S2456
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
cell carcinoma (HNSCC) represents a global health challenge. Radio(chemo)therapy (RCHT) is an effective primary treatment in many cases, and FDG-PET radiomics shows considerable potential for predicting clinical outcomes following RCHT [1]. However, different studies use different definitions of the region of interest (ROI) for radiomic feature extraction, and there is limited evidence on how this choice impacts predictive model performance. Common ROI definitions include clinically used, anatomy-based gross tumor volume (GTV) and PET-based segmentation of the metabolically active tumor volume (MTV). In this study, we evaluated the performance of PET-based outcome prediction models depending on the different ROI definitions. Material/Methods: The study cohort consisted of 739 patients with HNSCC who received primary RCHT at seven international Departments for Radiotherapy and Radiation Oncology. Two segmentations were generated either within clinical routine (GTV) or using a semi-automated delineation tool (MTV). From each segmentation, 182 FDG-PET–based radiomic features, compliant with the Image Biomarker Standardisation Initiative (IBSI) [2], were extracted and used in nine machine-learning pipelines – reducing the influence of pipeline selection on performance estimates. Model performance was evaluated using cross-validation (N=534) and an independent test set (N=205), with median C-indices and confidence intervals computed from identical bootstrap resamples. Patient stratification was performed using an ensemble of the nine models and evaluated with the log-rank test. Additionally, each radiomic model was combined with a clinical model to quantify the incremental value of PET features. Finally, the segmentation approaches were compared using the previously mentioned performance metrics. Results: Overall, the MTV-based models consistently showed higher performance in both cross-validation and the independent test set (Table 1). Similarly, the strongest combined models were obtained by integrating clinical models with MTV-based models (Table 1). For patient stratification, only the MTV-based model ensembles successfully distinguished the predefined test set risk groups for both outcomes (Fig. 1).
Conclusion: Our findings indicate that semi-automatic MTV segmentations are non-inferior to manual CT-based GTV segmentations for predicting HNSCC radio(chemo)therapy outcomes, and may even provide additional benefit in certain scenarios. Given the relatively simple segmentation of MTV (i.e., selecting PET voxels above an intensity threshold), these results may support the broader adoption of FDG-PET-based segmentation approaches in future HNSCC PET radiomics studies and may contribute to methodological standardization in the field. References: [1] Philip, Mahima Merin et al. “A systematic review and meta-analysis of predictive and prognostic models for outcome prediction using positron emission tomography radiomics in head and neck squamous cell carcinoma patients.” Cancer medicine vol. 12,15 (2023): 16181-16194. doi:10.1002/cam4.6278[2] Zwanenburg, Alex et al. “The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping.” Radiology vol. 295,2 (2020): 328-338. doi:10.1148/radiol.2020191145 Keywords: radiomics, machine learning, harmonization Diffusion-MRI hypoxia mapping with distortion correction in head and neck cancer radiotherapy Minoo Gandomi 1,2 , Patrik Brynolfsson 1,2 , Maria Gebre- Medhin 3,4 , Minna Lerner 1,2 , Emelie Lind 5,6 , Jenny Gorgisyan 2,6 , Lars E Olsson 1,2 1 Department of Translational Medicine, Medical Radiation Physics, Lund University, Malmö, Sweden. 2 Radiation Physics, Department of Hematology, Oncology, and Radiation Physics, Skåne University Hospital, Lund, Sweden. 3 Oncology, Department of Hematology, Oncology, and Radiation Physics, Skåne University Hospital, Lund, Sweden. 4 Oncology, Digital Poster Highlight 3021
Made with FlippingBook - Share PDF online