating characteristic(ROC)curves,accuracy,recall(sensitivity),precision,specificity,F1 score,and area under the ROC curve(AUC).Results The overall accuracy of the model was 89.4％and the average AUC was 0.96.The sensitivity of the model for PCNSL,GBM,BM were 0.893,0.936,and 0.778 respectively.The specificity were 0.996,0.877,and 0.955 respectively.The F1 score were 0.940,0.900,and 0.778 respectively.The AUC(95％CI)were 0.98(0.946-0.997),0.96(0.932-0.986)and 0.95(0.905-0.980)respectively.Conclusion The pre-trained deep learning neural network model trained on T1-weighted enhanced MRI scans has an outstanding diagnostic performance for PCNSL,GBM and BM.
Central nervous system neoplasmsDiagnosis,differentialLymphomaGlio-blastomaBrain metastasisDeep learning
NETL
NSTL
万方数据
