ESTRO 2020 Abstract book

S905 ESTRO 2020

Cavalieri 8 , S. Alfieri 8 , L. Licitra 8 , E. Pignoli 6 , L. Mainardi 9 , C. Fallai 1 , M. Bologna 7 1 Fondazione IRCCS Istituto Nazionale dei Tumori, Radiation Oncology 2, Milan, Italy ; 2 Fondazione IRCCS Istituto Nazionale dei Tumori, Department of Radiology, Milan, Italy ; 3 University of Milan, Department of Oncology and Haemato-oncology, Milan, Italy ; 4 Fondazione IRCCS Istituto Nazionale dei Tumori, Radiation Oncology 1, Milan, Italy ; 5 Fondazione IRCCS Istituto Nazionale dei Tumori, Prostate Cancer Program, Milan, Italy ; 6 Fondazione IRCCS Istituto Nazionale dei Tumori, Medical Physics, Milan, Italy ; 7 Politecnico di Milano, Dipartimento di Elettronica- Informazione e Bioingegneria DEIB, Milan, Italy ; 8 Fondazione IRCCS Istituto Nazionale dei Tumori, Head and Neck Medical Oncology Unit, Milan, Italy ; 9 Politecnico di Milano, Dipartimento di Elettronica- Informazione e Bioingegneria DEIB9, Milan, Italy Purpose or Objective To assess the ability of magnetic resonance imaging (MRI)- radiomics to provide a prognostic signature for overall survival (OS) in locoregionally advanced nasopharyngeal cancer (LANPC) patients (pts) in a non endemic area Material and Methods A mono-institutional series of 179 Epstein Barr Virus (EBV)– related non metastatic LANPC pts curatively treated with Intensity Modulated radiotherapy and chemotherapy (CHT) with or without induction CHT between 2004 and 2017 was considered for this retrospective analysis. The primary endpoint of the study was OS. All pts had pretreatment multi-MRI images. Texture features were extracted from the pretreatment T1- and T2-weighted images for each case. Regions Of Interest (ROIs) were considered in primary tumor and pathological nodes. A total of 2144 radiomics features was extracted from the ROIs by using software Pyradiomics (2.1.0). First, a set of stable, non redundant (r<0.85) and prognostic features (p<0.05 in univariate Cox models) was selected. The remaining radiomics features were sorted by hazard ratio and progressively added to a multivariate Cox regression to estimate the best features number for the final signature. A bootstrap sampling was used to select the pts of the training set (n=110), and the unsampled pts were used as a validation set (n=69). The association between the radiomics signatures and OS was investigated. The proportion of events in the training and validation sets was the same (12%). Feature selection and Cox model fitting were performed on the training set. Pts in the validation set were classified according to radiomics signature into high risk (HR) and low risk group (LR). The performance of the signature was evaluated in the validation set using the Harrel’s c-index (CI). Univariate analysis comparing Kaplan-Meier curves was performed using log-rank test. The possibility of a correlation with the primary gross tumor volume (GTV-T) was checked for the resulting radiomics signature. Results The median follow-up was 60.43 months (range: 44.23– 67.00). A signature including 2 features accounting for tumor intensity and texture ( T_T1w_waveletLLH_firstorder_Median and T_T2w_waveletLLH_glcm_Imc1 ) was identified. 5-years OS was 78.6% and 94.6% for HR and LR group, respectively. In the validation cohort the signature had a CI of 0.67±0.09 and a significant difference was found in the OS between HR and LR groups (p=0.04 for log-rank test, see Fig.1). A correlation was found between GTV-T and the radiomics signature (r=0.6), but GTV-T alone had no significant impact on OS curves (CI 0.57±0.1, log-rank p=0.28).

with 36.25 Gy delivered to the whole prostate and a concomitant boost of 37.5 Gy to the dominant intraprostatic lesion (DIL) identified by multiparametric MRI. T2-weighted (T2W) MRI sequences acquired with homogenous characteristics (0.59x0.59x3 mm 3 voxel size, XX Te, YY Tr) on a 1.5T Magnetom Avanto fit scanner (Siemens) were selected and the prostate gland contours were analyzed. The extraction of radiomic features (shape, first-order statistics and textural features) was performed using the IBEX software. We tested univariate and multivariate association of each radiomic features with T-stage (cT1 vs cT2), Gleason score (GS, 3+3 vs 3+4/4+3), extracapsular extension (ECE, 1/2 vs 3/4) score, Prostate Imaging – Reporting and Data System (PIRADS, 2/3 vs 4/5) score and risk class (intermediate vs low), and selected the feature with the lowest p-value in each cluster as representative. Statistical analysis was performed with SAS/STAT® software. Results Of the 65 prospectively enrolled patients, 49 T2W-MRI sequences fulfilled the inclusion criteria. Among the 1702 features extracted, 3 to 6 features with the highest predictive power were selected for each outcome. A logistic regression (machine learning) classifier was trained to predict clinical outcomes. Radiomic signature for prediction of high Gleason score included only GLCM3 texture features. Radiomic signature for prediction of cT2 stage as well as for 3/4 ECE score included first-order statistics intensity features. A GLCM3 texture feature was the most predictive feature for 4/5 PIRADAS score, with excellent predictive accuracy. Finally, radiomic signature for prediction of intermediate risk class included both GLCM3 texture and first-order statistics intensity features, with good predictive accuracy. Overall, the multivariable radiomic signature predicted oncological and radiological scores with AUC ranging from 0.74 to 0.94 (Table 1).

Conclusion MRI-based radiomics in PCa for the prediction of tumour phenotype is a feasible and promising approach. It might lead to a semi-automated definition of tumour characteristics and thus reduce the intra/inter-operator variability in the radiologic image interpretation. Although a significant association was found between the selected features and all the mentioned clinical and radiological scores, further validations on larger cohorts are needed before applying these findings in the clinical practice. PO-1577 Baseline MRI-radiomics can predict overall survival in non endemic nasopharyngeal cancer patients E. Orlandi 1 , G. Calareso 2 , C. Tenconi 3,4 , T. Rancati 5 , N.A. Iacovelli 1 , A. Cavallo 6 , N. Facchinetti 1 , R. Ingargiola 1 , E. Ivaldi 1 , D.A. Romanello 1 , V. Corino 7 , R. Valdagni 3,4,5 , S.