ESTRO 2025 - Abstract Book

S3754

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

and assessed the impact of PET image feedback timing and Human Papillomavirus (HPV) status on prediction accuracy.

Material/Methods: Serial FDG-PET/CT images were acquired from 50 HNSCC patients (35 HPV+, 11 HPV-, 4 unknown) at pretreatment (Pre-Tx) and during 2 nd and 4 th weeks of standard chemo-radiotherapy (35´2Gy). The tumor voxel DRM representing the spatial distribution of the average metabolic activity change ratio throughout the treatment course was constructed following deformable PET/CT image registration. 1 The reference DRM value for each voxel v , DRM ref ( v ), was determined voxel-by-voxel using linear regression on the logarithm of SUV i /SUV 0 (i = 2 and 4), with SUV i being the standard uptake values (SUVs) obtained at the i-th treatment week. A 3D residual U-net DL model with 3.4 million trainable parameters was employed. The DL models were trained to predict the DRM ref using various combinations of input images: SUV 0 alone, SUV 0 and SUV 2 , and SUV 0 with SUV 4 , resulting in predicted DRMs denoted as DRM 0 , DRM 2 , and DRM 4 , respectively. Model performance was evaluated using 10-fold cross-validation. The impact of PET image feedback timing and HPV status on the prediction accuracy was assessed. Results: Across all 34,612 tumor voxels, the mean (±SD) of DRM ref was 0.46 (0.2). The mean deviations within individual tumors were -0.042±0.103 for DRM 0 , -0.024±0.075 for DRM 2 , and -0.009±0.039 for DRM 4 . Detailed prediction deviations for resistant tumor voxels (DRM ref >0.6, accounting for 23.7% of all tumor voxels) and sensitive tumor voxels (DRM ref ≤0.6) in both HPV- and HPV+ tumors are provided in the Table.

Conclusion: The DL model can predict tumor voxel dose response using single PET feedback images acquired during the early weeks of treatment for HNSCC patients. For resistant tumor voxels, delaying PET image feedback improved dose response prediction accuracy, particularly in the HPV- group. For sensitive tumor voxels, the dose-response prediction was less dependent on the timing of PET image feedback across both HPV status groups. These findings offer valuable insights for the effective management of adaptive treatment for HNSCC.

Keywords: FDG-PET, Adaptive Radiotherapy, Deep Learning

References: 1 Yan, Di, et al. "Tumor voxel dose-response matrix and dose prescription function derived using 18F-FDG PET/CT images for adaptive dose painting by number." International Journal of Radiation Oncology* Biology* Physics 104.1 (2019): 207-218.