ESTRO 2025 - Abstract Book

S3426

Physics - Machine learning models and clinical applications

ESTRO 2025

Figure 2: Comparison of DVH of the original and the reconstructed plans for one brain irradiation recalculated in Monaco TPS (left panel) and VERIQA (right panel) Conclusion: The proposed model architecture's flexibility allows it to adapt to other delivery techniques and measurement modalities. Its utilization also reduces the requirements of the measurement devices for PSQA. The proposed unified framework could play a decisive role in automating QA workflow, especially in the context of real-time adaptive radiation therapy.

Keywords: PSQA, Deep learning, Detector arrays

3600

Digital Poster Automatic identification of trade-offs in head and neck cancer radiotherapy treatment planning Sanni Tähtinen, Hannu Laaksonen, Elena Czeizler Treatment Planning Technology Department, Varian Medical Systems, a Siemens Healthineers Company, Helsinki, Finland Purpose/Objective: Head-and-neck radiotherapy (RT) cases include a lot of variation in the number of targets, their prescription levels, location, size, and shape. This heterogeneity leads to a wide variation of trade-offs in RT plans. We developed an AI based automated method which can identify the key trade-off types. Then, we developed a classifier to predict the most likely trade-off, starting from anatomical data only. Material/Methods: The dataset includes 356 head-and-neck cancer patient cases treated with Volumetric Modulated Arc Therapy (VMAT). Clustering techniques were applied to Dose-Volume Histogram (DVH) data to automatically identify the clinically relevant trade-off categories. Clustering algorithms, including K-means, hierarchical clustering, and biclustering, were explored, all producing meaningful patient subgroups. These clusters served as labels to train a convolutional neural network (CNN) classifier to predict the trade-off group for new patients based on patient anatomy. The CNN model was based on the ResNet50 [1] architecture from MONAI [2] (Medical Open Network for AI) with three input channels for CT images, OAR masks, and PTV masks. The dataset was split into 200 for training, 67 for validation, and 89 for testing.

Results: A hierarchical clustering approach revealed the following trade-off patterns (Figure 1): • Cluster 1, low doses to all OARs.