RT Journal Article SR Electronic T1 Impact of intra-tumoral, peri-tumoral, and annular-tumoral radiomic features on prognostic modeling in multi-center head and neck tumor PET/CT imaging of HECKTOR challenge 2021 JF Journal of Nuclear Medicine JO J Nucl Med FD Society of Nuclear Medicine SP 3111 OP 3111 VO 63 IS supplement 2 A1 Hui Xu A1 Lijun Lu A1 Mathieu Hatt YR 2022 UL http://jnm.snmjournals.org/content/63/supplement_2/3111.abstract AB 3111 Introduction: Head and Neck cancers (HNC) are among the most common cancers worldwide, and the early prediction of prognosis is essential for improving patient outcome. Compared to the radiomic features extracted from the primary tumor gross tumor volume (GTV), we aimed to evaluate the impact of intra-tumoral, peri-tumoral, and annular-tumoral radiomic features in predicting the progression-free survival (PFS) for HNC patients.Methods: From the training dataset of the MICCAI challenge HECKTOR in 2021, 224 HNC patients from 5 centers were analyzed with pre-treatment 18F-FDG PET/CT images. Based on the contour of GTV (ground-truth provided with the training set), 2D morphologic operations of dilation and erosion with radial distance of 2, 4, 6, 8 and 10 mm were performed slice-by-slice, to generate 5 peri-tumoral regions and 5 intra-tumoral, respectively. Additionally, 5 corresponding hollow annular regions were determined by subtracting intra-tumoral region from peri-tumoral region (Fig.1). From the GTV and each of the 15 alternative volumes of interest (VOIs), 346 IBSI-compliant radiomic features were extracted from both PET and CT images. Following the screening for significant features by Lasso-Cox model with 10 times 5-fold repetitions, multi-variate Cox regression models were constructed for PFS prediction and the optimal one was chosen by Akaike information criteria. Three-fold cross-validation was adopted to split all patients into training and testing sets, and the results were evaluated by the mean concordance index (C-index).Results: The 5 peri-tumoral, intra-tumoral and annular-tumoral regions were denoted as D/E/R-2, -4, -6, -8, and -10mm, respectively (Fig. 1). Radiomics models for all generated volumes of interest, except for E-8mm, R-6mm and R-10mm, show similar or better prognostic value than those based on GTV. Radiomic features extracted from D-2mm, E-2mm and R-2mm led to the models with the best prognostic performance among the 5 radial distances. However, larger distances of dilation or erosion along the edge of GTV contour did not improve and even decreased their prognostic power (Table 1).Conclusions: Radiomics quantification based on intra-tumoral, peri-tumoral, and annular-tumoral regions with proper distance could provide slightly improved prognostic power compared to GTV for HNC patients. Our next step is to evaluate these models on the testing set of HECKTOR 2021.Acknowledgements: This work was supported by the National Natural Science Foundation of China under grants 81871437 and 12026601, and the China Scholarship Council under grant 202108440348. We thank the organizers of the MICCAI HECKTOR 2021 challenge for permission to use the training dataset.