ESTRO 2024 - Abstract Book

S5151

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Derivation of radiomic classifiers for radiotherapy response in locally advanced rectal cancer

Ross K McMahon 1 , Laura Grocutt 2 , Sean M O'Cathail 3 , Colin W Steele 1 , Martina Mori 4 , Jonathan J Platt 5 , Michael Digby 5 , Paul G Horgan 1 , Emiliano Spezi 6 , Campbell S Roxburgh 1 1 School of Cancer Sciences, College of Medical, Veterinary & Life Sciences, University of Glasgow, Academic Unit of Surgery, Glasgow, United Kingdom. 2 Cancer Research UK, Radnet Glasgow, University of Glasgow, Beatson West of Scotland Cancer Centre, Glasgow, United Kingdom. 3 School of Cancer Sciences, College of Medical, Veterinary & Life Sciences, University of Glasgow, Wolfson Wohl Cancer Research Centre, Glasgow, United Kingdom. 4 San Raffaele Scientific Institute, Medical Physics, Milan, Italy. 5 Glasgow Royal Infirmary, Radiology / Imaging Department, Glasgow, United Kingdom. 6 Cardiff University, School of Engineering, Cardiff, United Kingdom Neoadjuvant radiotherapy +/- chemotherapy (NAT) followed by surgery is the standard treatment in locally advanced rectal cancer (LARC). 1 A major challenge in LARC is the heterogeneity of radiation response - with a lack of robust predictive biomarkers. Those achieving pathological complete response (pCR) have an excellent long term outcome with 3-year DFS of 92.3%. 2 Furthermore, patients achieving clinical complete responses (cCR) to NAT can be entered into organ preservation protocols with active surveillance, thus avoiding the morbidity of rectal surgery. 3 This has brought into focus the need to better understand the processes governing response in order to develop novel predictive biomarkers and improve treatment allocation. Diagnostic imaging through Computerised Tomography (CT) and Magnetic Resonance Imaging are essential to LARC staging and radiotherapy planning. 4 Radiomics allows extraction of quantitative features from such imaging. The data that can be mined from these images could identify potential response biomarkers, as well as phenotype tumours when studied alongside histological samples. Combining features into ‘radiomic signatures’ has been shown to enhance prediction of pCR. 5 Here, we explore a radiomic model of NAT tumour response and regression in a large cohort of LARC patients. We included consecutive LARC patients from a single cancer centre who received curative-intent NAT between; June 2016–July 2021. Region-of-interests around the primary gross tumour volume (pGTV) were contoured on true-axial slices of radiotherapy planning CTs, with standardised patient positioning. pCR was defined as the absence of viable tumour locally and in nodes (ypT0N0). 6 Tumour regression grading (TRG) followed the recommended four-tier system; TRG0, no viable cancer cells, through to TRG3, extensive residual cancer with no regression. 6,7 The study aimed to define radiomic features (RF) in association with the following dichotomous endpoints; ‘complete response’ (pCR and cCR) vs ‘incomplete response’ (all others), and ‘good regression’ (pCR or TRG0-1 or cCR) vs ‘poor regression’ (TRG2-3). Radiomics was performed using SPAARC software 8 that includes 164 Image Biomarker Standardisation Initiative (IBSI) shape and texture features. 9-11 Spline interpolation was applied to achieve 2mm 3 isotropic voxel dimension. Fixed bin size of 25 Hounsfield units (HU) and CT segmentation range of [-1000, 1000] HU were used. Based on validated methodology 12 , the extracted RFs were combined into multivariate logistic regression (MVLR) on training cohort using Purpose/Objective: Material/Methods: