Abstract

Roman Ludwig¹, Bertrand Pouymayou¹, Panagiotis Balermpas¹, Jan Unkelbach¹

¹University Hospital Zurich, Radiation Oncology, Zurich, Switzerland

Currently, elective CTV definition in head & neck cancer is mostly based on the empiric prevalence of lymph node involvement for a given primary tumor location. However, an individual patient’s risk of harboring microscopic metastases in each lymph node level (LNL) varies depending on their T-stage and findings of macroscopic metastases in LNLs through imaging. We propose a probabilistic model for lymphatic metastatic spread that can quantify this risk of microscopic involvement based on the individual patient's state of cancer progression, which may allow for personalized CTV-N definition.

A patient’s state of disease is modelled via one hidden binary random variable for each LNL, which indicates if the LNL harbors tumor including occult metastases. This hidden state of a LNL is connected via sensitivity and specificity to an observed binary random variable that indicates if metastases are seen on imaging. Over one-time step, tumor cells may spread from the primary tumor to an LNL, or between LNLs, with some probability rate. Formally, this is described by a hidden Markov model (HMM). The directed arcs of the HMM's graph reflect the direction of lymph flow and spread probability rates are learned from a dataset of patients in whom involvement of each LNL was reported (Figure 1). T-stage can be incorporated into the model by assuming that on average late T-stage tumors had more time to progress than early T-stage tumors. We demonstrate the HMM model for ipsilateral spread of oropharyngeal head & neck squamous cell carcinoma, trained via MCMC sampling using a dataset reconstructed from [1], and under the assumption that the portion of N0 patients is 30% for early T-stage patients and 20% for late T-stage.

Figure 1:

[1] Sanguineti (2009) Int. J. Rad. Onc. Biol. Phys. 74(5) 1356–1364

Figure 2 shows the risk of occult metastases in LNL III and IV depending on T-stage and whether the upstream LNLs II and III harbor macroscopic metastases. For early T-stage, the expected risk for microscopic involvement of LNL III rises from 4.8 ± 1.0 % for N0 patients to 10.2 ± 1.8 % when macroscopic metastases in LNL II are observed. For late T-stage and macroscopic metastases in LNL II, the risk increases further to 13.7 ± 2.8 % (Figure 2 top). Analogous findings are observed for level IV (Figure 2 bottom).

Figure 2:

HMMs provide a statistical framework to model the spread of lymphatic metastases. It allows us to estimate the risk of microscopic involvement of LNLs depending on T-stage and location of macroscopic metastases. But larger datasets of detailed involvement patterns, rather than reports of only prevalence, are needed before the model may inform guidelines for CTV-N definition.

Acknowledgement:
This work was supported by the Clinical Research Priority Program Artificial Intelligence in Oncological Imaging of the University of Zurich.