ESTRO 36 Abstract Book

S863 ESTRO 36 _______________________________________________________________________________________________

1 University of Manchester, Division of Molecular and Clinical Cancer Sciences- Faculty of Biology- Medicine and Health, Manchester, United Kingdom 2 The Christie NHS Foundation Trust, Department of Infomatics, Mancehster, United Kingdom Purpose or Objective Radiomics aims to extract features from medical images that are prognostic for outcome and may help optimize treatment. As far as the tumour is concerned, most work has focused on pixel values inside the gross tumour volume (GTV). The aim of this work is to develop a generic methodology to also sample pixels outside a tumour volume, assuming that these may carry information about microscopic tumour spread and therefore might predict outcome. Material and Methods We analysed data from a cohort of 1101 non-small cell lung cancer patients treated with IMRT to 55 Gy in 20 fractions. To evaluate the CT pixel values at various distances inside and outside the GTV, we calculated a signed distance transform of the GTV, which was subsequently used to efficiently collect cross-histograms of the CT density versus distance from the GTV edge. Based on these cross- histograms various pixel statistics were determined as function of the GTV distance, here we report only on the mean pixel value, giving a curve of mean CT value versus distance. The mean of these curves was calculated for patients that were alive (652) and dead (449) at 12 months after start of therapy, censored for follow-up. Significance of the difference was tested by permuting the dead/alive labels 1000 times to create mock differences and counting how often the true difference exceeded the mock difference. Significant regions were defined and the mean pixel value from those regions used as variable in a cox proportional hazard model, splitting the patients on the median of the mean region density, while correcting for age and tumour size. As the outside of the tumour can also include chest wall and mediastinum, we repeated the analysis only analysing pixels inside the lungs. Results There was a significant different average pixel value in the region 0-1 cm outside the GTV for dead and alive patients (fig. 1) that translated to a hazard ratio (HR) of 1.4, p<10 - 5 (corrected for tumour size and age), survival curves split at median density value (fig. 2A). However when only pixels inside the lungs were analysed, the HR reduced to 1.1, p=0.15; i.e. no longer significant (fig. 2B). This finding indicates that the mean pixel values represent mediastinal attachment rather than microscopic disease. Not correcting for tumour size, both signals incorrectly predict outcome significantly (e.g. fig. 2C for lung pixels only).

radiomics, i.e. the change of radiomic features over time, has not yet been extensively explored. Cone-beam CT (CBCT) imaging can be performed daily for lung cancer patients and is therefore a potential candidate for delta radiomics, which may allow further treatment individualization. In this study we explored delta radiomics using CBCT imaging by investigating the number of features changing at a specific time point during treatment. Moreover, we investigated the differences between patients having an overall survival of less or more than 2 years. Material and Methods A total of 40 stage II-IV NSCLC patients, receiving curatively intended radiotherapy for a period of at least six weeks, were included in the study. The CBCT images used in this study were 1) CBCT prior to the first fraction of treatment (CBCTfx1), 2) CBCT prior the second fraction of treatment (CBCTfx2), 3) CBCT one week after the start of treatment (CBCTweek2), 4) CBCT three weeks after the start of treatment (CBCTweek4) and 5) CBCT five weeks after the start of treatment (CBCTweek6). For 38 patients CBCTfx1 and CBCTfx2 were available, whereas for 33 patients all weekly CBCTs were available. All patients had a minimal follow-up of 2 years. Per time point, a total of 1046 radiomic features were derived from the primary tumor volume. The images prior to the first and second fraction were used to calculate the variability in imaging features using the coefficient of repeatability (COR), defined as 1.96*SD. The weekly images were used to investigate the number of features changing more than the COR with respect to baseline (CBCTfx1). Results Figure 1 represents the total number of features that changed more than the COR, ranging from 0 to 999 features. The median number of features that changed for the group with overall survival <2 years was 279, whereas this was 500 for the group with overall survival >2 years (Mann-Whitney U test, p = 0.06). For 8 out of 10 patients that survived >2 years, more features (31.7%) changed one week after CBCTfx1 than for 13 out of 23 patients that did not survive two years.

Conclusion This study shows that a large proportion of the radiomic features derived from cone-beam CT images change significantly during the course of treatment, meaning that an interval of about two weeks is feasible for a radiomics study using CBCT imaging. The larger number of features that changed in the group with overall survival >2 years could reflect an early response of the tumor to the treatment. In future research, the prognostic value of changing radiomic features (delta radiomics) should be explored in a larger cohort. EP-1601 Do higher CT pixel values outside the GTV predict for poorer lung cancer survival? M. Van Herk 1 , J. Kennedy 2 , E. Vasquez Osorio 1 , C. Faivre- Finn 1 , A. McWilliam 1