S2470
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Digital Poster Highlight 3774
with an ICC>0.90, 0.90>ICC>0.75, 0.75>ICC>0.5, and ICC<0.50, respectively. The reproducibility was assessed using different numbers of contours. Results:
Internal validation of a diffusion-MRI-based survival prediction model in pancreatic cancer: Results from a large validation cohort Anne L. H. Bisgaard 1 , Pia Braagaard Hartfelt 2 , Carsten Brink 1,2 , Tine Schytte 2,3 , Rana Bahij 3 , Mathilde Weisz Ejlsmark 2,3 , Uffe Bernchou 1,2 , Anders S. Bertelsen 1 , Per Pfeiffer 2,3 , Faisal Mahmood 1,2 1 Laboratory of Radiation Physics, Department of Oncology, Odense University Hospital, Odense, Denmark. 2 Department of Clinical Research, University of Southern Denmark, Odense, Denmark. 3 Department of Oncology, Odense University Hospital, Odense, Denmark Purpose/Objective: Stereotactic body radiation therapy (SBRT) is used to treat locally advanced pancreatic cancer (LAPC), however, varying outcomes indicate a need for personalized radiotherapy. A previous study indicated that diffusion MRI (DWI) derived parameters had prognostic value in pancreatic cancer, showing ability to divide patients into risk groups of overall survival [1]. The aim of this study was to perform an internal validation of the developed model in an independent cohort of patients with LAPC treated with SBRT. Material/Methods: The existing model was a Cox proportional hazards model with overall survival as the endpoint, developed using a training dataset of 45 patients with five longitudinal DWI scans each. Parameter selection using bootstrap-based best-subset selection among both clinical and DWI-derived parameters resulted in a model with two parameters, both derived from DWI using monotonous slope non-negative matrix factorization [2]. Validation was performed using a validation dataset consisting of data from 101 patients with primary or locally recurrent pancreatic cancer treated with a similar treatment regimen to the training set (5 fractions of 10 Gy on a 1.5 T MRI-linac), with DWI acquired at each fraction. Model coefficients were applied to the validation dataset to calculate the linear predictors. Calibration plots were generated using the same baseline hazard function and the same cut-off values of linear predictors as in the training cohort (–0.66, 0.72) [3]. The model’s discriminative performance was assessed using Harrell’s C-index. Results: Model–data agreement was substantially lower in the validation cohort compared to the training cohort, although a weak agreement remained for the medium- and high-risk groups (Figure 1). A modest separation of the high-risk group from the other groups within the first 20 months suggests a limited but relevant prognostic value. Harrell’s C-index was 0.56 in the validation dataset, slightly better than a
The contours generated using the Osorio approach had more features with good/excellent reproducibility compared with the observer contours. The contours simulated using the max/min method had more features with moderate/poor reproducibility compared with the observer contours. For both sets of simulated contours, when using four contours or more to assess the reproducibility, variation in the number of features with good/excellent was minimal. Using different combinations of the observers’ contours led to different numbers of features with good/excellent reproducibility. For combination 1 (C1), variation between number of features with good/excellent features with the additional observers after four observers is minimal, however for combination 2 (C2), the number of features increases. Conclusion: Neither method for simulating observer contours generated the necessary variation to identify the same features to be reproducible as the real observer contours. Further work is required in determining a method to simulate contours for assessment of radiomic feature reproducibility. This may include further investigation into the use of a probabilistic autosegmentation tool. References: [1] R. Brown et al., ‘Impact of number of observers on reproducibility of gynaecological cancer MRI radiomics to interobserver contour variation’, Med. Phys., vol. 52, no. 11, p. e70085, 2025, doi: 10.1002/mp.70085.[2] D. Bernstein et al., ‘An Inter-observer Study to Determine Radiotherapy Planning Target Volumes for Recurrent Gynaecological Cancer Comparing Magnetic Resonance Imaging Only With Computed Tomography-Magnetic Resonance Imaging’, Clin. Oncol., vol. 33, no. 5, pp. 307–313, May 2021, doi: 10.1016/j.clon.2021.02.003.[3] E. V. Osorio, J. Shortall, J. Robbins, and M. van Herk, ‘Contour Generation with Realistic Inter-observer Variation’, Apr. 21, 2022, arXiv: arXiv:2204.10098. doi: 10.48550/arXiv.2204.10098. Keywords: Reproducibility, MRI, interobserver variation
Made with FlippingBook - Share PDF online