ESTRO 2025 - Abstract Book

S3759

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

1913

Mini-Oral Tumour-to-stroma ratio identified by artificial intelligence impacts cancer outcomes in a phase III randomised rectal chemo-radiotherapy trial Zhuoyan Shen 1 , Douglas Brand 1,2 , Mikael Simard 1 , Andre Lopes 3 , Rubina Begum 3 , Nicholas West 4 , Ying Zhang 1 , Sumeet Hindocha 2,1 , Gary Royle 1 , Tim Maughan 5 , David Sebag-Montefiore 4 , Charles Antoine Collins Fekete 1 , Maria Hawkins 2,1 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. 3 Cancer Institute, University College London, London, United Kingdom. 4 School of Medicine, University of Leeds, Leeds, United Kingdom. 5 Department of Oncology, University of Oxford, Oxford, United Kingdom Purpose/Objective: Neoadjuvant chemoradiotherapy (nCRT) has demonstrated improved local control in locally advanced rectal cancer (LARC) in clinical trials. However, the survival benefits and optimal patient selection criteria remain under investigation. The tumour-associated stroma plays a critical role in tumour progression and therapy resistance in rectal cancer. This study explores the impact of the tumour-to-stroma ratio (TSR) identified by artificial intelligence (AI) on outcomes in LARC patients after nCRT. Material/Methods: ARISTOTLE (ISRCTN09351447) [1] is a phase III clinical trial randomly assigning LARC patients to standard nCRT (CRT) – preoperative radiotherapy with capecitabine or experimental nCRT (IrCRT) – preoperative radiotherapy with dose reduced capecitabine plus irinotecan. Curative surgical resection was intended 8-10 weeks post-nCRT. This study utilised clinical data and digitised Haematoxylin-and-eosin-stained whole slide images (WSIs) from patients with available pre-treatment biopsies (N=414/589). We trained an AI model to quantify the TSR automatically ( Figure 1 ). In a randomly selected hold-out cohort (N=174) we identified the optimal cut-off value to classify patients’ TSR into low (TSR – ) or high (TSR + ). This value was determined as the point that maximises the Youden index in a receiver operating characteristic analysis for predicting three-year disease progression. The remaining cohort (N=240) was used to compare disease-free survival (DFS) and overall survival (OS) between the TSR-stratified groups. Kaplan-Meier estimation was used for survival analysis, with log-rank tests to compare survival curves between subgroups. Hazard ratios (HR) and 95% confidence intervals (CI) were given by a Cox proportional hazards model.