ESTRO 2025 - Abstract Book

S3397

Physics - Machine learning models and clinical applications

ESTRO 2025

2340

Digital Poster Data efficiency and prolonged prediction horizons for cyclic long short-term memory networks Julius Arnold 1,2 , Barbara Knäusl 1,2 , Martin Heilmann 1,2 , Joachim Widder 1,2 , Dietmar Georg 1,2 , Andreas Renner 1,2 1 Department of Radiation Oncology, Medical University of Vienna, Vienna, Austria. 2 Christian Doppler Laboratory for Image and Knowledge Driven Precision Radiation Oncology, Medical University of Vienna, Vienna, Austria Purpose/Objective: Non-learning-based breathing motion prediction models face difficulties for prediction horizons beyond 200ms. This limits their application for standard C-arm linacs having longer latency times. Recently, Long Short-Term Memory (LSTM) models showed effectiveness for a prediction horizon 500ms after hyperparameter tuning [1]. Dividing the breathing signal into its cyclic nature for model training enabled further improvement of the prediction accuracy [2]. The purpose of this study was to investigate the performance of the cyclic LSTM for prediction horizons up to one second. Additionally, we explored the number of training datasets required for low prediction errors. Material/Methods: Breathing motion datasets from 207 patients were acquired with a Catalyst-HD surface scanner (C-Rad, Sweden) (IRB: 1239/2022). Preprocessing included interpolation to achieve a constant acquisition frequency, cropping to remove setup artifacts and low-pass filtering for noise reduction. Each breathing pattern segment was dynamically scaled to [-1,1]. The scaled breathing signal was used as input to predict the signals decomposed baseline, deflection, and phase components, which were re-combined and re-scaled for evaluation. 20% of patient datasets were used for testing, while the remaining ones were 90:10 split for training and validation. Model training was conducted for 100 epochs with the model checkpointing based on minimum validation loss. For the first evaluation, prediction horizons from 100ms to 1000ms, in 100ms increments, were used with an 8s input window, training ten models per horizon. For the second evaluation, breathing signals of 5, 10, 20, 30, 50, 70, 90, 110, 130, and 150 patients were each ten times randomly selected. Training:validation splits were performed and the models trained with a 500ms prediction horizon. Root mean square errors (RMSEs) between re-combined predicted values and true signals of the identical held-out test set were calculated. Results: RMSEs increased moderately with median values below 0.15mm for horizons up to 500ms and increased more rapidly for longer horizons, remaining below 0.5mm (Figure 1). Figure 2 shows RMSEs for different sizes of the training data and a prediction horizon of 500ms. A plateau in the prediction error was observed with as little as 30 patients.