ESTRO 36 Abstract Book

S33 ESTRO 36 _______________________________________________________________________________________________

2 Zuyderland Medical Centre, Department of Surgery, Heerlen, The Netherlands 3 Zuyderland Medical Centre, Department of Nuclear Medicine, Heerlen, The Netherlands Purpose or Objective For several tumour sites it was shown that quantitative radiomics features, derived from CT images, unravel valuable prognostic information. However, the large number of available features increases the risk of overfitting. Typically, test-retest scans allow to reduce this number by selecting a robust feature subset. However, these test-retest scans are not available for all tumour sites. Hence we hypothesized that different phases of respiratory-correlated 4D CT-scans (4DCT) can be used as alternative to test-retest imaging to select robust features. To test this hypothesis, we assessed the repeatability of 542 radiomics features in a test-retest and two 4DCT datasets of lung and oesophageal cancer patients. Material and Methods The publically available RIDER dataset, consisting of test- retest CT-scans of 27 non-small cell lung cancer (NSCLC) patients, and 4DCT-scans of 22 NSCLC (4D-Lung) and 20 oesophageal cancer patients (4D-OES) were analysed. The 4DCT-scans contained 8 phases of the breathing cycle. The gross tumour volume (GTV) of the primary tumours were manually delineated. In total, 70 radiomics features describing the tumour shape, intensity and texture, and 472 wavelet-filtered features were calculated within the GTVs. A concordance correlation coefficient (CCC) ≥ 0.85 was used to identify robust features, either between the test-retest scans or over all phase pairs in the 4DCT scans. Results Unfiltered features in general showed a higher robustness than wavelet-filtered features. In total 34/70 (49%) unfiltered features and 122/472 (26%) wavelet features were stable in both the test-retest dataset and the 4D- lung dataset. The four features selected previously to be prognostic in lung and head-and-neck cancer (Aerts et al 2014), had a minimum CCC > 0.95 in both datasets. In the 4D-OES dataset 205/542 (38%) features showed a high robustness, of which 42 unfiltered and 99 wavelet-filtered features were also stable in the 4D-lung dataset. Due to the fact that the image acquisition settings and hardware were exactly the same in the 4D-lung and 4D-OES scans, this partial disconcordance suggests that the remaining stable features might be tumour site specific.

Material and Methods This study is based on 4D-CBCT and 4D-CT scans of 19 non- small-cell lung cancer (NSCLC) patients subjected to curatively intended radiotherapy initiated between March 2012 and August 2014. Lung ventilation was measured as voxel-wise Jacobian determinants (JD) computed by deformable image registration between expiration phases and inspiration phases. All image registrations were carried out by the freeware tool elastix (elastix.isi.uu.nl). 4D-CT scans acquired before treatment were chosen as gold standard for ventilation. The clinical 4D-CBCT projection images of the first treatment fraction were improved by the procedure described in [PMB, 15, 5781, 2016], which corrected projections for image lag, detector scatter, body scatter, and beam hardening. Clinical and improved projection images were binned and FDK- reconstructed by software in the RTK-package (www.openrtk.org). All CBCT reconstructions were rigidly resampled into CT-space. Before deformable image registration a 1x1x1cm wide median filter was applied on all images. For each patient the clinical 4D-CBCT JDs and the improved 4D-CBCT JDs were voxel-wise compared to 4D-CT JDs by Spearman correlation and the resulting correlation coefficients were analysed by a paired t-test. Results The clinical projection images were improved successfully and both versions of projection images were reconstructed to 4D-CBCT. Deformable registrations were carried out on clinical 4D-CBCT, improved 4D-CBCT, and 4D-CT. The sample mean for correlations between clinical 4D-CBCT and 4D-CT JDs was 0.297±0.154 while the sample mean for clinical 4D-CBCT and 4D-CT JDs was 0.362±0.106 (±1 SD). A paired t-test of the Fisher transform correlation values showed the difference to be statistically significant (p=0.047).

Conclusion Ventilation computed from clinical 4D-CBCT and improved 4D-CBCT were compared to 4D-CT ventilation. The improved 4D-CBCT ventilation correlated better with 4DCT ventilation, suggesting that the improved 4D-CBCT are superior to clinical 4D-CBCT for measuring lung ventilation. Further improvements on measuring lung ventilation from 4D-CBCT may possibly be achieved by adding iterative reconstruction. OC-0067 4DCT imaging to assess radiomics feature stability: an investigation for thoracic cancer R.T.H.M. Larue 1 , L. Van De Voorde 1 , J.E. Van Timmeren 1 , R.T.H. Leijenaar 1 , M. Berbée 1 , M.N. Sosef 2 , W.M.J. Schreurs 3 , W. Van Empt 1 , P. Lambin 1 1 Maastricht University Medical Centre - GROW-School for Oncology and Developmental Biology, Department of Radiation Oncology - MAASTRO, Maastricht, The Netherlands

Figure 1. Venn chart visualizing the overlap of stable features with CCC>0.85 in the RIDER, 4D-lung and 4D-OES dataset.