ESTRO 38 Abstract book

S1028 ESTRO 38

both the maximum dose to the PTV as well as the conformity of the prescribed dose to the PTV.

Picture 1: Dose comparison before hotspot correction (right) and after (left)

Results For GTVT, four radiomics features met all criteria. For GTVN, none of the radiomics had an AUC>0.65 at each week, so that we selected the two features with the highest and significant AUC. The number of patients having an absolute delta features variation out of the boundaries was the highest at the third week, thus, this time point was chosen to drive our models. For GTVT, AUCs for predicting the therapeutic response were 0.782 (p<0.001), 0.701 (p = 0.0059) and 0.740 (p<0.001) for clinical, CBCT3-radiomics (Run Length Non-Uniformity (RLNU) noramlized)(Fig.2) and combined models, respectively. For GTVN, the clinical based model (AUC 0.736, p<0.001) did better than the combined model (AUC 0.563, p=0.143) or the radiomics based model (Elongation, AUC : 0.569). On the validation cohort, the best prediction was given by radiomics for GTVT (AUC:0.639) and by clinical parameters for GTVN (AUC:0.685).

Electronic Poster: Physics track: Radiobiological and predictive modelling, and radiomics

EP-1892 Predictive Response To Radiotherapy Of Head And Neck Cancer Using Radiomics Analyses Of Cbct. S. Sellami 1 , D. Bouzid 2 , M. Hatt 2 , D. Visvikis 2 , F. Lucia 1 , O. Pradier 1 , U. Schick 1 1 University Hospital- Brest Morvan, Radiation Oncology Department, Brest, France ; 2 LaTIM- University of Brest, Imaging Research Laboratory- INSERM- UMR 1101, Brest, France Purpose or Objective Radiotherapy for head and neck cancer (HNC) is now guided by cone-beam computed tomography (CBCT). Advanced imaging features (radiomics) extracted from diagnostic imaging have already been shown to predict outcome in several tumor models. The aims of this study were to investigate a methodology for feature selection of a longitudinal radiomics approach and to develop a radiomics signature based on CBCT to predict response to radiotherapy. Material and Methods In 102 HNC patients grouped in a training (=68) and validation (=34) cohorts, Gross Tumor Volumes of the primary tumor (GTVT) and metastatic lymph nodes (GTVN) used for the planning were registered to the weekly CBCT images using a deformable registration followed by manual correction. 88 features were extracted from the GTVs on each CBCT. Receiver operating characteristic (ROC) curves were plotted at each week to evaluate the predictive power of response of each feature. Only significant features at each week and independent of volume were pre-selected (AUC) >0.65). Absolute differences (CBCTn-CBCT1, called delta-radiomics) were calculated between features from each weekly CBCT images and the baseline CBCT1 performed before the first fraction. The smallest detectable change (C) with his confidence interval (95%) was determined for each radiomics using the formula C=1.96*SD, SD being the standard deviation of différences between features values calculated on CBCT1 and CBCT2. We then selected the features for which the change was more than C for at least 10% of patients at least for one week (Fig.1). A radiomics-based model was built at the time-point that showed most changes. Finally, we compare the prognostic performance of 3 models: clinical, radiomics, and combined.

Conclusion We described a feature selection methodology for longitudinal radiomics that is able to select reproducible delta radiomics features which are informative due to their change during treatment. Nonetheless, the prognostic value of the selected delta radiomics features did not seem to improve the prediction already given by the clinical data. EP-1893 A machine learning based stain-free method for classification of cell apoptosis stages J. Feng 1 , P. Wang 2 , N. Zhang 1 , S. Yu 1 1 Tianjin University, Biomedical Engineering, Tianjin, China ; 2 Tianjin Medical University Cancer Institute and Hospital, Radiation Therapy, Tianjin, China