S2462
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Digital Poster Highlight 3351
Daily MR-linac ADC measurements enable patient- specific uncertainty estimation of ADC changes during chemoradiation in esophageal cancer Koen M. Kuijer, Astrid L.H.M.W. van Lier, Lieke T.C. Meijers, Stella Mook, Gert J. Meijer Department of Radiotherapy, University Medical Center Utrecht, Utrecht, Netherlands Purpose/Objective: Accurate and reliable assessment of response to neoadjuvant chemoradiation could enable organ preservation in esophageal cancer. The apparent diffusion coefficient (ADC) is a promising biomarker, as changes in ADC have been associated with treatment response [1]. However, ADC measurements are subject to high uncertainty [2], leading to large uncertainty in ADC changes as these are typically based on only two time points. The extent to which this uncertainty varies between patients is unknown. Patient-specific information about the uncertainty could improve the interpretation of ADC changes. This study investigated whether daily ADC measurements enabled by MR-linac systems allow quantification of patient-specific uncertainty in estimated ADC changes to improve their interpretability in individual patients. Material/Methods: Seventeen esophageal cancer patients (10 receiving 23 fractions, 7 receiving 28 fractions) underwent chemoradiation on a 1.5 T MR-linac. Diffusion- weighted scans were acquired daily, except once a week. ADC maps were calculated using a mono- exponential fit of b-values 150 and 500 s/mm2. A volume-of-interest was manually delineated on each b500 image, and the median ADC was extracted.For each patient, the ADC change between the first and 23rd fraction was derived using linear regression across all measurements. Patient-specific uncertainty in the estimated ADC change was deduced from the variance of the predicted ADC in the linear regression model (see Equation below), where the variances and the covariance of the intercept ( β ₀ ) and slope ( β ₁ ) directly follow from the residuals and measurement distribution. Finally, potential determinants of uncertainty in ADC changes were explored. Results: The median ADC was successfully acquired for 312 fractions (15-22 per patient), with trends for two patients shown in Figure 1. Linear regression-derived ADC changes and their uncertainty, expressed as the interquartile range (IQR) of the estimated change, are presented in Figure 2. The uncertainty varied
Figure 2: Cross-validated Kaplan-Meier curves from the clinical model (A) and the clinical + PET ensemble model (B). Conclusion: In this study, we demonstrate the effectiveness of a 3D deep learning framework using pretreatment FDG-PET imaging for generating clinically useful predictive models for CCRT resistance in LACC. Future work will externally validate our ensemble model and investigate the explainability of its 3D CNN component’s risk score. References: 1. Gennigens, Christine, et al. "Optimal treatment in locally advanced cervical cancer." Expert Review of Anticancer Therapy 21.6 (2021): 657-671.2. Chung, Hyun, et al. "Pembrolizumab treatment of advanced cervical cancer: updated results from the phase II KEYNOTE-158 study." Gynecologic Oncology 162 (2021): S27.3. Gadducci, Angiolo, and Stefania Cosio. "Neoadjuvant chemotherapy in locally advanced cervical cancer: review of the literature and perspectives of clinical research." Anticancer research 40.9 (2020): 4819-4828.4. Simon, Richard M., et al. "Using cross-validation to evaluate predictive accuracy of survival risk classifiers based on high-dimensional data." Briefings in bioinformatics 12.3 (2011): 203-214. Keywords: cervical cancer, FDG-PET, deep learning
Made with FlippingBook - Share PDF online