Abstract Book

S1084

ESTRO 37

The NTCP model of grade 2+ rectal bleeding was as follows: S = −17.49 + Platelets (1000/µL) * (-0.025) + Risk group* Corresponding coefficient (low-risk group = 0; intermediate-risk group = 19.07; high-risk group = 20.41) + V 65 * 0.045. Conclusion A LASSO-based multivariable NTCP model comprising three important predictors (platelet count, risk group and V 65 ) was established to predict the incidence of grade 2+ late rectal bleeding after IMRT. EP-1993 Evaluation of the pertinence of CT-based radiomics shape features with 3D printed phantoms E.J. Limkin 1 , S. Reuzé 1 , R. Sun 1 , A. Schernberg 1 , A. Alexis 1 , A. Dirand 1 , E. Deutsch 1 , C. Ferté 2 , C. Robert 1 1 Institut Gustave Roussy, Radiotherapy, Villejuif, France 2 Institut Gustave Roussy, Medical Oncology, Villejuif, France Purpose or Objective Studies have shown that radiomics shape features correlate with patient outcomes. However, how these features vary in function of tumor complexity and volume remain unknown. We created 3D printed shape phantoms with increasing degrees of spiculatedness to determine the relationship of shape features with shape complexity and volume. Material and Methods 28 tumor models were mathematically created using spherical harmonics, with the degree l being increased in increments of three (range: 11-92). Models were then 3D printed with identical base diameters of 5cm and heights of 3cm, CT scanned, and contoured semi-automatically with an open source software. First, segmented shapes were spatially resampled to have constant volume and eliminate volume-dependence in calculations. Next, 3 representative phantoms (l = 11,47,86) were resampled to have 25, 50, 75, 125 and 150% of the original volume to see how features behave with volume differences. 23 3D features (Table 1) were then extracted in MATLAB with an in-house code. Statistical analyses were performed with R ver. 3.3.2. Z-score transformation was applied to standardize feature values. Spearman's rank- order correlation coefficients (CC) were computed to determine the relationship of each shape feature with tumor complexity and with changes in volume. Principal component analysis (PCA) was performed to determine which among the features contribute most to the variance of the data. Pairwise correlations were computed to determine which among the features are highly correlated.

spiculatedness, 6 had an inverse relationship, while 9 algebraic ellipsoid points had no significant correlation with complexity (Fig 1A) and were thus excluded from subsequent analysis. PCA showed that 83% of the variance is explained by the first principal component (PC), the second adding only 11.5%. Most features contribute significantly to the first PC, with seven having correlations >0.90. Pairwise correlations showed that majority of the features have high CCs >0.75, except for b and c ellipsoid radii and chi^2. For varying volumes, surface to volume ratio was strongly inversely correlated with increasing volume (r=-0.94) and compactness1 had r=1, p<2.2 -16 . Five features had non-significant direct correlations with volume. Correlations were identical for l=47 and 86 (r=0.68, p=0.14) but not for l=11 (r=0.12, p=0.82) ( Fig 1B).

Table 1. Radiomics shape features

Results For constant volume, 13 features had strong (Spearman’s rho >0.50) correlations with the variations of the complexity of the shape phantoms, all with p-values <0.05. 7 features increased with increasing