ESTRO 2022 - Abstract Book

S1564

Abstract book

ESTRO 2022

Conclusion Our results suggest that D-TA has a significant role in the prediction of pCR, thus this method may lead to select patients who may potentially avoid surgery. However, further analysis with prospective and multicenter trials is warranted.

PO-1760 Independent validation of a PET radiomic model predicting outcome after Radiotherapy for HN cancer

M. Mori 1 , C. Deantoni 2 , M. Olivieri 1 , E. Spezi 3 , A. Chiara 2 , S. Baroni 2 , M. Picchio 4 , A. Del Vecchio 1 , N. Di Muzio 2 , C. Fiorino 1 , I. Dell'Oca 2 1 IRCCS San Raffaele Scientific Institute, Medical Physics, Milan, Italy; 2 IRCCS San Raffaele Scientific Institute, Radiotherapy, Milan, Italy; 3 Velindre Cancer Centre, Medical Physics, Cardiff, United Kingdom; 4 IRCCS San Raffaele Scientific Institute, Nuclear Medicine, Milan, Italy Purpose or Objective To externally validate 18 Fluorodeoxyglucose Positron Emission Tomography ( 18 F-FDG-PET ) radiomic scores recently suggested by Martens et al. (EJNMI Research (2020) 10:102) for the prediction of local recurrence (LRFS), distant metastasis (DRFS) and overall survival (OS) in patients with oropharyngeal cancer treated with chemo-radiotherapy. Materials and Methods Outcome data and pre-radiotherapy PET images of 78 patients (G1/G2=3; G3=13; G4=62) treated at our Institute between 2006 and 2020 according to an internal protocol delivering moderate hypo-fractionation (66 Gy in 30 fr) were available. Image processing (voxelization and binning) was performed according to the procedures described by Martens et al. and tumors were segmented using a previously validated semi-automatic method with the software MIM (v.6.9.6). Four-hundred and fifty radiomic features (RF) were extracted according to IBSI (Image Biomarker Standardization Initiative) guidelines, using the SPAARC software (Whybra et al. Sci Rep 9:9649(2019)). Martens et al. condensed the predictive RF in 8 independent meta-factors (F), consisting of a combination of selected RF with variable importance weight. In particular, two F named F1 (combining least-axis-length, non-uniformity, high-dependence-of-high-grey-levels) and F5 (asphericity, major-axis-length, inverse-compactness and inversed-flatness) were suggested as highly predictive of OS. Then, F1 and F5 were tested in predicting OS on our population by Cox Regression. An index combining both F was also computed and used to stratify risk according to the best cut-off derived from the ROC analysis. Kaplan Meier test was finally performed. Results Among 78 patients, 3, 11 and 22 events for LRFS, DRFS and OS were registered in our external population. In Martens et al, RF selected as predictive were found only for LRFS and OS prediction. Due to the small number of LRFS events in our population, only OS prediction F may be tested. F1 and F5 derived from our population were confirmed to be associated to OS (p=0.028 and 0.080 respectively). The prognostic index derived from Multivariate Cox analysis combining F1 and F5 was tested with ROC analysis, showing AUC=0.68 (95%CL=0.56-0.78) and p=0.006. According to the Youden criterion the best cut-off for the combined index was >0.48, showing a high ability in patients stratification, as depicted by the Kaplan Meier curve reported in the Figure below (p=0.001, HR=4.1, 95%CI=1.78-9.54). The 5-year OS were 91% (95%CI: 82-100%) vs 63% (48-80%) for low and high-risk groups respectively.

Conclusion Previously published radiomic scores were independently confirmed in predicting OS, despite different Institute/scanners, delivered dose and patient’s characteristics. Multiple validations like the current one may help in corroborating the generalizability of radiomic-based models, allowing for personalized risk stratification and optimization of personalized cancer care.