ESTRO 2025 - Abstract Book

S1277

Clinical – Lower GI

ESTRO 2025

4240

Digital Poster A novel CT radiomics-based machine learning approach: predicting NAR and overall survival in rectal cancer patients treated with TNT Constanza Hormazábal González 1 , Isabella Liedtke Grau 1 , Gabriel Lazcano Álvarez 1 , Tomás Walter Martin 2 , Darlett Folch Mora 2 , Francisco Pérez Peña 2 , José Solís Campos 1 1 Oncology, Hospital Carlos Van Buren, Valparaíso, Chile. 2 Radiation Oncology, Universidad de Valparaíso, Valparaíso, Chile Purpose/Objective: Emerging technologies, such as artificial intelligence and radiomics, offer potential to enhance the predictive power of medical imaging. The Neoadjuvant Rectal (NAR) score has shown superior predictive value over pathologic complete response (pCR) for overall survival (OS) in locally advanced rectal cancer (LARC) patients undergoing total neoadjuvant therapy (TNT) [1]. However, these biomarkers rely on postoperative data. Predicting treatment response using pre-treatment CT scans could enable more accurate preoperative decision-making and personalized treatment strategies, improving patient care. This study aims to develop two machine learning (ML) models using preoperative clinical data and simulation CT radiomics: 1. A model to predict three-year overall survival (3-year OS). 2. A model to predict a binary NAR score category, with a threshold of 8.4 based on a local study [2]. Material/Methods: This retrospective study included LARC cancer patients treated with TNT at a single centre between 2017 and 2021. TNT consisted of short-course pelvic radiotherapy followed by FOLFOX or CAPOX chemotherapy.The manually segmented CTV included the mesorectum and nodal regions. This study received local ethics committee approval. Two ML models were developed to predict 3-year OS and NAR binary score using preoperative clinical data and CT radiomics. Radiomic features were extracted using PyRadiomics v3 [3]. Dimensionality reduction involved: 1. Removing highly correlated features (Pearson correlation coefficient >0.99). 2. Feature selection via iterative Random Forest [4] algorithm. A 10-fold cross-validation method was used to evaluate ten Scikit-learn [4] supervised classification algorithms. Results: The 3-year OS model included 100 patients (45 female) with a median age of 61 years (IQR 52.75–70). Clinical staging included 7 cT2, 68 cT3, and 25 cT4 patients, with 86 cN+. The 3-year OS rate was 54%. Two models were developed: 1. A simplified model using demographic and clinical features, achieving an AUC of 0.65 (SD 0.03) with K Nearest Neighbors [4]. 2. A model incorporating radiomic features, achieving an AUC of 0.81 (SD 0.07) with Support Vector Machines