S2137
Physics - Inter-fraction motion management and daily adaptive radiotherapy
ESTRO 2026
but struggled with thin collections (<2-3 mm), more commonly observed in older patients (Fig.2).
Hospitals NHS Foundation Trust, London, United Kingdom
Purpose/Objective: The performance of deep-learning (DL) in image- guidance radiotherapy workflows relies on large and diverse datasets1. To address difficulties in amassing data from young patients, we used publicly available datasets to train DL-models for radiograph-based monitoring of inter-fractional change during abdominal radiotherapy. These models were tested on multi-centre data from young patients to investigate generalisability and feasibility of cross-institutional DL- model training and sharing to promote equitable and unbiased care. Material/Methods: Two supervised DL networks were trained to predict global gastrointestinal (GI) gas volumes (CNN/U-Net) and/or local GI pathlength maps (U-Net) from digitally reconstructed radiographs (DRRs) derived from CTs using TIGRE2; the paired labels (volume and pathlength maps) resulted from forward-projecting GI gas segmentations. The collection of comprehensive young people radiotherapy datasets is challenging due to the rarity of cancer in early life, diversity in disease and developmental stages, inter-institutional variability in imaging protocols, data governance and patient privacy regulations. Model development and internal evaluation were therefore performed using The Cancer Imaging Archive Paediatric-CT-SEG dataset (Ntrain-validation=315, Nheld-out-test=34, 0–16y) containing diagnostic CTs from subjects with unknown conditions (North-America)3. This dataset was augmented fivefold by simulating GI gas shrinkage/expansion, organ motion, and anatomical growth with population-derived parameters4–6. For each network, predictions were obtained from averaging five independent models generated through five-fold partitioning and training. External evaluation was performed on DRRs generated from planning CTs of multi-institutional radiotherapy cohorts: children and young people in Europe (CYP; N=21; 2–19y), and teenagers and young adults in Asia (TYA; N=14; 10– 23y). Model performance was assessed via mean absolute error (MAE) of GI gas metrics and Dice Similarity Coefficient (DSC) of GI segmentations. Results: MAE for GI gas volume and pathlengths were similar for internal and external datasets, demonstrating the good generalisability of both DL-models (Fig.1a-c). On the held-out Paediatric-CT-SEG test set, the outliers were linked to extreme GI air volumes (>1.1 L). DSC values were lower for TYA than for Paediatric-CT-SEG and CYP (Fig.1d, 0.77±0.12 vs 0.88±0.08 and 0.82±0.12, respectively) despite lower MAE, indicating poorer generalisability due to cohort-specific gas morphology. The U-Net model reliably detected large gas pockets
Conclusion: The proposed image-guidance models trained on public augmented datasets generalised relatively well to young radiotherapy patients of different ages, demographics, and patient groups. Our approach promotes equity and fairness in deploying DL methodology across institutions. References: 1.Yamashita R, et al. Convolutional neural networks: an overview and application in radiology. Insights Imaging. 2018;9(4):611-629. 2.Biguri A, et al. TIGRE: a MATLAB-GPU toolbox for CBCT image reconstruction. Biomed Phys Eng Express. 2016;2(5):055010. 3.Jordan P, et al. Pediatric chest - abdomen - pelvis and abdomen - pelvis CT images with expert organ contours. Med Phys. 2022;49(5):3523- 3528. 4.Guerreiro F, et al. Intra- and inter-fraction uncertainties during IGRT for Wilms’ tumor. Acta Oncol. 2018;57(7):941-949. 5.Kuczmarski RJ, et al. CDC growth charts: United States. Adv Data. 2000;(314):1- 27. 6.Chlap P, et al. PlatiPy: Processing Library and
Made with FlippingBook - Share PDF online