ESTRO 2025 - Abstract Book

S3427

Physics - Machine learning models and clinical applications

ESTRO 2025

• Cluster 2, balanced doses to the paired structures (parotids, submandibular glands) but higher doses in the submandibular glands compared with larynx and trachea. • Cluster 3, elevated doses for the left submandibular gland. • Cluster 4, elevated doses for the right submandibular gland. • Cluster 5, balanced doses to the paired structures but higher doses in the larynx and trachea compared with the submandibular glands. Complementary to Cluster 2. Using these clusters, we trained a CNN classifier to predict the expected trade-off. The model accuracy was evaluated on the test set, achieving 91% accuracy, with most misclassifications indicating borderline cases, see Figure 2. Conclusion: We implemented an automated method to identify the key trade-offs encountered in the population of head and neck cancer cases. Out method is general and can be applied to any cancer site. Using the labels generated by these clusters, we trained a classifier to predict the expected trade-off. By identifying these likely compromises in advance, the system accelerates the planning process, which may enable clinicians to reach high-quality treatment decisions more efficiently.

Figure 1: Average DVHs for key OARs across five clusters.