S2418
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
concurrent chemoradiotherapy. ALCs were measured before and during RT. A nine-compartment physiological model based on Jin et al. [1], representing bone marrow, spleen, lungs, liver, gut, mediastinal lymph nodes, other lymphoid/vessel compartments, non-lymphoid tissues, and circulating blood was implemented to simulate lymphocyte circulation, radiation-induced depletion, and repopulation. Radiation effect followed exponential survival governed by α , while lymphocyte production was modelled by a constant rate (kR). Ten repeated five-fold cross-validations compared two α -estimation strategies: (1) Fixed (cohort-median) α — a single α (median of fitted training values) applied to all validation patients; (2) Patient-specific (PS) α — a fold- specific linear regression trained to predict α from baseline ALC, applied to validation patients. The training-median was used for both. Model accuracy was evaluated by per-patient mean absolute error (MAE) between predicted and observed ALC, averaged over repetitions. Results: Baseline ALC significantly predicted fitted α in all training folds (p < 0.05), with a representative relationship: α = 0.062 + 0.000426 × ALC (R ² = 0.29) (Figure 1). Median fitted α and kR were 0.750 [/Gy] and 0.170 [%/day], respectively. The PS α model yielded lower prediction error than the fixed α model: median (IQR) MAE: 167 (107–245) vs 171 (106–252) [/µL]). The mean Δ MAE (PS – Fixed) was –10 [/µL], and the median Δ MAE was –4.3 [/µL] (IQR –30.5 to 11.9 µL) (Figure 2). Although the difference was modest and did not reach statistical significance (Wilcoxon p = 0.053), the overall trend favored the patient-specific α approach.
Conclusion: The nested CV and feature reduction framework enabled unbiased and reproducible normal tissue complication probability (NTCP) modeling, with TB D45% as a potential RN predictor. This methodology mitigates optimistic bias and provides a solid framework for future radiobiological model development in SRS, warranting prospective validation. Keywords: NTCP A compartment model predicting lymphocyte depletion using patient-specific radiosensitivity Takahiro Kanehira 1 , Fuki Koizumi 2 , Koichi Miyazaki 3 , Taisuke Takayanagi 3 , Norio Katoh 4 , Keiji Kobashi 5 , Takayuki Hashimoto 5 , Hidefumi Aoyama 4 1 Department of Medical Physics, Hokkaido University Hospital, Sapporo, Japan. 2 Department of Radiation Oncology, Hokkaido University Hospital, Sapporo, Japan. 3 Research and Development Group, Hitachi, Ltd., Ibaraki, Japan. 4 Department of Radiation Poster Discussion 811 Oncology, Hokkaido University Faculty of Medicine and Graduate School of Medicine, Sapporo, Japan. 5 Global Center for Biomedical Science and Engineering, Faculty of Medicine, Hokkaido University, Sapporo, Japan Purpose/Objective: Radiation-induced lymphopenia (RIL) has been associated with poor outcomes in non–small cell lung cancer (NSCLC). Jin et al. proposed a physiological compartment model predicting lymphocyte depletion during radiotherapy (RT) by fitting patient-specific radiosensitivity ( α ), revealing inter-patient variability [1]. For prospective prediction, α is often replaced by a cohort-median value, which may be suboptimal. This study tested whether estimating patient-specific (PS) α from pre-treatment information—specifically baseline absolute lymphocyte count (ALC)—improves prediction of ALC dynamics during thoracic RT. Material/Methods: We analyzed 150 NSCLC patients treated with definitive, conventionally fractionated RT or
Made with FlippingBook - Share PDF online