ESTRO 38 Abstract book

S208 ESTRO 38

Oncology-Pathology, Stockholm, Sweden ; 3 Politecnico di Milano, - Department of Electronics- Information and Bioengineering, Milan, Italy; 4 Stockholm University, Department of Physics, Stockholm, Sweden Purpose or Objective Recently, numerous studies have developed advanced imaging biomarkers in order to capture cancer imaging phenotypes. By extracting large numbers of quantitative features, radiomics offers new opportunities in cancer outcome modeling. However, radiomic features are quite sensitive to different factors, and it is often difficult to interpret their physiological meanings. The objective of this study was to develop a novel physiologically interpretable feature set describing intra-tumor heterogeneity in patients with lung cancer and test it as an imaging biomarker for survival prediction. Material and Methods Longitudinal PET-CT images from 30 patients diagnosed with non-small cell lung cancer were analyzed. After preprocessing the images, gross target volumes (GTV) were segmented semi-automatically. Inspired by the assumption that partial tumor response to the therapy would be explained by the existence of sub-regions within the GTV, we introduced the size-aware longitudinal pattern (SALoP), which aims at quantifying variations in the structure and function of the tumor. To compute SALoP, we partitioned the GTVs into separate subregions based on the distance of each voxel inside the GTV from the tumor border, i.e. for every 0.5cm of distance from the tumor border, one region was added. Then, the change in average intensity of each subregion between the two scans was calculated. For comparison, radiomic analysis was performed by extracting 451 features. Reproducibility of SALoP and radiomic features were investigated on an external test-retest dataset. Dimensionality reduction was performed by applying a forward feature selection (FFS) algorithm, and a support vector machine (SVM) was employed as the prediction model.

Figure 1. CONSORT diagram of the patient cohort.

Results Reproducibility of the SALoP set was substantiated by achieving a high agreement when it was applied on test- retest dataset. Without FFS, SALoP features outperformed radiomics significantly. Feeding the prediction model with only selected features, the combination of SALoP and radiomics resulted in the highest predictive values. For SALoP, a combination of PET and CT features led to higher predictive values than CT and PET features separately either with or without FFS. Applying FFS contributed an improvement of 15 percent (0.71 to 0.86) in predictive power of radiomics and 5 percent (0.90 to 0.95) for the SALoP. This implies that large numbers of radiomic features were either redundant or lacking informative value, whereas the SALoP features were more consistent

Figure 2. Upper panel: Isocurves of the predicted 90-day mortality rates 1% and 5% as a function of the mean lung dose and patient age. The shaded areas corresponds to the 95% confidence intervals. Lower panel: Calibration plot comparing observed and modelled 90-day mortality rates. The vertical error bars are the 95% binomial confidence interval and the numbers above are the observed data in each bin. Conclusion We demonstrated registry based outcome modeling as a means to predict the infrequent endpoint of early mortality following curative intended radiotherapy for NSCLC. Age and MLD were found to be associated with a higher risk of early mortality. OC-0406 Early survival prediction in non-small cell lung cancer with PET/CT size aware longitudinal pattern M. Astaraki 1,2 , C. Wang 1 , G. Buizza 3 , I. Toma-Dasu 2,4 , M. Lazzeroni 2,4 , Ö. Smedby 1 1 KTH Royal Institute of Technology, Department of Biomedical Engineering and Health Systems, Stockholm, Sweden ; 2 Karolinska Institutet, Department of