ESTRO 37 Abstract book

S1096

ESTRO 37

chemo-radiotherapy, was downloaded from NSCLC- Radiomics collection in The Cancer Imaging Archive (TCIA). The latter for test, received Targeted treatment (35) or Non-Targeted treatment (19) for ALK+ mutation, were all reported as ALK+ in Fudan University Shanghai Cancer Center. An in-house code-set based on Matlab 2015b (Mathworks, Natick, MA, USA) was used for feature extraction from the pre-treatment computed tomography (CT) images. Totally 203 features were extracted. A test- retest was conducted on RIDER NSCLC dataset and a single-measurement, absolute-agreement, 2-way mixed- effects model was selected for ICC calculation. Features with ICC value larger than 0.9 was considered as stable, which were then incorporated in the least absolute shrinkage and selection operator (LASSO) Cox regression and a leave-one-out cross-validation was practiced. C- index was calculated also in each therapy group. And log- rank test was performed for stratified analysis. Rad_index was calculated as the sum of feature multiplied by its coefficient. (Figure 1).

Figure 2 risk group identification based on Rad_index (Low risk: Rad_index ＜ -0.00903; High risk: Rad_index≥- 0.00903). Conclusion Compactness, Wavlet_GLCM_Std, and Wavlet_GLRMS_RLN might have strong association with prognostics of inoperable NSCLC patients with ALK+ mutation, of which the radiomics signature was comprised, demonstrating well-validated value for predicting and identifying high or low risk patients in non-targeted treatment group. These finding indicate that the radiomics signature should be used with cautions in ALK+ mutated patients, and is prospectively applicable as prognostic indicator for inoperable ALK+ mutated patients received non-targeted treatment. However, in terms of patients received targeted therapy, further investigation is needed. EP-2008 Image based data mining using per-voxel Cox regression A. Green 1 , A. McWilliam 1 , E. Vasquez-Osorio 1 , W. Beasley 1 , M. Van Herk 1 1 Manchester Cancer Research Centre- Division of Cancer Science- School of Medical Sciences- Faculty of Biology- Medicine and Health- University of Manchester- Manchester Academic Health Sciences Centre- Manchester- UK- M13 9PL- UK, Radiotherapy Related, Purpose or Objective Recent work has shown the efficacy of image based data mining (IBDM) relating planned radiotherapy dose to treatment outcome for hypothesis generation in dose- effect modelling. However, all techniques used to date do not account for confounding variables in the image based statistic used to identify dose sensitive regions, instead relying on correction after identifying a region of interest. In this work, we introduce a novel IBDM technique in which a Cox regression is performed in each voxel of the 4D dose distribution obtained by mapping all patients to a reference patient. By using a Cox regression in this way, it is possible to control for other variables when assessing the impact of radiation dose on tissue on a per-voxel basis providing unbiased estimates of the location of dose-response, which is particularly important for variables that affect dose and outcome such as GTV volume. Material and Methods A cohort of 870 lung cancer patients was collected, including planning CT, planned dose distribution, outcome and clinical data. Dose distributions were spatially normalised using deformable image registration to a common frame of reference. A mask was produced such that only regions receiving a mean dose greater than 10 Gy were included in the analysis to remove low-dose regions which would lead to spurious Cox-regression fits. Hazard ratio for dose per Gy and the other variables was

Figure 1 workflow

Results Three-feature signature was developed and C-index was calculated in test set (CI = 0.65). In each treatment group, C-index of Non-Targeted therapy was 0.85. Nonetheless, poor result was observed in Targeted therapy group (CI = 0.54). Rad_index = 0.04435912* Compactness - 0.01378887* Wavlet_GLCM_Std + 0.01888021* Wavlet_GLRMS_RLN. And Patients were divided into high or low risk according to Rad_index. log- rank p-value was 0.036 in Non-Targeted therapy group (Figure 2).

Table 1 Patient characteristics of two cohort of patients