PT - JOURNAL ARTICLE AU - Gurvan Dissaux AU - Dimitris Visvikis AU - Ronrick Do-ano AU - Olivier Pradier AU - Enrique Chajon AU - Isabelle Barillot AU - Loig Duverge AU - Ingrid Masson AU - Ronan Abgral AU - Maria-Joao Santiago Ribeiro AU - Anne Devillers AU - Amandine Pallardy AU - Vincent Fleury AU - Marc-André Mahé AU - Renaud De Crevoisier AU - Mathieu Hatt AU - Ulrike Schick TI - Pre-treatment <sup>18</sup>F-FDG PET/CT Radiomics predict local recurrence in patients treated with stereotactic radiotherapy for early-stage non-small cell lung cancer: a multicentric study AID - 10.2967/jnumed.119.228106 DP - 2019 Nov 01 TA - Journal of Nuclear Medicine PG - jnumed.119.228106 4099 - http://jnm.snmjournals.org/content/early/2019/11/15/jnumed.119.228106.short 4100 - http://jnm.snmjournals.org/content/early/2019/11/15/jnumed.119.228106.full AB - Purpose: The aim of this retrospective multicentric study was to develop and evaluate a prognostic FDG PET/CT radiomics signature in early-stage non-small cell lung cancer (NSCLC) patients treated with stereotactic radiotherapy (SBRT). Material and Methods: Patients from 3 different centers (n = 27, 29 and 8) were pooled to constitute the training set, whereas the patients from a fourth center (n = 23) were used as the testing set. The primary endpoint was local control (LC). The primary tumour was semi-automatically delineated in the PET images using the Fuzzy locally adaptive Bayesian algorithm, and manually in the low-dose CT images. A total of 184 IBSI-compliant radiomic features were extracted. Seven clinical and treatment parameters were included. We used ComBat to harmonize radiomic features extracted from the four institutions relying on different PET/CT scanners. In the training set, variables found significant in the univariate analysis were fed into a multivariate regression model and models were built by combining independent prognostic factors. Results: Median follow-up was 21.1 (1.7 – 63.4) and 25.5 (7.7 – 57.8) months in training and testing sets respectively. In univariate analysis, none of the clinical variables, 2 PET and 2 CT features were significantly predictive of LC. The best predictive models in the training set were obtained by combining one feature from PET, namely information correlation 2 (IC2) and one from CT (Flatness), reaching a sensitivity of 100% and a specificity of 96%. Another model combining 2 PET features (IC2 and Strength), reached sensitivity of 100% and specificity of 88%, both with an undefined hazard ratio (HR) (p&lt;0.001). The latter model obtained an accuracy of 0.91 (sensitivity 100%, specificity 81%), with a HR undefined (P = 0.023) in the testing set, however other models relying on CT radiomics features only or the combination of PET and CT features failed to validate in the testing set. Conclusion: We showed that two radiomic features derived from FDG PET were independently associated with LC in patients with NSCLC undergoing SBRT and could be combined in an accurate predictive model. This model could provide local relapse-related information and could be helpful in clinical decision-making.