S2436
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Conclusion: Xerostomia was significantly associated with the mean dose to the parotid glands, submandibular glands, and oral cavity, with partial recovery over time. The model exhibited robust calibration and moderate discriminative ability, capturing toxicity grades, dose– responses and time-dependent patterns following radiotherapy. The observed between-patient variability highlights the value of including random effects in the model. References: 1. Agresti, A. (2010). Analysis of Ordinal Categorical Data (2nd Edition). Wiley.2. Van Calster B, Van Belle V, Vergouwe Y, Timmerman D, Van Huffel S, Steyerberg EW. Extending the c-statistic to nominal polytomous outcomes: the Polytomous Discrimination Index. Stat Med. 2012;31(23):2610-2626. doi:10.1002/sim.5321 Keywords: Ordinal logistic regression, Xerostomia external validation of a multimodal AI-based model for prostate cancer radiation therapy outcome prediction from baseline MRI Benjamin D Simon 1,2 , Avani D Rao 3 , Stephanie A Harmon 2 , Lei Clifton 4 , Anshul Thakur 1 , Simon Berents 5 , Krishnan R Patel 6 , Amar Kishan 7 , Tommy Jiang 7 , Matthias Ojo 8 , Peter L Choyke 2 , Luca F Valle 7 , David A Clifton 1 , Deborah E Citrin 6 , Baris Turkbey 2 1 Department of Engineering Science, University of Oxford, Oxford, United Kingdom. 2 Molecular Imaging Branch, NCI, NIH, Bethesda, USA. 3 Inova Schar Cancer Institute, Inova Health System, Fairfax, USA. 4 Nuffield Department of Primary Care Health Sciences, University of Oxford, Oxford, United Kingdom. 5 College of Science, Clemson University, Clemson, USA. 6 Radiation Oncology Branch, NCI, NIH, Bethesda, USA. 7 Radiation Oncology, UCLA, Los Angeles, USA. 8 Radiation Oncology, Charles R. Drew University of Medicine and Science, Los Angeles, USA Purpose/Objective: Prostate cancer (PCa) risk prognostication methods such as NCCN risk groups and Gleason grading have limited reproducibility for prognosis in a given patient (1,2). As a result, physicians are limited in their ability to evaluate risk. We demonstrate the development of an automated deep learning pipeline using radiomics from MRI across two centers to predict biochemical failure (BF) of PCa for patients receiving radiation therapy (RT) with or without ADT and validate the model in a third external cohort. Material/Methods: Patients treated between 2005-2024 with RT ± ADT from three centers were aggregated. Two centers were selected for training (n=151, n=68) and the third Digital Poster 2088
the total variance was attributable to between-patient differences (intraclass correlation coefficient (ICC) = 0.39 [1]). The calibration curve (figure 2) showed good agreement between predicted and observed outcomes. Overall discrimination was moderate (PDI = 0.63, 0.62–0.63) and varied across grades: 0.72 (0.71– 0.73), 0.55 (0.53–0.56), 0.60 (0.58–0.61) and 0.64 (0.61– 0.67) for grades 0-3, respectively.
Made with FlippingBook - Share PDF online