ESTRO 38 Abstract book

S162 ESTRO 38

Results The employed 4D-CT phases were of an acceptable quality. A few patients had breathing motion artefacts close to the diaphragm. The DRA performed well in regions of reasonable image contrast. Some patients with very low contrast lung regions showed several deformation artefacts influencing the FL segmentation especially near the diaphragm. The V-SPECT segmentations also often indicated regions near the diaphragm. The V LUNG varied widely, with a median volume [range] on the 4D-ex of 3159cm 3 [2093–5446] (see Table1). The two segmented FL volumes for each patient were comparable in size, with a median difference between the two volumes (V FL-SPECT -V FL- 4D ) of 33cm 3 [-80–142] (see Figure 1a). The median OF of the two volumes was 38%[27%–69%]. The volume of lung indicated as inferiorly functional by both methods, corresponded to a median 47%[42-58] of V LUNG (see The ventilation FL segmentations obtained using 4D-CT and V-SPECT did not identify the same lung voxels. However, though the overlap of the two FL volumes was low, the two methods indicated the same general areas as inferiorly functional. It is possible that the 4D-CT ventilation FL segmentations may highly depend on the quality of the 4D-CT scans and the quality of the deformation in low contrast areas. Figure1b). Conclusion

1 Maastricht Radiation Oncology Maastro Clinic, Radiotherapy, Maastricht, The Netherlands; 2 Princess Margaret Cancer Centre, Radiation Physics, Toronto, Canada ; 3 Princess Margaret Cancer Centre, Radiation Oncology, Toronto, Canada Purpose or Objective Imaging biomarkers derived from computed tomography (CT) scans (“radiomics”) have been used to generate predictive models for clinical outcomes such as overall survival (OS) following radiation therapy. Models based on imaging features have been shown to achieve similar performance to traditional models based on semantic features (e.g. tumor volume and staging). However, several challenges remain including: generalizability and reproducibility of radiomics-based models, risk of overfitting due to high number of model covariates, and strong correlations between features and clinical variables. These issues limit usage of radiomics-based prediction models in decision support systems. In our study, we applied unsupervised and supervised learning techniques to investigate the relationships between radiomics features, clinical features, and clinical outcomes with a special focus on the relationship between radiomics features and tumor volume. Material and Methods We used two publicly available datasets: 420 Non Small Cell Lung Cancer patients and 130 Squamous Cell Carcinoma oropharynx patients. CT scans, manually delineated GTV (Gross Tumor Volumes), clinical variables were available. Radiomics features were extracted from GTVs using the open source Pyradiomics library. Hierarchical clustering was used to discover groups of patients with similar radiomic signatures. Dependencies between radiomics features and tumor volume were evaluated with the Spearman concordance correlation coefficient ( ϱ ). Overall survival was compared between clusters using Kaplan-Meier analysis. Tumor volume distributions were also compared. Bootstrap-based methods were used to evaluate the stability and importance of radiomic features in predicting 2-year OS. Results As shown in Figure1A, it is possible to identify two groups of patients with different OS by using all radiomics features. However, when only using volume-independent features ( ϱ <0.2) the groups cannot be distinguished (1B). PCA (1C) revealed that tumour volume is highly correlated with the first and second dimensional PCA, which explains the higher variability in the data. Semantic features and volume explained most of OS variability in both datasets (1D).

PV-0314 Machine learning helps identifying relations and confounding factors in radiomics-based models A. Traverso 1,2 , M. Kazmierski 1 , L. Wee 1 , A. Dekker 1 , M. Welch 2 , A. Hosni 3 , D. Jaffray 2 , A. Hope 3