ESTRO 38 Abstract book

S514 ESTRO 38

characteristics of RILF. This study aims to empirically test and propose an appropriate HU range which corresponds to RILF. Material and Methods Voxel HU values were extracted from 68 patient lung CTs at pre-treatment (simulation CT) and 6-months post- treatment (diagnostic CT). A team of 2 radiation oncologists and 2 radiologists scored the patient population for RILF severity on a five-grade scale: 0 = no fibrosis, 1 = 1 pulmonary segment equivalent of fibrosis (PSEF), 2 = 2 PSEF, 3 = 3 PSEF, 4 = 4 PSEF, 5 = >5 PSEF. We defined PSEF as the volume of half the right middle lung lobe. Post-treatment images were deformably registered to pre-treatment CTs using an in-house algorithm developed in MevisLab TM . The registered image was corrected using the non-irradiated lung to account for scanner differences between images acquired at baseline and follow-up. For the irradiated lung parenchyma (defined as total treated lung volume minus PTV) the voxel HU counts were binned as histograms. The difference between follow-up and baseline histograms was computed and its area normalized to unity. We then computed the integral (HU int ) of the difference histogram between a lower (HU L ) and upper (HU U ) HU range. HU L was varied between -860 HU (below which is health lung parenchyma) to -10 HU and HU U was varied between 0 HU to 460 HU (above which is no longer soft tissues). Each combination of HU L and HU U yielded a unique HU int for each patient. For the patient cohort, the Spearman correlation (ρ) was calculated between these unique HU int and the mode of the physician assigned scores. See Figure 1 for an illustration of the method.

Conclusion For a quantitative analysis of RILF, we found a value of ≈- 270HU to be appropriate for HU L . We were unable to establish a HU U recommendation given HU U choice is less impactful on the predictive ability of HU int . We suspect this is due to the reduced number of voxels that occupy the higher densities. These recommendations need to be externally validated on larger populations. PO-0951 How to build accurate prediction models without sharing patient data across hospitals? I. Zhovannik 1,2 , Z. Shi 1 , F. Dankers 1,2 , T. Deist 3 , A. Traverso 1 , P. Kalendralis 1 , R. Monshouwer 2 , J. Bussink 2 , R. Fijten 1 , H. Aerts 4,5,6 , A. Dekker 1 , L. Wee 1 1 Maastricht University Medical Center +, Radiation Oncology - Maastro Clinic, Maastricht, The Netherlands ; 2 Radboud University Medical Center, Radiation Oncology, Nijmegen, The Netherlands ; 3 Maastricht University, The D-Lab: Decision Support for Precision Medicine, Maastricht, The Netherlands ; 4 Harvard Medical School, Brigham and Women's Hospital, Boston, USA ; 5 Maastricht University Medical Center +, Radiology, Maastricht, The Netherlands ; 6 Dana-Farber Cancer Institute, Computational Imaging and Bioinformatics Laboratory, Boston, USA Purpose or Objective Imagine if you had an intelligent model that could select the most optimal treatment for a cancer patient, i.e. optimizing the tradeoff between tumor control and side effects. Are we close to having such a model? Patient data are unstructured and distributed in patient databases all over the world – that is the greatest challenge. In addition, there are bureaucratic and legal barriers to share (even anonymized) patient data. To avoid the difficulties of data-sharing we proposed to use a distributed learning approach where the privacy-sensitive data never leaves the hospital. In this approach, only a model is shared and optimized. Material and Methods We used the pioneering approach of Aerts et al (2014) to build a model in a distributed fashion without data sharing between two hospitals. The developed Cox proportional hazards radiomic model splits a patient cohort into two groups by the median mortality risk. To train and validate the model we used two cohorts of non-small cell lung cancer patients of the two hospitals (441 and 221 patients respectively). We extracted radiomic image biomarkers from the Gross Tumor Volume (GTV) of both cohorts. Radiomics ontology (RO) and radiation oncology ontology

Results As seen in Figure 2, for a fixed HU U

value of 200 HU,

changes in the HU L showed dramatic effects on ρ, with values of 0.07 (p<0.001) at -850 HU increasing to 0.63 (p<0.001) at -270 HU. With HU L fixed at -270 HU, we observed no more than 20% variation in r when HU U was varied. The best performance (ρ=0.63) was for HU L = -270 HU and HU U = 195 HU.