ESTRO 38 Abstract book

S1054 ESTRO 38

EP-1936 PET/CT Radiomics predict local recurrence in patients treated with SBRT for early-stage NSCLC G. Dissaux 1 , M. Hatt 2 , D. Visvikis 2 , O. Pradier 1 , E. Chajon 3 , I. Barillot 4 , L. Duverge 3 , I. Masson 5 , R. Abgral 6 , M. Santiago-Ribeiro 7 , A. Devillers 8 , F. Kraeber-Bodéré 9 , M. Mahé 5 , R. De Crevoisier 3 , U. Schick 1 1 University Hospital, Radiation Oncology department, Brest, France ; 2 Univ Brest, LaTIM- INSERM- UMR 1101, Brest, France ; 3 Centre Eugène Marquis, Radiation Oncology department, Rennes, France ; 4 University Hospital, Radiation Oncology department CORAD, Tours, France ; 5 Institut de Cancérologie de l'Ouest Cancer Center, Radiation Oncology department, Nantes, France ; 6 University Hospital, Nuclear Medicine department, Brest, France ; 7 University Hospital, Nuclear Medicine department, Tours, France ; 8 Centre Eugene Marquis, Nuclear Medicine department, Rennes, France ; 9 Institut de Cancérologie de l'Ouest Cancer Center, Nuclear Medicine department, Nantes, France Purpose or Objective The aim of this French multicentric study was to develop and validate an FDG PET/CT radiomics signature with prognostic value in patients treated with stereotactic radiotherapy (SBRT) for early-stage non-small cell lung cancer (NSCLC) and to assess its incremental value with respect to the standard clinical and imaging features. Material and Methods Patients from Rennes (n=27), Tours (n=29) and Brest (n=8) were pooled to constitute the training set, whereas the patients from Nantes (n=23) were used as the testing set. The primary endpoint of study was local control. For each patient, the primary tumour was delineated automatically in the PET images using the FLAB algorithm, and manually in the low-dose CT images, and IBSI-validated radiomic features were extracted. A total of 184 radiomic features (92 in both FDG PET and low-dose CT), 7 clinical and histopathological parameters (age, gender, tumour size, stage, localization, OMS status, histology) and one treatment parameter (biologically equivalent dose, BED) were included. In order to pool together radiomics features extracted from the four institutions relying on different PET/CT scanners and associated protocols, we used the posteriori harmonization method ComBat. In the training set, variables found significant in the univariate analysis were fed into a multivariate Cox proportional hazard regression model. The area under the ROC curve (AUC) was used to evaluate the performance of the resulting model in the testing set. Results Median follow-up was 22.7 (3.2 – 63.4) and 22.2 (1.7 – 58.1) months, in training and testing sets respectively. In univariate analysis, none of the clinical variables, 2 PET features and 3 CT features were significant. The best performance in the training set was obtained with the model combining the two PET features, reaching an AUC of 0.94 (sensitivity 100%, specificity 88%) to predict local recurrence, with a HR undefined (p<0.001). This model obtained an accuracy of 0.91 (sensitivity 100%, specificity 81%), with a HR undefined (p = 0.023) in the testing set. The models relying on CT radiomics features or the combination of PET and CT features reached lower accuracy. Conclusion We showed that two radiomic features derived from FDG PET were independently associated with local control in patients with NSCLC undergoing SBRT and could be combined in an accurate predictive model. This model could provide recurrence-related information and could be helpful in clinical decision-making, especially regarding dose escalation.

1 Fondazione Policlinico A. Gemelli IRCCS- Università Cattolica Sacro Cuore, Dipartimento Scienze Radiologiche- Radioterapiche ed Ematologiche- Istituto di Radiologia, Rome, Italy ; 2 Università di Perugia- Ospedale S. Maria della Misericordia- Sezione di Radioterapia Oncologica, Dipartimento di Chirurgia e Scienze Biomedicali, Perugia, Italy ; 3 Fondazione Policlinico A. Gemelli IRCCS, Dipartimento Scienze Radiologiche- Radioterapiche ed Ematologiche- Istituto di Radiologia, Rome, Italy Purpose or Objective Different investigations have focused on the GBM’s heterogeneous features to develop an individualized patient management. A multi-institutional study, the GLI.F.A. (Glioblastoma: advanced Imaging Features Analysis) Project, was performed for a comprehensive analysis of GBM heterogeneity in order to create a multidimensional map for predictive models (PM) and decision support systems (DSS) in GBM Material and Methods Adult patient with newly diagnosed GBM, that undergo to surgery and chemo-radiotherapy according to EORTC 26981-22981-NCIC trial were analyzed in this first phase of the study. Gross Tumor Volume (GTV) was contoured in the T1 post contrast and T2-FLAIR weighted images. A brain ontology and a platform for sharing and combining multiple datasets (BOA-WEB System) were created in order to standardize data. MRI features were extracted by the dedicated software. Two analysis were conducted: the study of imaging features of MRI at diagnosis and a delta radiomics study oriented to evaluate the evolution of imaging features, considering all MRI performed. In both of the studies, the Wilcoxon Mann Whitney test and the Log-rank test for Kaplan-Meier curves were applied to evaluate the significance of the radiomic features on the T2-Flair and T1 images, using the median value of the radiomic features to categorize the continue variables. We considered as main outcomes the local control (LC), the progression fee survival (PFS) and the response to radio-chemotherapy (RTCT). Results We enrolled in this study 43 patients, treated from July 2014 to February 2018. Median age was 63 years (45-80) and 27 patients were still alive at the time of the analysis. The MODDICOM software analyzed image features. Significant features of MRI images at diagnosis divided by MRI sequence and outcome are reported in Fig1. The second order, textural features order, describing the spatial correlation between images voxel, resulted significant for PFS, LC and response to RTCT on the contrast-enhanced T1 weighted MRI. Regarding the delta radiomics features analysis, 18 patients were considered and the CTV segmentation and feature extraction was performed on 68 post contrast T1 weighted MRI. After extracting 92 features, we observed a significa nt correlation, defined by the adjusted p value, for PFS and LC on pre- RT MRI and on MRI after the 6th chemotherapy cycle. The significant features belong to the first, the second and the third order are reported in Fig2. Almost significant features are morfological and could describe quantitatively something that usually is only qualitatively described Conclusion This preliminary univariate analysis suggests that the radiomic features relates to survival and clinical outcomes and that is possible to stratify patients according MR based quantitative imaging. A higher number of patients, multivariate analysis and external validation are next steps for getting reliable predictive model.