ESTRO 38 Abstract book

S209 ESTRO 38

Models’ discriminatory performance was assessed using receiver operating characteristic (ROC) area under the curve (AUC) analysis, accuracy, and Youden index ( YI = specificity + sensitivity - 1 ).

with

each

other.

Conclusion Assuming that the sensitivity of tumor cells to treatment may vary as a function of distance from the tumor border, dividing the tumor volume into subregions can characterize tumor cells and identify them from more homogeneous regions. The proposed novel feature set was capable of describing intra-tumor heterogeneity with only 20 features and outperformed conventional radiomics. OC-0407 CT-based Radiomics for Risk Stratification in Prostate Cancer S. Osman 1,2 , R.T.H. Leijenaar 3 , A.J. Cole 1,4 , A.R. Hounsell 2 , K.M. Prise 1 , J.M. O'Sullivan 1,4 , P. Lambin 3 , C.K. McGarry 1,2 , S. Jain 1,4 1 Queen's University Belfast, Centre for Cancer Research & Cell Biology, Belfast, United Kingdom ; 2 Northern Ireland Cancer Centre, Medical Physics, Belfast, United Kingdom; 3 The D-Lab: Decision Support for Precision Medicine- GROW- Maastricht University Medical Centre- The Netherlands, School for Oncology and Developmental Biology, Maastricht, The Netherlands ; 4 Northern Ireland Cancer Centre, Department of Clinical Oncology, Belfast, United Kingdom Purpose or Objective The prognostic and predictive value of MRI- based radiomics imaging features for prostate cancer (PCa) has been demonstrated in a number of publications. However, there is no clear consensus on the most important imaging biomarkers for prostate or their clinical applicability. This could be due to the lack of interoperability (i.e. variable imaging protocols, scanners, software or uncertainties in the manual definition of tumours). Planning CT scan protocols for PCa are typically standardised for all patients treated with external beam radiation therapy (EBRT). This study explored the role of CT-based radiomics features in PCa Gleason score (GS) and risk group classification. Material and Methods The study population consisted of 506 PCa patients from a clinically annotated database all treated in a single centre using EBRT. After applying exclusion criteria (available CT scans with 2.5mm slice thickness and no artefact), 342 patients were included in the final analysis. CT-based radiomics features were extracted for prostate gland only (PO) structure. CT scans were re-sampled to 2.5mm isotropic voxels and the range of Hounsfield Units (HU) discretised to 10 HU bins prior to feature extraction. For 20 patients with scan and re-scan data the interclass correlation coefficient ICC was used to identify a set of robust features to use for our analysis. Features with ICC > 0.8 were considered reproducible. Pairwise correlation testing was also employed to remove redundant features. Penalized regression analysis using LASSO generalized linear model was used. Optimal lambda was estimated using 10-fold cross validation (CV) repeated 100 times, the lambda value one standard error away was used. Models were evaluated through 10-fold CV repeated 100 times.

Results Classifiers employing CT-based radiomics features distinguished between GS ≤ 6 vs. GS ≥ 7 with (AUC = 0.83, YI = 0.16) and GS=7(3+4) vs. GS=7(4+3) with (AUC = 0.98, YI = 0.80). Excellent performance was observed for radiomics-based classification of PCa risk groups for low vs. high risk group (AUC = 1.00, YI = 0.81), low vs. Intermediate risk (AUC = 0.97, YI = 0.58). Models showed poor performance in distinguishing between GS 7 vs. GS > 7 (AUC = 0.65, YI = 0.02) as well as intermediate vs. high risk patients (AUC = 0.66, YI = 0.00). Applying augmentation methods to balance the data improved classifiers’ performance in all cases.

Conclusion Results show that radiomics features from routinely acquired planning CT scans may provide insights into prostate cancer aggressiveness (i.e. Gleason score and risk-group) in a non-invasive manner. Our classifiers were especially accurate in identifying high-risk patients. External validation, and prospective studies are warranted to verify the presented findings.

Proffered Papers: PH 8: Proffered paper: Handling intra-fraction motion in MR guided RT

OC-0408 Impact of bladder filling on the magnitude of prostate intra-fraction motion assessed in 3D Cine-MR H. De Boer 1 , D.M. De Muinck Keizer 1 , J.R.N. Voort van Zyp 1 , N.A.T. Van den Berg 1 , F.J. Pos 2 , U.A. Van der Heide 2 , B.W. Raaymakers 1 , J.J.W. Lagendijk 1 , L.G.W. Kerkmeijer 1