S2311
Physics - Machine learning and AI algorithms
ESTRO 2026
average reduction of 47%, while tumours >4cm ³ exhibited minimal response. At 12Gy, tumours <4cm ³ showed consistent decrease of 8-27%, while those >4 cm ³ reduced by up to 77%. At 12.5Gy, shrinkage was most pronounced, ranging from 28% tumours <4 cm ³ to 58% for large lesions. Cystic composition and NF2 status were dominant morphological predictors of non-linear tumour volume change, while intracanalicular involvement was associated with early stabilization. The inclusion of temporal MITD data improved the model’s ability to capture subtle early morphological changes preceding measurable regression. Conclusion: Machine learning and artificial intelligence provide a robust and interpretable framework for suggesting tumour response after SRS. By integrating morphological, anatomical, and dosimetric data, the models offer a more specific, data driven understanding of how prescription and tumour morphology jointly affect response. This approach supports adaptive dose planning and individualized follow up strategies for patients with vestibular schwannoma. Keywords: radiosurgery, tumour morphology, predictive models Poster Discussion 3841 Epistemic and aleatoric uncertainty in automated planning Antony Carver, Stuart Green Medical Physics, University Hospital Birmingham, Birmingham, United Kingdom Purpose/Objective: AI based dose prediction is evolving into a standard tool for improving the quality and consistency of treatment plan quality. A key limitation over simpler methods is that interval estimation is challenging. Consequently AI based tools rarely produce a range of achievable outcomes or estimate of uncertainty.This study presents a Bayesian neural network which outputs both epistemic (uncertainty due to parameters) and aleatoric uncertainties (uncertainty due to data variance) by replacing weights with random variables. This is demonstrated by quantifying the variation in automated plan quality due to model uncertainty. Material/Methods: A dose prediction model was trained using 340 prostate cases. Two hundred were used for training, 40 for on-line evaluation and the final 100 were held back for offline validation and analysis. A pair of cascaded, 3-dimensional UNets formed the basis for the dose prediction model. The structure is shown in
Digital Poster 3796
ML and AI based predictions in vestibular schwannoma dynamics after radiosurgery integrating morphological, anatomical, and dosimetric insights Eleni Rozaki 1 , Michael Amoo 2 , Eoin Minnock 2 , Mohsen Javadpour 2 , David Fitzpatrick 3 , Caroline Hayhurst 4 , Nachi Palaniappan 5 , Christina Skourou 3 1 1. School of Marketing and Entrepreneurship, Technological University Dublin, Dublin, Ireland. 2 2. Department of Neurosurgery, Beaumont Hospital, Dublin, Ireland. 3 Radiation Oncology, St Luke's Radiation Oncology Network, Dublin, Ireland. 4 4. Department of Neurosurgery, University Hospital of Wales, Cardiff, United Kingdom. 5 Clinical Oncology, Velindre Cancer Centre, Cardiff, United Kingdom Purpose/Objective: Predicting tumour response following stereotactic radiosurgery (SRS) for vestibular schwannoma (VS) remains a challenge due to patient specific differences in tumour morphology and prescribed dose. Conventional statistical approaches fail to capture the complex and non-linear interactions between these features and long term volumetric outcomes. This study developed and evaluated machine learning (ML) and artificial intelligence (AI) models for predicting tumour size changes over 1-, 2-, and 3-year intervals after SRS, aiming to determine how anatomical, morphological, and dosimetric factors together influence this response. Material/Methods: A longitudinal dataset of 148 patients with VS treated with SRS was analysed. Predictive models were trained using features such as laterality, intracanalicular involvement, NF2 status, cystic composition, mean intracanalicular tumour diameter (MITD), and marginal dose (11Gy, 12Gy, and 12.5Gy) on 70% of the cohort. Several supervised algorithms including Random Forest (RF), Gradient Boosting (GB), Support Vector Machines (SVM), and Deep Neural Networks (DNN) were compared against baseline regression models for predictive performance on the remaining 30% of the cohort. Model performance was evaluated using accuracy, precision, and area under the curve metrics. Explainable AI methods, including SHAP analysis, were used to interpret the influence of each predictor on volumetric outcomes. Results: AI based models demonstrated superior predictive accuracy compared to traditional approaches. RF and GB emerged as the most reliable algorithms for integrating all factors. Both morphology and marginal dose were significant predictors of tumour size reduction. At 11Gy, tumours ≤ 1 cm ³ showed an
Made with FlippingBook - Share PDF online