Abstract Book

S267

ESTRO 37

OC-0512 Space based normal tissue complication probability modeling G. Palma 1 , A. Buonanno 2 , S. Monti 3 , R. Pacelli 4 , L. Cella 1 1 National Research Council, Institute of Biostructures and Bioimaging, Naples, Italy 2 Università degli Studi della Campania "Luigi Vanvitelli", Ingegneria Industriale e dell'Informazione, Aversa CE, Italy 3 IRCCS SDN, Image Processing Department, Naples, Italy 4 Federico II University School of Medicine, Department of Advanced Biomedical Sciences, Naples, Italy Purpose or Objective Normal Tissue Complication Probability (NTCP) models have been historically built from Dose-Volume Histogram (DVH) of critical organs. However, an increasing awareness have been triggered in the scientific community regarding the value of including the spatial information of dose distributions in the analysis of radiation-induced toxicity (RIT). Up to now, statistical inference on spatial signature of RIT was addressed in recent studies [Acosta 2013, Chen 2013, Dean 2016], but no insight on actual NTCP models including full dose spatial info has been achieved. The purpose of this study is to propose a new formalism to fill this gap of knowledge and to develop a space based NTCP (SpNTCP). Material and Methods A SpNTCP approach was devised assuming the availability of a training set of N patients’ dose distributions (sampled on m voxels each) classified according to a binary RIT. A logistic regression on the dose values is performed voxelwise, and, accordingly, a collection of m RIT probabilities is computed on each voxel for a test dose distribution. Similarly, a collection of m weights is obtained as the inverse of each logistic Confidence Interval (CI) length. The SpNTCP is then computed as the weighted mean of the m RIT probabilities. The SpNTCP was first demonstrated on a set of synthetic plausible dose maps, classified by thresholding the generalized Equivalent Uniform Dose (gEUD) simulated for given Radio-Sensitivity Map (RSM) and n volume-effect parameter. Moreover, an estimate of the RSM was reconstructed by computing the previous SpNTCP on a new set of hot spots evenly spanning the field of view. Then the performance was evaluated on a cohort of 98 RT thoracic cancer patients classified for lung fibrosis (18 events). For this purpose, each dose map was normalized to a common anatomical reference via a log- diffeomorphic demons registration tool, as described in [Monti 2017]. Lyman-Kutcher-Burman (LKB) models were trained for comparison. Results The learning curves show that the SpNTCP model outperformed LKB one, as shown in Figs. 1-2, with an increasing model bias at lower n values.

between PBT and VMAT showed large differences in both training and validation cohort, especially in the remaining brain. The mean brain dose of the PBT plans was significantly lower compared to VMAT (median training: 6.9 Gy vs 18.6 Gy, median validation: 8.5 Gy vs 16.0 Gy; p < 0.001). For alopecia grade 2, plan comparison between PBT and VMAT predicted a potential median NTCP reduction for PBT of approx. 5% (range: - 39% – 32%) in the training and 1% (range: -25% – 37%) in the validation cohort. A reduction of NTCP for alopecia grade 2 for PBT by more than 10% was observed for 12/113 patients in training and for 9/71 patients in validation.

Conclusion Plan comparison showed a large reduction in dose to the brain using PBT instead of VMAT. We found significant correlations between the occurrence of early side effects and DVH parameters of associated OARs for patients with brain tumours receiving PBT. A relevant reduction of NTCP (>10%) for PBT was calculated for approx. 10 % of the patients. However, due to the large range of NTCP reduction, patient individual calculations are mandatory. After inclusion of more relevant late side effects and neurocognitive changes, these models may be used to identify patients who are likely to benefit most from PBT [1]. [1] Langendijk JA et al. (2013) Radiother Oncol 107, 267- 273.