ESTRO 2025 - Abstract Book

S3810

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Material/Methods: A PubMed search in Oct 2024 of (myelopathy OR neuropathy) AND (reirradiation OR re-irradiation) identified 306 evaluable cases of spine stereotactic body radiation therapy (SBRT) reirradiation from 13 studies. The inclusion criteria required reporting of initial spinal cord dose, reirradiation spinal cord dose, and the time interval between radiation courses. Prior spinal cord doses were within clinically accepted tolerances, preventing risk extrapolation and relevant application to cases that have already exceeded normal tolerance. Full 3D dose distributions, that would have allowed analysis of a spectrum of dosimetric factors, were not available. For cumulative maximum dose, BED was calculated with α/β=2Gy combined with a recovery function described using an exponential model. Residual dose from the first treatment was calculated assuming a varying discount factor ranging from 1 (no recovery) to a plateau value of 0.2 in increment of 0.05. A logistic model was used to describe dose-response. The best-fitting values of D50 and γ50 were calculated using maximum likelihood analysis, and 95% confidence intervals obtained using the profile likelihood method. Results: All models of dose-response were significant (p<0.001) with likelihood ratio tests. For limiting cases, model parameters were BED 50 =246.1 Gy (95% CI 215.1, 335.6Gy) and γ 50 =2.98 (1.80, 4.85), for no recovery model, parameters were BED 50 =152.2 Gy (132.5, 208.3) and γ 50 =2.49 (1.59, 4.08). Although statistically, it was not feasible to identify the best recovery model, these models can be readily used to assess the risks of myelopathy. Exploring alternate recovery models is in progress. Conclusion: Dose-response was observed for the incidence of spinal cord toxicity as a function of cumulative dose under varying assumptions to describe recovery. It was not possible to identify a preferred model. The raised exponential model used to describe recovery approximates the Gompertzian function and may be most intuitive. Digital Poster Artificial intelligence quantification of tumour lymphocyte infiltration enables colorectal cancer patients stratification to predict survival Zhuoyan Shen 1 , Mikael Simard 1 , Douglas Brand 2,1 , Ying Zhang 1 , Sumeet Hindocha 2,1 , Gary Royle 1 , Maria Hawkins 2,1 , Charles-Antoine Collins Fekete 1 1 Department of Medical Physics and Biomedical Engineering, University College London, London, United Kingdom. 2 Department of Radiotherapy, University College London Hospitals NHS Foundation Trust, London, United Kingdom Purpose/Objective: High levels of lymphocyte infiltration in the tumour microenvironment have been associated with improved outcomes in colorectal cancer (CRC). However, manual assessment of the lymphocyte infiltration is time-consuming, subject to inter- and intra-observer variability, and limited by region selection bias. This study analyses the survival impact of the tumour-infiltrating lymphocyte (TIL) density quantified using an artificial intelligence (AI) framework. Material/Methods: This study utilised the clinical data and the initial diagnostic H&E-stained whole slide images (WSIs) of 428 subjects from the TCGA-COAD and TCGA-READ datasets. To quantify TIL density within the WSIs, each tile (448 × 448 pixels, pixel size = 0.25 µm/pixel) was first classified by an EfficientNet model [1], trained on the NCT-CRC-HE-100K dataset [2], to identify tumour regions. Then, the TILs were detected by a YOLOv10 model [3], trained on the ImmunOcto dataset [4]. TIL density was calculated as the number of lymphocytes per 2.5 mm² within the identified tumour regions. Keywords: Reirradiation, NTCP, spinal cord toxicity 3078