PT - JOURNAL ARTICLE AU - Haddadi Avval, Atlas AU - Hajianfar, Ghasem AU - Amini, Mehdi AU - Salimi, Yazdan AU - Oveisi, Mehrdad AU - Shiri, Isaac AU - Zaidi, Habib TI - <strong>Progression-free survival prediction in head and neck cancer: A comparative study between conventional PET indices and radiomics models. </strong> DP - 2022 Aug 01 TA - Journal of Nuclear Medicine PG - 3173--3173 VI - 63 IP - supplement 2 4099 - http://jnm.snmjournals.org/content/63/supplement_2/3173.short 4100 - http://jnm.snmjournals.org/content/63/supplement_2/3173.full SO - J Nucl Med2022 Aug 01; 63 AB - 3173 Introduction: Radiomics is an emerging era of quantitative imaging which aids in the management and prediction of patients’ prognoses. In this study, we aimed to predict head and neck cancer progression-free survival (PFS) using two methods; radiomics features and conventional PET quantitative indices.Methods: Our study consisted of the data of 322 patients with head and neck cancer collected from six different centers. From the PET images of each patient, four conventional quantitative indices namely SUVmax, SUVmean, Metabolic tumor volume (MTV), and total lesion glycolysis (TLG) in addition to 107 radiomics features were extracted. Four models were constructed using each of the conventional indices. In addition, A comprehensive model was established using radiomics features. The latter utilized univariate C-index as the feature selection method and Cox proportional hazard (Coxph) as the machine learning algorithm. Finally, the performance of each of the five models was evaluated using C-index. The performance of the four models was compared with Our comprehensive radiomics model via student t-test analysis (p-value&lt;0.05 was considered significant). Leave-one-center-out cross-validation was used in this study. Therefore, models were trained on the data of five centers and then tested on one center in each iteration (bootstrapping = 1000).Results: The most repeatedly selected features for our comprehensive model were sphericity, max 2D diameter, and flatness from the shape set, minimum from the first-order set, small dependant low gray-level emphasis from the GLDM set, and small area low gray-level emphasis from the GLSZM set. Our SUVmax, SUVmean, MTV, and TLG models performed well with C-indices (mean±SD) of 0.63±0.12, 0.62±0.12, 0.64±0.09, and 0.64±0.10, respectively. Our comprehensive model reached a C-index of 0.67±0.08 which was significantly better than all of our previous models (all p-values&lt;0.001).Conclusions: Our findings show that radiomics applications are significantly superior to conventional quantification of PET images for the purpose of head and neck cancer patients’ prognostication. Further studies might be conducted to prove the potential of radiomics models to be applied to the clinical setting.