ESTRO 2025 - Abstract Book

S3471

Physics - Optimisation, algorithms and applications for ion beam treatment planning

ESTRO 2025

University Hospital Essen, Essen, Germany. 8 Department of Radiotherapy and Radiation Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, TUD Dresden University of Technology, Dresden, Germany

Purpose/Objective: Neglecting the increase in relative biological effectiveness (RBE) with rising linear energy transfer (LET) might result in an underestimation of side effects in proton beam therapy. To analyse the clinical relevance of RBE effects, large retrospective datasets are required. The inclusion of patients, however, can be impeded if the LET is not available from the treatment-planning system, which is particularly relevant for older datasets. To address this challenge, a convolutional neural network (CNN) predicting LET from dose distributions has recently been developed and validated [1], but its performance on datasets lacking ground truth LET is unknown. Therefore, this study aims to explore uncertainty estimation techniques to assess the applicability of the CNN-based LET model on independent datasets, verifying the use of LET predictions for further RBE research. Material/Methods: The study utilized several sets of proton therapy treatment plans that were not considered for model training: 46 pencil beam scanning (PBS) plans and 45 double scattering (DS) plans from centre 1 and 122 PBS plans from centre 2. Two uncertainty estimation techniques were explored: 1. Latent space distance: Feature maps from the input dose distribution using the CNN's final encoder were flattened to define the latent space. The Euclidean distance to the 25 nearest neighbours of the CNN’s training set in the latent space and its Spearman correlation with the prediction error (median absolute difference between LET prediction and ground truth LET) were calculated. 2. Ensemble models: An ensemble of five CNN models, previously generated through five-fold cross-validation [1], was used to assess prediction uncertainty. The median variance of the LET predictions was calculated and correlated with the prediction error (median absolute difference between LET prediction and ground truth LET). Results: A moderate Spearman correlation was observed between the latent space distance and the prediction error (ρ = 0.45, p<0.001), indicating that greater distance was associated with higher errors (Table 1, Figure 1). The variance across ensemble models correlated strongly with the prediction error (ρ = 0.72, p<0.001), suggesting ensemble based uncertainty is a reliable predictor of model performance.