Peri-tumoral Ratliorics Analysis On PET/CT Images:application For Improved Prognosis Of Head And Neck Cancer

The mortality rate of head and neck(H&N)cancer ranks 7th in the world for malignant tumors.Local recurrence(LR)and distant metastasis(DM)are the main causes of treatment failure and death.18F-FDG PET/CT has been used as an important tool for diagnosis,prognosis,and staging in head and neck cancer(H&N cancer).Therefore,pretreatment prediction of LR and DM to achieve accurate risk stratification is clinically invaluable for developing a personalized cancer treatment regimen.In recent years,PET/CT radiomics has been used to predict the prognosis of patients with H&N cancer.However,most of these studies focused on quanting the region of whole tumor,without considering the role of peritumoral region in tumor growth,proliferation,invasion and metastasis.We designed an experiment to study the value of quantifing the peritumoral region in prediction of LR and DM of H&N cancer based on a suppose that there are prognostic factors in peritumoral region of H&N cancer’s PET/CT imagesData obtained from The Cancer Imaging Archive(TCIA)that include 297 H&N cancer patients who had been treated with radiation or chemoradiotherapy from four different centers were retrospectively analyzed.For the aim of prediction of LR and DM of H&N cancer,6294 3D radiomic features were extracted from the intratumoral region and three groups of peritumoral region respectively for both PET and CT components of PET/CT images.Eight ensemble logistic regression models trained with the optimal combination of features which chosen by using a stepwise forward feature selection in terms of the Gain equation and the model’s performance were used to predict the outcome(LR or DM)of patients and stratify the patients into different risk groups.The models based on different modalities(PET and CT)and different regions(intra-tumor and peri-tumor)were fusion at the score-level by average strategy to evaluate the complementary value of multi-modalities and multi-regions.The model’s performance was evaluated by area under the receiver operating characteristic curve(AUC),sensitivity,specifity,concordance index(C-index)and Kaplan-Merier curve’s analysisFor the prediction and prognosis of LR,the combined model PETCT-Peri-5 which based on peritumoral regions achieved a higher performance significantly compared with those models CT-Intra or PET-Intra based on intratumor alone(PETCT-Peri-5 vs.PET-Intra/CT-Intra:AUC:0.76 vs.0.45/0.60,p<0.0001/p=0.02,Delong’s test;C-index:0.73 vs.0.45/0,60,p=0.001/p = 0.02,Noether’s test).Model PETCT-Peri-5 yielded the optimal risk stratification with the p-value of log rank test of 0.00015.For the prediction and prognosis of DM,the value of peritumoral radiomic models is limited.The combined model PETCT-Intra&Peri-5 and PETCT-Intra achieved the best performance of prediction and prognosis respectively(PETCT-Intra&Peri-5 vs.PETCT-Intra:AUC:0.90 vs 0.88,p=0.62,Delong’s test;C-index:0.88 vs 0.90,p=0.17,Noether’s test).In both PET and CT modalities,as well as in predicting DM and LR,peritumoral distance of 5mm showed the best predictive performance among the three groups of peritumoral regions.This study suggests that quantification the tumor and peritumoral microenvironment of PET/CT images can help distinguish whether patients with H&N cancer will develop to LR and DM,and help to risk stratification of patients before treatment.However,the complementary value of peritumoral radiomic features to intratumoral features in DM prediction and prognosis is not obvious.