ESTRO 2025 - Abstract Book

S3367

Physics - Machine learning models and clinical applications

ESTRO 2025

tumors (C05) have a 53% oropharynx-model weight and a 47% oral cavity-model weight, consistent with its anatomical location between these regions. Unspecified oropharynx subsites (C10) combine 53% oropharynx, 20% oral cavity, and 27% hypopharynx components. For comparison, independent models were trained for oral cavity (C02 – C06), oropharynx ((C05),C01,C09,C10), and hypopharynx (C12,C13). For subsites mainly assigned to one model component, such as base of the tongue (C09), predictions from the mixture and independent models were similar. However, for subsites located in-between, the mixture model provided greater accuracy. For example, for palatine tumors (C05), LNL II involvement was observed in 48% (29/61) of cases. The mixture model predicts 46%, compared to 32% by the independent oral cavity-model and 72% by the oropharynx-model where C05 is included. For C10, observed LNL II involvement was 62% (106/169), closely matched by the mixture model's 65% prediction, as opposed to the oropharynx model's 76% without C05.

Conclusion: The mixture model is a natural approach to include different subsites into a single unified model of lymphatic spread. It captures the gradual change in the lymph node involvements between ICD codes, and thereby facilitates accounting for tumor subsite in estimating risk of occult metastases.

Keywords: Hidden Markov Model, Head and Neck, elective CTV

References: 1. J. Biau et al. “Selection of Lymph Node Target Volumes for Definitive Head and Neck Radiation Therapy: A 2019 Update”. Radiotherapy and Oncology 134 (May 2019), pp. 1– 9. doi: 10.1016/j.radonc.2019.01.018. 2. R. Ludwig et al. “A Hidden Markov Model for Lymphatic Tumor Progression in the Head and Neck”. Scientific Reports 11.1 (Dec. 2021), p. 12261. doi: 10.1038/ s41598-021-91544-1.