S1465
Interdisciplinary - Other
ESTRO 2026
1 Radiation Oncology & Institute for Data Science in Oncology, MD Anderson Cancer Center, Houston, USA. 2 Integrated Mathematical Oncology Department, Moffitt Cancer Center, Tampa Bay, USA. 3 Department of Radiation Oncology, Inselspital, Bern University Hospital and University of Bern, Bern, Switzerland Purpose/Objective: Accurate modeling of tumor and biomarker dynamics is critical for predicting responses to radiotherapy and immunotherapy. Yet clinical time-series data are often incomplete, left-censored (true progression precedes its measurement), or right-censored (progression not observed within follow-up). These gaps hinder mechanistic and data-driven modeling. We developed a unified Generative Adversarial Network (GAN)– Diffusion framework that learns temporal correlations from incomplete trajectories, transforming them into Gramian Angular Fields (GAFs), enabling robust reconstruction and biologically consistent imputation of oncologic time series. Material/Methods: A class-conditional half-GAF GAN was implemented to synthesize the upper-triangular GAF segment from latent noise (z) and class label (y) corresponding to tumor-growth model type (exponential, logistic, Gompertz). The discriminator combined an adversarial and an auxiliary-classification head and was trained using binary and categorical cross-entropy losses. A frozen inverter Ψ (GAF → time series → GAF) enforced weak cycle-consistency and temporal reversibility. Training employed 30,000 synthetic tumor-growth trajectories (10,000 per model) corrupted with Gaussian noise (σ = 0.05–0.2) and 10–50 % missingness patterns (random, clustered, and left/right-censored). Validation used real-world prostate-specific-antigen (PSA) trajectories from 70 patients receiving androgen-deprivation therapy, demonstrating clinical applicability. Five-fold cross- validation quantified generalization. Reconstruction performance was assessed by R², and hit-rate (±15 % error tolerance).
82% were obtained for CTVT1, CTVT2 and CTVN, respectively. For rectum, bladder and bowel bag, the corresponding median DSC values were 88%, 90% and 86%. For patient-2, median DSC values of 87%, 80% and 95% were observed for the CTVT, rectum and bladder, respectively. Considerable delineation discrepancies were identified, particularly at the cranial edge of the CTVN and at anatomical boundaries such as the cranial limit of the rectum (sigmoid take-off) and the prostatic apex. These variations led to reduction of D98% of the PTVT1, PTVT2 and PTVN of up to 7%, 25% and 37%, relative to the clinical objective for patient-1. The dose-volume constraints for the OIs were exceeded, particularly for the rectum and bowel bag. A summary of the results is presented in Table 1.
Conclusion: Substantial inter-observer variability in structure delineation was observed among the ROs, resulting in inadequate target dose coverage and excessive OI doses beyond clinical limits. These findings highlight the need for harmonization and continuous quality assurance in prostate cancer radiotherapy contouring. Regular contouring audits or dummy-runs are essential to ensure consistent delineation practices over time. Keywords: Prostate cancer, delineation variation, dummy-run Digital Poster 4918 Ab Initio Generative Framework for Time-Series Imputation in Radiotherapy Modeling Nahum Puebla 1 , Barnett Barnett 2 , Pirmin Schlicke 1 , Mohammad Zahid 1 , Sarah Brüningk 3 , Heiko Enderling 1
Results:
Made with FlippingBook - Share PDF online