ESTRO 2020 Abstract book

S900 ESTRO 2020

compared to the volume of the image at the end of treatment. Results The models performed well in 225 of the patients, for 15 patients were the total prediction ‘error’ was double the average RMSE (Figure 1a). For all patients, the linear, second- and third- degree regression models achieved an average RMSE of 22, 10 and 3 respectively. As expected, the higher order polynomial regression models performed better. However, the linear regression model performed acceptably. Figure 1b) shows the predicted tumour volume using a linear fit against the measured tumour volume across all patients. The linear model predicted an average tumour volume of 72 cubic centimetres (cm 3 ) compared to the measured tumour volume of 79 cubic cm 3 . The delta range between measured and predicted tumour volume is -37.34 – 40.15 cm 3 . The visual representation of the predictions applying a linear fit for an example patient can be seen in Figure 2. Conclusion Our results show that regression models built per-voxel on the intensity values from on-treatment CBCT can predict tumour shape and volume at the end of treatment. If we can identify tumours responding to treatment early, patients that will benefit from plan adaption can be identified early. Future work will explore clustering patients to identify patients with similar changes and linking tumour change with survival information. PO-1570 Robustness of dosomic features extraction on grid resolution and algorithm model calculation L. Placidi 1 , J. Lenkowicz 1 , D. Cusumano 1 , N. Dinapoli 1 , R. Gatta 1 , V. Valentini 1 1 Fondazione Policlinico Universitario A. Gemelli IRCCS - Università Cattolica Sacro Cuore, KBO Labs, Rome, Italy Purpose or Objective Patients’ 3D dose distributions can be considered as images with spatial and statistical distributions of dose levels that can be investigated with texture analysis (TA). Dosomics, can be considered as a method to quantify the heterogeneity of regions of interest (ROIs) and produce consistently defined, shape-based dose features, that encode the spatial distribution of dose at a higher resolution than organ-level dose-volume histograms. A relevant step to be investigated in advance before introducing dosomics to improve predictive modelling, is to evaluate the robustness and repeatability of the extracted dosomics features. In this study, on calculation grid resolution and algorithm calculation models. Material and Methods 144 different 3D dose distributions of 18 patients were analyzed, each characterized by a specific delivery techniques and/or by a different irradiated anatomical site, as reported in Fig.1. For each of these 18 patients, 8 different 3D dose distributions were calculated, based on all possible combinations of the calculation grid resolution (1, 2, 2.5 and 3 mm) and on the algorithm calculation models (Analytical Anisotropic Algorithm and Acuros XB). Dosomic features extraction has been performed by a routine in MODDICOM library, R language, optimized for the dose distribution texture analysis. A total number of 56 IBSI standardized features have been extracted, belonging to the statistical, morphological and gray level co-occurrence (glcm) textural families from 4 ROIs: planning target volume (PTV), ring (defined around the PTV – 0.5 cm from the PTV and 3 cm thick) and the two closest OARs. Robustness of each extracted features (R fe ), has been analyzed in terms of standard deviation normalized with the mean value of the features.

Conclusion In this study, we investigated the impact of variations in the number of OSEM subsets on pre-clinical PET radiomic features. Our results indicate that varying the OSEM subsets changes the radiomic output derived from these images. Therefore, more research into the impact of OSEM reconstruction parameters on clinical PET needs to be undertaken in order to further understand its impact on radiomic features. References : [1] M. E. Juweid et al,“Positron-emission tomography and assessment of cancer therapy” N. Engl. J. Med. , vol. 354, no. 5,pp. 496–507, 2006. [2] A. M. Morey et al,“Effect of Varying Number of OSEM Subsets on PET Lesion Detectability” J. Nucl. Med. Technol. , vol.41, no.4, pp.268–273,2013. [3] G. J. R. Cook et al,“Radiomics in PET:principles and applications” Neuroimage , pp.269–276,2014. [4] P. Whybra et al, “Assessing radiomic feature robustness to interpolation in F-FDG PET imaging” Sci. Rep. , no. Dec, pp.0–10, 2019. [5] I. Shiri et al,“The impact of image reconstruction settings on 18F-FDG PET radiomic features: multi-scanner phantom and patient studies” Eur.Radiol. , vol. 27, no. 11, pp. 4498–4509, 2017. PO-1569 Early prediction of tumour-response to radiotherapy in NSCLC patients. L.M. Amugongo 1 , E. Vasquez Osorio 1 , A. Green 1 , D. Cobben 1 , M. Van Herk 1 , A. McWilliam 1 1 University of Manchester, Division of Cancer Sciences- Faculty of Biology- Medicine and Health- University of Manchester, Manchester, United Kingdom Purpose or Objective Given the heterogeneity of tumour response to radiotherapy in lung cancer, there is a clinical need to identify patients who are responding to treatment early. Early identification of good responders will enable better treatment personalisation through adaptation of radiotherapy. Advanced imaging analyses and machine learning techniques are increasingly being used for prediction to aid clinical decisions. In this study we apply a novel technique, using on-treatment cone-beam computed tomography (CBCT) images to predict tumour volume and shape at the end of treatment. Material and Methods CBCTs of 240 non-small cell lung cancer (NSCLC) patients, treated with 55Gy in 20 fractions, were collected. CBCTs were rigidly registered to the planning CT (pCT) and their intensities corrected. All CBCTs were cropped to focus on the tumour using a 5 mm extension around the gross tumour volume contour defined on the pCT. Next, intensity values were extracted in each voxel across all CBCTs from day 1, 2, 3 and 7 (standard imaging protocol). For each patient and each voxel three regression models were fitted; linear, second and third-degree polynomials. These models were used to predict the intensity value for each voxel at the end of treatment (day 20). To evaluate the performance of each model, the root mean square error (RMSE) in pixel value was calculated for every patient. Finally, the volume of the predicted tumour was