| Objective:To investigate the diagnostic performance of conventional radiomics models based on 18F-FDG PET/CT,68Ga-NOTA-PRGD2 PET/CT and 99mTc-3PRGD2 SPECT/CT images,and deep learning-based radiomics models based on 18F-FDG PET images in the identification of benign or malignant pulmonary space-occupying lesions and the pathological subtypes of lung cancer.Methods:Retrospective analysis was performed on 64 cases with pulmonary space-occupying lesions from Peking Union Medical College Hospital from 2011 to 2013,48 of which enrolled in the clinical trials of 68Ga-NOTA-PRGD2 PET/CT and 18F-FDG PET/CT,and 16 of which enrolled in the clinical trials of 99mTc-3PRGD2 SPECT/CT and 18F-FDG PET/CT.All the cases were confirmed by surgical pathology or puncture biopsy.LIFEx v5.10 software was used to manually depict VOI and extract image features.For the 48 cases which received 68Ga-NOTA-PRGD2 PET/CT and 18F-FDG PET/CT scanning,the diagnostic performance of image features in the identification of benign or malignant pulmonary space-occupying lesions was evaluated.The difference of image features between the benign and malignant groups was compared by Mann-Whitney U test.Random forest models for identification of benign or malignant lesions were established,and cross-validation method was used to evaluate and compare these models.For the 30 cases among them with lung adenocarcinoma or squamous carcinoma,the diagnostic performance of image features in the identification of lung adenocarcinoma or squamous cell carcinoma was evaluated.The difference of image features between the lung adenocarcinoma and squamous cell carcinoma groups was compared.Random forest models for identification of lung adenocarcinoma or squamous cell carcinoma were established,and cross-validation method was used to evaluate and compare these models.For the 16 cases which received 99mTc-3PRGD2 SPECT/CT and 18F-FDG PET/CT scanning,random forest models for identification of benign or malignant lesions were established,and leave-one-out method was used to evaluate these models.For the 9 cases among them with lung adenocarcinoma or squamous carcinoma,random forest models for identification of lung adenocarcinoma or squamous cell carcinoma were established,and leave-one-out method was used to evaluate these models.For all 64 cases,2D images of the lesions on 18F-FDG PET were captured.Then a deep learning model and a random forest model were established for identification of benign or malignant lesions.Delong test was performed to compare the difference of their ROC curves.For 39 cases among them with lung adenocarcinoma or squamous carcinoma,a deep learning model and a random forest model were established for identification of lung adenocarcinoma or squamous cell carcinoma with the same method,and the difference of their ROC curves was compared.Results:There were significant differences(P<0.05)in 18F-FDG PET STLG?18F-FDG PET GLZLM HGZE?18F-FDG PET GLZLM SZE?18F-FDG PET GLZLM SZLGE?68Ga-NOTA-PRGD2 PET Sphericity between the benign and malignant groups.And there was no statistical difference(P>0.05)in CT HUmin.Their AUC values were 0.836,0.834,0.825,0.816,0.811,and 0.833,respectively.There were significant differences(P<0.05)in CT GLZLM LZE?CT GLZLM LZHGE?68Ga-NOTA-PRGD2 PET HISTO Entropy log10?68Ga-NOTA-PRGD2 PET HISTO Energy Uniformity?18F-FDG PET TLG?18F-FDG PET NGLDM Coarseness between the lung adenocarcinoma and squamous cell carcinoma groups.Their AUC values were 0.892,0.892,0.864,0.864,0.852 and 0.852,respectively.The AUC values of the models to predict benign or malignant lesions based on CT features,18F-FDG PET features,68Ga-NOTA-PRGD2 PET features,CT and 18F-FDG PET features,CT and 68Ga-NOTA-PRGD2 PET features,18F-FDG PET and 68Ga-NOTA-PRGD2 PET features,or all features,were 0.884,0.920,0.862,0.951,0.895,0.919 and 0.951,respectively,and the AUC value of the three-modal model was significantly higher(P<0.05)than that of the single-modal models.The AUC values of the models based on the same features to classify lung adenocarcinoma or squamous cell carcinoma were 0.921,0.952,0.920,0.947,0.944,0.956 and 0.960,respectively,and the AUC value of the three-modal model was significantly higher(P<0.05)than that of the single-modal models.The accuracy rates of the models to predict benign or malignant lesions based on CT features,18F-FDG PET features,99mTc-3PRGD2 SPECT features,CT and 18F-FDG PET features,CT and 99mTc-3PRGD2 SPECT features,18F-FDG PET and 99mTc-3PRGD2 SPECT features,or all features were 0.875,0.875,0.875,0.875,0.938,0.938,0.938,respectively,and the accuracy rates of the models to classify lung adenocarcinoma or squamous cell carcinoma were 0.889,0.889,0.778,0.889,0.889,0.889 and 0.889,respectively.The AUC value of the deep learning model to predict benign and malignant lesions based on 18F-FDG PET images was 0.886,while the AUC value of the conventional radiomics model was 0.975,without significant difference(P>0.05)between their ROC curves.The AUC value of the deep learning model to classify lung adenocarcinoma or squamous cell carcinoma was 0.780,while the AUC value of the conventional radiomics model was 0.867,without significant difference(P>0.05)between their ROC curves.Conclusion:Conventional radiomics models and deep learning-based radiomics models based on nuclear medical images have application value in the identification of benign or malignant lung space-occupying lesions and lung adenocarcinoma or squamous carcinoma. |
PDF Full Text Request