ESTRO 38 Abstract book

S159 ESTRO 38

Department of Radiotherapy, Utrecht, The Netherlands ; 3 Università Cattolica del Sacro Cuore, Istituto di Radiologia, Roma, Italy ; 4 Fondazione Policlinico Universitario A.Gemelli IRCCS, U.O.C. Radioterapia Oncologica- Dipartimento di Diagnostica per immagini- Radioterapia Oncologica ed Ematologia, Roma, Italy Purpose or Objective Magnetic Resonance Imaging (MRI) is a complex and extreme versatile imaging technique. This may potentially jeopardize the generalizability of radiomics of images acquired with different field strengths and scan protocols. Aim of this study was to develop a generalized radiomics model for predicting pathological complete response (pCR) after neoadjuvant chemo-radiotherapy (CRT) in locally advanced rectal cancer (LARC) patients (pts) using pre-CRT T2-weighted (T2-w) images acquired at a 1.5 T and a 3 T scanner. Material and Methods In two institutions 195 patients were scanned with a T2- weighted imaging protocol. In institution A 142 patients were scanned on a 1.5 T MR scanner (GE Signa Exite, Little Chalfont, United Kingdom) whereas in institution B 59 pts were scanned on a 3-T MR-scanner (Philips Medical System, Eindhoven, The Netherlands). The heterogeneity between the two cohorts of pts was evaluated in terms of Wilcoxon Mann Whitney (WMW) and Pearson's χ 2 test. Gross Tumor Volumes (GTV) were delineated on the MR images and all the images were resampled to a fixed spatial planar resolution of 0.7x0.7 mm 2 before to perform features extraction. A total of 225 radiomic features belonging to four families (fractal, statistical, textural and morphological features) were extracted, applying two image filters (Laplacian of Gaussian (LOG) and Intensity Based (IB)). Features were standardized with Z-score normalization and an initial feature selection was carried out using WMW test: the most significant features at 1.5 T and 3 T were selected as main features. Several logistic regression models combining the main features with a third one selected by those resulting significant in both datasets were elaborated and evaluated in terms of Area Under Curve (AUC) of the Receive Operative Curve (ROC). Features that combined together maximised the AUC value and minimised the Aikake Information Criteria (AIC) were selected. A 10-fold cross validation was repeated 300 times to evaluate the model robustness. Results Table 1 reports the clinical characteristics: no statistical difference was observed between the two cohorts of patients demonstrating matched populations.

Three features were selected based on their performance: maximum fractal dimension with IB=0-50, energy and grey level non uniformity calculated on the run length matrix with IB=0-50. Figure 1 reports the ROC curves of the model: the AUC of the model tested on the whole dataset after cross-validation was 0.72, while 0.70 and 0.83 were obtained when only 1.5T and 3T patients were respectively considered.

Conclusion A MR radiomics prediction model for pCR after neoadjuvant therapy in locally advanced rectal was developed . The model showed good performance, even when data from patients scanned on 1.5T and 3.0T were merged. This shows that magnetic field intensity variability can be overcome by means of selecting appropriate images features. PV-0311 MRI-based tumour control probability model in particle therapy G. Buizza 1 , S. Molinelli 2 , E. D'Ippolito 3 , G. Fontana 4 , L. Anemoni 5 , L. Preda 6 , G. Baroni 1 , F. Valvo 3 , C. Paganelli 1 1 Politecnico di Milano, DEIB - Department of Electronics Information and Bioengineering, Milano, Italy ; 2 National Centre of Oncological Hadrontherapy CNAO, Medical Physics Unit, Pavia, Italy ; 3 National Centre of Oncological Hadrontherapy CNAO, Radiation Oncology Unit, Pavia, Italy; 4 National Centre of Oncological Hadrontherapy CNAO, Clinical Bioengineering Unit, Pavia, Italy; 5 National Centre of Oncological Hadrontherapy CNAO, Medical Radiology Technicians