Abstract Book

S237

ESTRO 37

instance by analyzing DVH by intra organ segments, it has been shown that the dose on some subregions correlated better with toxicity (Stenmark et al. 2014; Ebert et al. 2014). Some other models have sought to geometrically describe spatial dose distribution on the organ wall via dose surface maps (DSM) and found correlations between geometric features and toxicity events (Buettner et al. 2009; 2011) or performed pixel-wise comparisons (Palorini et al. 2016) to show differences across a population. More recently, approaches with in-depth analysis of 3D dose distribution using voxel-wise population analysis techniques have emerged (Drean et al. 2016; Acosta et al. 2012). These techniques aim to analyze and classify the dose-effect relationship in each voxel across a population in a common coordinate system with the objective of identifying anatomical regions that may be responsible for the toxicity event. Voxel-based methods face nevertheless several challenges and require different preprocessing steps to be anatomically meaningful: (i) the mapping of a population of individuals in terms of their anatomy to a common coordinate system; (ii) the propagation of dose distributions according to the anatomical transformation thus obtained; (iii) a reliable methodology to perform local statistical analysis of dose-effect relationship. The dose mapping may rely on a parametric representation of the anatomy in a spherical coordinate system (Heemsbergen et al. 2010; Witte et al. 2010) or can be more precisely obtained through non-rigid registration (NRR) (Drean et al 2016), which appears as particularly difficult given the very high interindividual anatomical variability (organ volume, presence of gas, air, etc.) and low contrast of soft tissues. The selection of an adequate template and the validation of registration are fundamental tasks which precede any voxel-wise analysis. In this work, we present a generic workflow (Figure 1) which relies on NRR allowing the identification of generic subregions at risk which resulted to be more predictive when applied to the rectal toxicity use case. Anatomical mapping methods that employs 3D structural models of the organs which help steering registration by exploiting the delineations from the planning CT scans are introduced (Drean et al. 2016; Acosta et al 2017). To cope with the interindividual variability, several templates spanning different anatomies may be used. The process is then followed by voxel-wise comparisons or classification thereby producing a predictive model. This last step may include for example machine learning techniques. Eventually, the acquired knowledge from a population data may help identifying sub organs regions to be spared during the planning. Some of the limits of voxel-based studies involve not only the validation of the registration but also the reproducibility on larger and external cohorts. Further validation for application in different locations is required. Besides, the nature of deformable organs whose planning dose is not representative of the real delivered dose may appear as an issue as well as the interindividual radiosensibility which recall the importance of bringing complimentary data to the models (multimodal imaging, clinical and biological data or in silico simulations). As any relationship between local dose and toxicity does not necessarily means causality, further investigation is needed to understand the involved physio pathological process. Finally, this predictive approach doesn't include non-dosimetric parameters such as patient history or individual radiosensitivity, which may also impact on the risk of toxicity. As the way is open for further investigation, we discuss also the perspectives in terms of the use of generic voxel- based models for knowledge-based patient-specific planning helping to devise tailored radiation treatments.

Figure 1. Population voxel-based model based on Non- Rigid Registration and mapping of 3D planning dose distributions allowing the identification of sub regions at risk. SP-0459 Issues in pixel/voxel wise dose maps comparison: significance and robustness F. Palorini 1 1 Fondazione IRCCS Istituto Nazionale dei Tumori, Prostate Cancer Program, Milan, Italy Abstract text The search of spatial dose effects is an important emerging field. It is well recognised that common dose descriptors, such as DVHs, still have a pivotal role in treatment plan optimisation but suffer the lack of spatial information. Various methods aimed at analysing local effects are being proposed. Here we will concentrate on voxel-wise analyses of dose distributions, highlighting some critical aspects of dose map building and comparison. We will discuss what kind of information can be really extracted from pixel/voxel wise analysis. And, finally, we will discuss the multiple testing problem, trying to understand if, and when, it really matters in in this context. SP-0460 Evidence for spatially-based bladder dose constraints N.A.B. Yahya 1 , M. Ebert 2 1 The National University of Malaysia, Faculty of Health Sciences, Kuala Lumpur, Malaysia 2 The University of Western Australia, School of Physics, Nedlands, Australia Abstract text Dose maps overcome the limitation of dose-volume histograms that ignore the spatial distribution of the dose. This allows dose-toxicity associations in specific anatomic regions to be explicitly assessed. In the realm of urinary toxicity, the use of dose maps has only lately received more attention. In this presentation, first, we will review the evidence of spatially-variable dose-toxicity associations of the bladder. To date, the evidence is largely focusing on the importance of dose to the bladder trigone, an area distinct in the bladder. Second, improvements of the analyses for dose-toxicity associations will be reviewed based on RADAR dataset. This includes the inclusion of important clinical features as covariates, simplification of multiple comparison adjustment and separation of endpoints based on the potential endpoint architecture. Third, we will discuss on potential ways for the dose maps to be incorporated into treatment planning as spatially-variable dose constraints and critically assess whether these dose constraints may offer a significant advantage over simpler dose-surface/volume histograms.