¹⁸F-FDG PET/CT Radiomics Based Classification Of High-grade Glioma And Brain Metastasis

ObjectiveTo evaluate the diagnostic value of radiomic features derived from ¹⁸F-fluorodeoxyglucose(¹⁸F-FDG)positron emission tomography/computed tomography（PET/CT）images in differentiating high-grade glioma（HGG）from brain metastasis（BM）.MethodsSubjects with pathologically confirmed brain tumors between July 2013 and June2019 were retrospectively reviewed.A total of 97 brain tumor lesions（HGG,n=41;BM,n=56）of 74 patients who received preoperative ¹⁸F-FDG PET/CT and MRI examinations were retrospectively included.Tumors were segmented on T1-weighted contrast enhanced MRI images and ¹⁸F-FDG PET/CT based features were extracted using LIFEx packages.A procedure of stratified 10-fold cross validation with 10 repeats was used to construct and test radiomic models of PET,CT and PET/CT.A PETconv model using conventional features was constructed for comparison.The diagnostic performance of all models were evaluated in test set using receiver operator characteristic curve（ROC）method.The occurrence frequency of each feature in all feature selection procedures was counted.The Spearman rank correlation coefficients between texture features and tumor volume were calculated.P<0.05 was considered as statistically significant.ResultsArea under the receiver operating characteristic curve（AUCs）in classifying HGG and BM were 0.704,0.745,and 0.747 for PET,CT and PET/CT models,respectively.Prediction performances of radiomic models were superior to the AUC of 0.578 for PETconv model（all P<0.001）.CT radiomic model demonstrated a tendency to outperform PET radiomic model（P=0.08）,while PET/CT model was significantly better than PET radiomic model（P<0.05）.No significant difference was observed between PET/CT radiomic model and CT radiomic model（P=0.94）.During the model construction,statistical features derived from tumor edge（mean RIM3＿Hustdev,mean RIM3＿Humin and mean RIM3＿SUVmin）,features derived from grey level co-occurrence matrix（GLCM＿Correlation）and neighborhood grey-level different matrix（NGLDM＿Contrast）showed robustness in all feature selection procedures.The Spearman rank correlation coefficients between texture features and tumor volume were generally low to medium（|rs|<0.8）.ConclusionRadiomic features extracted from ¹⁸F-FDG PET/CT images can be used to better classify HGG and BM.The differential diagnostic value of radiomic approach significantly outperformed that of conventional PET model ¹⁸F-FDG PET/CT based radiomics may pave a new avenue for differentiating brain tumors.