A Novel Framework for Spatial Normalization of Dose Distributions in Voxel-Based Analyses of Brain Irradiation Outcomes

Abstract

Voxel-based population analyses (VBAs) are gaining momentum in radiation therapy (RT) outcome modeling. VBAs rely on the accurate spatial normalization (SN) of different patients’ anatomies on a common coordinate system (CCS) by inter-patient elastic registration of planning CTs. VBA has been applied to different toxicity endpoints and anatomical regions. However, as for brain applications, the cerebral anatomy, while being well characterized by several MRI modalities, shows a poor CT contrast, thus resulting in potentially large inaccuracies of VBAs based on the planning CT alone. Aim of the present study was to devise and evaluate a novel SN framework for VBA in brain irradiation based on multimodal imaging. We analyzed 43 brain tumor patients treated at a single institution with proton therapy. All patients underwent a 3D T1-weighted (T1w) MR sequence. The Montreal Neurological Institute (MNI) space [Evans et al. 1993] was chosen as CCS. The SN pipeline was: 1. Intra-patient affine registration of CT to T1w; 2. Inter-patient B-spline registration of T1w to MNI; 3. doses propagation to MNI by applying the deformation fields of steps (1) and (2). To specifically evaluate the accuracy of dose SN, the dose organ overlap (DOO) scores [Acosta et al. 2013] were computed on White Matter (WM), Gray Matter (GM) and liquor (CSF) before (DOOpre) and after (DOOpost) SN. In addition, the trends in the DOOs distribution were investigated: a Gaussian mixture model of the 6D data (DOOpre and DOOpost for each structure) was selected by choosing the number of clusters that maximizes the average silhouette over all data, and each 6D point (i.e. patient) was assigned to the cluster corresponding to the maximum posterior probability. The selected Gaussian mixture model included 3 clusters. While clusters 2 and 3 exhibited acceptable DOOpost scores on parenchyma (> 0.5), cluster 1 was associated to very low values of DOOpost. The paired comparison of the DOOpre and DOOpost showed a significant improvement of organ dose matching, both in the whole cohort and in clusters 2 and 3. The presented study represents the first attempt to assess a SN procedure specifically tailored to the needs of brain dose VBAs. The DOO scores show the overall accuracy of the devised SN procedure. In addition, the Gaussian mixture model provides a formal criterion for excluding from the statistical analysis the few patients (7%) on whom the SN failed to produce acceptable results. We believe that the implemented framework lays the groundwork for future reliable VBAs in brain irradiation studies.Abstract 214; Table 1Patients (n)DOOWMGMCSFPrePostpPrePostpPrePostpAll (43)0.14 ± 0.110.62 ± 0.18<10-50.16 ± 0.080.60 ± 0.12<10-50.12 ± 0.090.42 ± 0.13<10-5Cluster 1 (3)0.03 ± 0.010.09 ± 0.090.320.19 ± 0.090.36 ± 0.130.100.10 ± 0.040.19 ± 0.060.27Cluster 2 (4)0.41 ± 0.060.62 ± 0.180.050.32 ± 0.030.56 ± 0.110.020.33 ± 0.080.52 ± 0.090.01Cluster 3 (36)0.12 ± 0.060.66 ± 0.11<10-50.14 ± 0.060.63 ± 0.10<10-50.09 ± 0.060.42 ± 0.12<10-5 Open table in a new tab