S181
Clinical - Biomarkers of clinical response
ESTRO 2026
Digital Poster 462 Daily MRI-based bone marrow fat fraction quantification as a surrogate marker for radiation- induced lymphopenia Luka C. Liebrand 1,2 , Ben Neijndorff 1 , Vivian W.J. van Pelt 1 , Marijn Kruiskamp 3 , Celia Juan-Cruz 1 , Uulke A. van der Heide 1 , Alice Couwenberg 1 , Petra J. van Houdt 1 , Tomas Janssen 1 1 Department of Radiation Oncology, The Netherlands Cancer Institute, Amsterdam, Netherlands. 2 Department of Radiotherapy, Amsterdam UMC, Amsterdam, Netherlands. 3 MR Clinical Science, Philips Healthcare, Best, Netherlands Purpose/Objective: Severe radiation-induced lymphopenia has been associated with poorer oncological outcomes [1]. Dixon-MRI based proton-density fat fraction (PDFF) quantification has shown that bone marrow fattiness increases as blood lymphocyte counts decrease following chemoradiation for cervical cancer [2]. The time-scale and dose-effect relation at which these changes occur remains largely unknown. Elucidating the dose-time-effect relationship could prove useful in adapting bone marrow sparing strategies during the course of the treatment, potentially reducing lymphopenia. In this initial work, we seek to fill in the temporal gap by measuring the PDFF during each daily fraction on a 1.5T MRI-linac. Material/Methods: We obtained data from 12 patients who received radiotherapy treatment for rectal cancer enrolled in the MOMENTUM study (NCT04075305).Patients were treated with either short course radiotherapy (SCRT) 5x5Gy (N=8) or concurrent chemoradiation (CCRT) 25x2Gy with capecitabine (N=4). For each fraction, we acquired anatomical (T2w) and voxel-wise PDFF data (T1 mDixon Quant, adapted from [3]) on a 1.5T MRI- linac (Elekta AB). TotalSegmentator [4] was used to delineate seven bony structures on the synthetic planning CT. All delineations were isotropically eroded by 0.5cm to exclude cortical bone.PDFF maps were rigidly registered to the synthetic CT and structure- mean values were extracted for each fraction for longitudinal assessment. We determined the Pearson correlation coefficient between the structure-wise mean planning doses and difference in PDFF between first and last fraction. Results:
Digital Poster 312
Cross-Cancer Identification of a Radiotherapy- Specific Prognostic Gene Signature Using Machine Learning Yo-Liang Lai 1,2 , Chia-Hsin Liu 3 , Pei-Chun Shen 4 , Yu-De Wang 5 , Wei-Chung Cheng 4,6 1 Department of Radiation Oncology, China Medical University Hospital, Taichung, Taiwan. 2 School of Medicine, College of Medicine, China Medical University, Taichung, Taiwan. 3 Graduate Institute of Clinical Medicine, College of Medicine, Kaohsiung Medical University, Kaohsiung, Taiwan. 4 Cancer Biology and Precision Therapeutics Center, China Medical University, Taichung, Taiwan. 5 Department of Urology, China Medical University Hospital, Taichung, Taiwan. 6 Graduate Institute of Biomedical Sciences, China Medical University, Taichung, Taiwan Purpose/Objective: Radiotherapy (RT) is central to cancer care, yet responses vary widely and current biomarkers rarely capture RT-specific effects. We aimed to derive a robust, RT-focused prognostic signature using multi- cohort genomics and machine learning (ML). Material/Methods: We analyzed a large-scale, multi-cohort genomic resource (DriverDBv4[1]; ~24,000 samples across 70 cohorts). Patients were grouped as RT-alone, chemoradiotherapy (CRT), or no-RT. Five cancer types with ≥ 50 RT-alone cases were included. An ML pipeline screening >100 model/feature-selection combinations was used to identify reproducible gene signatures, with cross-cohort validation. Results: We identified multiple candidate signatures for RT response. A 28-gene model showed consistent risk stratification across multiple RT-alone cohorts and maintained performance in CRT cohorts. Notably, the signature showed no association in no-RT groups, supporting RT specificity across cancer types and treatment settings. Conclusion: We report a cross-cancer, RT-specific 28-gene signature that generalizes across independent cohorts and treatment contexts. These findings support integrating genomics and ML to advance personalized radiotherapy and motivate prospective validation. References: 1.Liu C-H, Lai Y-L, Shen P-C, et al. DriverDBv4: a multi- omics integration database for cancer driver gene research. Nucleic Acids Res. 2024;52(D1):D1246– D1252. Keywords: radiotherapy, machine learning, gene signature
Figure 1. PDFF over time (bands:SD).Most structures in
Made with FlippingBook - Share PDF online