S2471
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Proffered Paper 3822 Multimodal analysis of clinical, radiomic and dosiomic features for prognosticrisk assessment in two large prospective lung cancer trials Stefan M. Fischer 1,2 , Johannes Kiechle 3,2 , Lukas M. Reuter 3,4 , Danai Pletzer 3,4 , Kim M. Kraus 3,5 , Stephanie E. Combs 3,6 , Denise Bernhardt 3,6 , Julia A. Schnabel 1,7 , Jan C. Peeken 3,6 1 School of Computation, Information and Technology, Technical University of Munich, Munich, Germany. 2 Munich Center of Machine Learning, Munich Center of Machine Learning, Munich, Germany. 3 Department of Radiation Oncology, School of Medicine and Klinikum Rechts der Isar, Technical University of Munich, Munich, Germany. 4 Institute of Machine Learning in Biomedical Imaging (IML), Helmholtz Center Munich, Munich, Germany. 5 nstitute of Radiation Medicine (IRM), Helmholtz Center Munich, Munich, Germany. 6 Partner Site Munich and German Cancer Research Center (DKFZ), German Cancer Consortium (DKTK), Munich, Germany. 7 School of Biomedical Engineering and Imaging Sciences, King’s College London, London, United Kingdom Purpose/Objective: Various retrospective studies have analyzed the value of machine learning based prognostic assessment in Non-Small Cell Lung Cancer (NSCLC) patients. In this study, we leverage prospective data from the RTOG 0617 and REQUITE trials to compare the prognostic performance of clinical variables, radiomics, and dosiomics in NSCLC patients. Material/Methods: Prospective data from 680 lung cancer patients were analyzed, including RTOG 0617 (n = 441, training) and REQUITE (n = 239, external validation). The RTOG cohort consisted exclusively of stage III patients treated with 60 or 74 Gy, whereas REQUITE included patients with stages I–III, receiving doses between 45 and 70 Gy.Radiomics and dosiomics features were extracted from pre-treatment CT images, 3D dose distributions, and 3D EQD2 maps, following the IBSI guidelines. Label maps were defined for GTV, PTV, PTV+2 cm, and lungs-GTV. Clinical variables (age, sex, histology, smoking status, chemotherapy, and radiotherapy technique) were included. Highly correlated features (r > 0.5) were removed.Modeling was performed using Random Survival Forests within a five-fold nested cross-validation. For each feature set, the best configuration was retrained on the full training set and validated externally. Performance was quantified using the concordance index (C-index), with statistical significance assessed by one-sided T-tests and log-rank tests versus random guessing.
random prediction, compared to a value of 0.75 in the training dataset. The training and validation cohorts showed similar demographics (Table 1), and similar distributions of linear predictors (25th–75th percentiles: training -0.66–0.72; validation -0.61–0.63).
Conclusion: The model showed poor performance with the validation dataset, and had a considerably lower C- Harrell index, likely due to overfitting in the small training dataset. The low discriminating power limits the model’s clinical value for patient stratification for personalized local treatment. However, the slight separation of risk groups suggests that DWI retains prognostic potential, encouraging further exploration and model training in larger cohorts to mitigate overfitting. References: [1] Bisgaard ALH, Brink C, Tine S, et al. Front Oncol 2024;14:1401464. https://doi.org/10.3389/fonc.2024.1401464.[2] Rahbek S, Mahmood F, Tomaszewski MR, et al. Phys Med Biol 2023;68. https://doi.org/10.1088/1361- 6560/acaa85.[3] Royston P, Altman DGA. BMC Res Methodol 2013;13. https://doi.org/10.1186/1471- 2288-13-33 Keywords: Diffusion-weighted MRI, Imaging biomarker, LAPC
Made with FlippingBook - Share PDF online