Abstract Book

S1089

ESTRO 37

with population-based statistics on the lymphatic progression patterns. Bayesian networks (BN) are well suited to quantitatively describe this problem. A BN consists of nodes (which represent random variables) and edges (which represent probabilistic relationships). We developed a BN model in which each LN level is associated with two binary nodes (Figure 1). One corresponds to the detected macroscopic state (e.g. visible LN metastases on imaging) while the second corresponds to the microscopic involvement and cannot be observed. The edge linking the two nodes models the detection probabilities and is used to infer the microscopic state. PT sites are modeled by additional binary nodes. The edges connecting the LN levels and the PT nodes represent the probability for tumor cells to spread from one site to another. The graph structure and the model parameters are chosen based on current guidelines [2] and to match the risks of microscopic involvement reported in [1].

Conclusion Continuous IBDM was more sensitive to dose-response relationships than binary IBDM with a dichotomised endpoint, with binary IBDM strongly influenced by the choice of threshold. Continuous IBDM allows anatomy with significant dose-response relationships to be spatially localised, and improved understanding of these dose-response relationships will help optimise radiotherapy. We are currently applying this technique to study trismus in head and neck cancer patients. EP-2001 A Bayesian network model for personalized elective CTV definition in head and neck cancer B. Pouymayou 1 , O. Riesterer 1 , M. Guckenberger 1 , J. Unkelbach 1 1 University Hospital Zurich and University of Zurich, Department of Radiation Oncology, Zürich, Switzerland Purpose or Objective Definition of the clinical target volume (CTV) is one of the largest sources of uncertainties in radiotherapy. In the case of head & neck cancer, the CTV contains lymph node (LN) levels that are at risk of harboring microscopic metastases despite negative findings on imaging. Thereby, a large portion of the neck is irradiated prophylactically, adding to treatment-related toxicity. Currently, population-based guidelines are being used to determine the LN levels to be included in the CTV, which typically only incorporate the site of the primary tumor (PT) (e.g. oropharynx, hypopharynx or larynx) and N- stage. We present a statistical model to estimate the probability of microscopic involvement of LN levels based on the individual patient's state of lymphatic tumor progression. Material and Methods To estimate the probability of microscopic involvement of LN levels, patient specific characteristics (site of the PT and macroscopic LN metastases) have to be combined

Results Once the BN is defined, it can be applied to personalize CTV definition for a newly diagnosed patient. To that end, an inference algorithm retrieves the probabilities of the microscopic state, given the observed macroscopic state. In Figure 1, we illustrate this for an oropharynx tumor patient in whom levels II and III harbor LN metastases. The BN model predicts that level IV is microscopically involved with a probability of 11%. An advantage of the BN model is that it describes how the risk of microscopic involvement depends on the macroscopic progression (Table 1). For example, when level III has no observed metastases, the involvement probability of level IV drops to 5%.