ESTRO 38 Abstract book

S1035 ESTRO 38

Nijmegen, The Netherlands ; 2 Maastricht University Medical Center +, Radiation Oncology - Maastro Clinic, Maastricht, The Netherlands Purpose or Objective Every cancer patient treated with radiotherapy undergoes imaging. This is imaging intended primarily for staging and treatment planning, however, there is increasing interest to use this data to support clinical decision making based on imaging. To build a reliable model ‘big data’ are necessary, involving data from hospitals located in different locations and countries to generalize well across the patient population. Cancer prediction modeling based on image biomarkers is referred to as radiomics. Although radiomics have predictive power, generalizability is often poor in cross- institutional studies because the institutions have different scanners with various acquisition protocols. The question in this study is whether there is a way to correct for the influence of these image acquisition parameters on the radiomics? Material and Methods We used a Gammex 467 CT phantom to build a training (default configuration) and validation (home-made plugs) sets for radiomics. As an example for scanner variation, we varied the X-ray tube exposure (mAs) used in the scans. To extract radiomic features we used an open-source tool PyRadiomics. Linear regression with X-ray tube exposure as a predictor was used to predict the relationship between X-ray tube exposure and a radiomic feature target value (TRV). Spearman rank was used to evaluate the relationship monotony between a radiomic feature and the X-ray tube exposure. Mean variance ratio (MVR) – variance ratio before and after correction averaged for each delineation – was used to evaluate the correction model. Results In 88 out of 92 radiomic features we found a positive correction in both training and validation sets (MVR > 1). Spearman rank test showed to be a good metric of correctability. Conclusion There is a straightforward way to correct on redundant radiomics variance caused by X-ray tube exposure variation. Radiomics correction modeling may be used to pre-process local institutional data to reduce scanner induced noise in cross-institutional studies involving radiomics. EP-1904 3T CE-MRI (peri)tumoral radiomics for prediction of lymphovascular invasion in early breast cancer M. Avanzo 1 , L. Vinante 2 , G. Pirrone 1 , J. Stancanello 3 , A. Revelant 2 , A. De Paoli 2 , A. Drigo 1 , L. Barresi 1 , L. Balestrieri 4 , M. La Grassa 4 , M. Urbani 4 , N. De Pascalis 4 , S. Massarut 5 , M. Mileto 5 , G. Franchin 2 , G. Sartor 1 1 Centro di Riferimento Oncologico di Aviano CRO IRCCS, Medical Physics Department, Aviano, Italy ; 2 Centro di Riferimento Oncologico di Aviano CRO IRCCS, Department of Radiation Oncology, Aviano, Italy ; 3 Oncoradiomics SA, Oncoradiomics SA, Liege, Belgium ; 4 Centro di Riferimento Oncologico di Aviano CRO IRCCS,

Department of Radiology, Aviano, Italy ; 5 Centro di Riferimento Oncologico di Aviano CRO IRCCS, Breast Surgery Unit, Aviano, Italy Purpose or Objective The presence of lymphovascular invasion (LVI) in early stage breast cancer patients is related to worse outcome. For patients undergoing partial breast irradiation with intraoperative radiotherapy (IORT), if LVI is found in pathological examination, then the local treatment is completed with external beam whole breast irradiation. The purpose of this investigation was to predict presence of LVI in patients with early-stage invasive breast cancer by use of radiomics of tumor and peritumoral volume (PV) in 3.0-T, contrast enhanced MRI (CE-MRI). Material and Methods 50 patients diagnosed with early-stage invasive breast carcinoma had bilateral 3.0-T breast CE-MRI before surgery and IORT, which was delivered using low-energy photon source, the Intrabeam System (Carl Zeiss, Oberkochen, Germany). The gross target volume (GTV) was contoured by an experienced radiation oncologist in the T1–MRI series with maximum contrast enhancement. The PV was automatically contoured by generating a 1 cm thickness shell around the GTV. The 3D images were pre-processed with re-sampling and 3-D filtering using Gaussian, Laplacian of Gaussian, and Median filters. A total of 228 radiomic histogram-based and textural features were calculated in the GTV and PV (456 in total). Sequential feature selection was used to identify a subset of features that best predicts the data and remove redundant or not significant predictors. A support vector machine machine learning classifier was trained on the patient dataset for prediction of presence of LVI in the treated site assessed by pathology (positive classifier for presence of LVI). The predictive power of the model was assessed using sensitivity (probability that test is positive on patients with LVI), specificity, and Youden 's index in five-fold cross validation. Results In 20 out of 50 patients (40.0%) LV was found in pathology. The features selected for prediction of LV were 8 from the PV, 3 from GTV, all of which were textural features derived from the GLRLM, GLCM, NGTDM, and GLSZM matrices. The classifier scored sensitivity 90.0%, specificity 80.0%, and Youdenìs Index of 0.7 in the cross These preliminary findings show that radiomic variables extracted from the PV and GTV in 3.0-T CE-MRI can predict LVI and may help to better select patients candidate to exclusive partial breast irradiation with IORT. EP-1905 CT /PET based dosiomics and radiomics model predicts local control of nasopharyngeal carcinoma M. Avanzo 1 , G. Pirrone 1 , C. Avigo 2 , G. Fanetti 3 , J. Stancanello 4 , A. De Paoli 3 , P. Elisa 3 , A. Drigo 1 , P. Chiovati 1 , A. Dassie 1 , E. Borsatti 5 , T. Baresic 5 , G. Franchin 3 , G. Sartor 1 1 Centro di Riferimento Oncologico di Aviano CRO IRCCS, Medical Physics Department, Aviano, Italy ; 2 ULSS 1 Dolomiti- S. Martino Hospital, Medical Physics Department, Belluno, Italy ; 3 Centro di Riferimento Oncologico di Aviano CRO IRCCS, Department of Radiation Oncology, Aviano, Italy ; 4 Oncoradiomics SA, validation. Conclusion

Oncoradiomics SA, Liege, Belgium ; 5 Centro di Riferimento Oncologico di Aviano CRO IRCCS, Department of Nuclear Medicine, Aviano, Italy

Purpose or Objective to develop and validate a model predictive for local control, based on CT-PET radiomics and planning CT