S2454
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Conclusion: Our study showed that radiomic features, particularly shape sphericity, can predict post-SBRT liver toxicity. Selecting average or stable phases improves prediction, with visualizations highlighting their deviations. Future work should combine radiomic signatures with clinical and dose data, as well as delta- radiomics across phases. References: [1] Davey et al. Optimising use of 4D-CT phase information for radiomics analysis in lung cancer patients treated with stereotactic body radiotherapy. Phys.Med. Biol. 66, 11 (2021).[2] Griethuysen et al. Computational radiomics system to decode the radiographic phenotype. Cancer research 77, 21 (2017).[3] Ranstam and Cook. LASSO regression. British Journal of Surgery, Volume 105, Issue 10 (2018).[4] Breiman. Random forests. Machine learning 45.1 (2001).[5] Lundberg and Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30 (2017). Keywords: Liver toxicity, phase stability, SBRT prostate–rectum spacing metric in prostate SBRT Owen McLaughlin 1 , Stephen J McMahon 1 , Suneil Jain 1,2 , Conor K McGarry 3,1 1 Johnston Cancer Research Centre, Queen's University Belfast, Belfast, United Kingdom. 2 Department of Clinical Oncology, Northern Ireland Cancer Centre, Belfast Health and Social Care Trust, Belfast, United Kingdom. 3 Radiotherapy Physics, Northern Ireland Cancer Centre, Belfast Health and Social Care Trust, Belfast, United Kingdom Purpose/Objective: Spacers reduce gastrointestinal (GI) adverse events (AE) in patients undergoing prostate radiotherapy by increasing prostate-rectum distance [1]. Limited methods exist to quality assure these implants [2]. This study aims to develop an automated spacer metric that is predictive of AE, utilising manual or auto- contours already existing in the patient pathway. Material/Methods: A cohort of high-risk prostate cancer patients who received hydrogel spacers prior to imaging were analysed [3]. Patients received five fraction prostate and pelvic node SBRT (n=25) or prostate-only SBRT (n=15) with a PTV dose of 36.25 Gy. Planning structures were manually contoured on CT using MRI- CT fusion, whilst MRI and pre-treatment CBCT images were auto-contoured using Limbus Contour (v1.8.1). Auto-contours were not edited after their generation.Distance arrays were computed for each Digital Poster 2926 Prediction of adverse events using a novel
Results: Feature Selection—Feature selection identified 8-11 radiomic features across liver toxicity. Shape and GLSZM features were the most consistently selected, with shape sphericity emerging as a key predictor. Removing unstable features improved balanced accuracy (+0.04±0.02), while excluding redundant ones decreased performance (-0.03±0.01). Model Evaluation—Logistic regression (Fig.1) achieved the best performance, especially for patients with a change in the CP class (AUC=0.80) or an increase in CP score by at least 2 points (AUC=0.83). Random forests performed better for enzyme elevation (AUC=0.71) and slightly better for liver failure (AUC=0.94). Optimal and stable phase selection marginally improved model performance (5%).
Phase Stability Exploration—As opposed to the averaged CT, the stable phase shows minimal radiomic variation. Radial plots (Fig.2) indicated deviations between the two (up to 20%). Restrictive gates (30-70%) presented smaller average-to-stable deviations, meaning greater breathing stability. However, associations with liver failure and toxicity were inconclusive.
Made with FlippingBook - Share PDF online