| Objective N To establish machine learning models based on multimodal PET/MR radiomics for predicting early response to concurrent chemoradiotherapy(CCRT)in locally advanced cervical cancer.Methods NA total of 154 patients with locally advanced cervical squamous cell carcinoma were enrolled and PET/MR scans were performed before CCRT.109 patients scanned by GE Signa PET/MR machine were randomly divided into training dataset and validation dataset at 7:3 ratio.The Philips Ingenuity TF PET/MR machine scanned 45 patients as an external dataset.Patients with residual tumor>7.5cm~3 after external irradiation were evaluated as non-good early response(NGER),and tumor residual volume≤7.5cm~3 was considered as good early response(GER).The volumes of interest of primary tumor in PET,MR-T2 and MR-T1 sequences were manually segmented,and radiomics features were extracted using AK software.Maximal correlation minimum redundancy(m RMR)and Lasso algorithms were used to select and calculate the PET Rad-score(Rad-Score＿PET),MR Rad-score(Rad-Score＿MR)and PET/MR Rad-score(Rad-Score＿PET/MR).Logistic regression algorithm was used to build models of PET model,MR model,PET/MR model and combined model that was PET/MR radiomics combined with clinicopathological in the training dataset,and the models were tested in the validation and external dataset.Kruskal Wallis rank sum test and Pearson’s Chi-squared test were used to test the difference of clinicopathological features distribution between the training and validation set.The Wilcoxon rank-sum test was used to analyze the distribution differences of Rad-Score＿PET,Rad-Score＿MR and Rad-Score＿PET/MR between NGER and GER patinets in the three datasets.ROC curve and decision curve were used to evaluate the prediction efficiency of the model.Pairwise comparison of ROC curves was performed by the Delong test.Results NThe results of rank sum test and chi-square test showed that there was no significant statistical difference of clinicopathological features between the training and validation set(P>0.05).And Rad-score＿PET/MR of NGER patients were higher than GER patients in the training and external dataset(P<0.001,p=0.004,respectively).But the difference was not statistically significant in validation dataset(P>0.05).The AUC values of PET/MR model in the training dataset(AUC=0.906,P<0.001)and external dataset(AUC=0.880,P<0.001)were all higher than that of PET model(AUCs:0.793,0.691,P<0.001)and MR(AUCs:0.901,0.776,P<0.001).The decision curve showed that the combined model had the highest and stable net benefit in all datasets.Conclusion NThe PET/MR multimodal radiomics machine learning model can predict the early response of CCRT in locally advanced cervical squamous cell carcinoma,which provides a new tool for clinical adjustment of radiotherapy dose,improvement of radiotherapy sensitization drugs and combined adjuvant therapy.Moreover,the model combined with clinicopathological information has the higher robustness.The Rad-Score＿PET/MR of the patients with NGER of concurrent chemoradiotherapy was significantly higher than that of the patients with NGER. |
