ESTRO 2024 - Abstract Book

S4416

Physics - Machine learning models and clinical applications

ESTRO 2024

We extended a previously developed hidden Markov model (HMM) 2 to better reflect the nature of lymph node metastasis diagnosis. In the extended model, each LNL is described by a trinary state: healthy, microscopically involved, and macroscopically involved. Lymphatic spread of the tumor is represented by a directed graph (Figure 1) where edges represent lymphatic drainage pathways that are parametrized with probabilities for the primary tumor to spread to a LNL (b-parameters) and from one LNL to the next (t-parameters). Spread to a LNL initially results in microscopic involvement of the level. We then introduce a growth parameter (g-parameter), which describes the probability for a LNL to transition from microscopic involvement to the clinically diagnosable macroscopic state. It is assumed that pathology after neck dissection detects lymph node metastases with sensitivity and specificity of one but does not distinguish microscopic and macroscopic involvement. Imaging is assumed to detect macroscopic involvement with sensitivity one but does not detect microscopic involvement. We trained the model on a dataset of 264 oropharyngeal SCC patients from Bauwens et al 3 , where both clinical and pathological data were available. The model's predicting of the risk of occult metastases (corresponding to probability of the microscopic involvement state) is compared to the prevalence in the dataset. Figure 1

Results:

Figure 2 shows the risk of microscopic involvement in ipsilateral LNLs III and IV for three common diagnoses. The observed prevalences of microscopic involvement are depicted with a solid line, whereas the computed risk of involvement is shown by a histogram. The distributions represent the uncertainty in both data and the model. Figure 2a) considers patients without clinically diagnosed metastases and early T-categories (T1 & T2), showing a risk of microscopic metastases in LNL III around 15% and below 3% in LNL IV. In Figure 2b) LNL II was diagnosed as clinically positive. Consequently, the risk for LNL III involvement increases, while LNL IV is still comparably low around 5%. Figure 2c) shows the case of a positive clinical diagnose for LNLs II and III and late T-category (T3 & T4) where we now observe an increased risk and prevalence for LNL IV involvement around 20%.