ESTRO 2025 - Abstract Book

S3378

Physics - Machine learning models and clinical applications

ESTRO 2025

1289

Digital Poster An adaptive approach to enhance external-to-internal motion prediction by combining long short-term memory networks and time-domain cross-correlation Andrew R Milewski 1,2 , Uddin Fayed 2 , Vyas Grupta 2 , Xingyu Nie 3 , Guang Li 2 1 Department of Anesthesiology, Weil Cornell Medicine, New York, USA. 2 Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, USA. 3 Department of Radiology, Kentucky University College of Medicine, Lexington, USA Purpose/Objective: To evaluate the ability of long short-term memory (LSTM) deep-learning neural networks to accurately predict internal organ motion using an external surrogate in radiotherapy timeframes and to investigate an adaptation strategy that corrects residual phase shifts without retraining the LSTM models. Material/Methods: Under an IRB-approved study, 10 subjects received two 4DMRI scans: Each lasted 3-10min, and ~20min intervened between them. Concurrent waveforms from an internal navigator at the diaphragm and external bellows on the abdomen were acquired during each scan. Subject-specific models for external-to-internal motion prediction were built by training LSTM networks on the first half of the mid-term (3-10min) concurrent-waveform datasets. The models were deployed to predict the internal waveforms from the external waveforms for the remaining mid-term data and the entire long-term (20- 30min) datasets. The accuracy of the LSTM models’ predictions was evaluated by calculating the root mean square error (RMSE) and Pearson correlation (C) between the predicted and actual internal waveforms, and by calculating the error in predicting the temporal occurrence of respiratory peaks in the internal waveforms (∆t). RMSE, C, and ∆t values were calculated for the native concurrent waveforms and the phase shift-corrected waveforms by maximizing the time-domain cross-correlation (TCC) 1,2 . An adaptive approach using TCC to identify and correct residual phase shifts between the LSTM-predicted and actual internal waveforms was evaluated as a hybrid LSTM-TCC approach. Results: Compared to the native waveforms (C=0.42±0.28), the LSTM models yield predictions that correlate more strongly with the actual internal waveforms (C=0.89±0.07) in the mid-term datasets and surpass the performance of the TCC method (C=0.77±0.09). Over 20-30min, the correlations produced by the LSTM predictions were poor (C<0.63) for two subjects, but the LSTM-TCC approach rescued the performance (C>0.80) and yielded high correlations across all subjects (C=0.90±0.09). The LSTM models’ peak - prediction error (∆t=0.18±0.15sec) was also improved by the LSTM TCC approach (∆t=0.15±0.11sec). The improvement in RMSE (∆RMSE) achieved by the LSTM models in the training datasets correlated highly with the correlation enhancement (∆C) found for the mid - term dataset: C’(∆RMSE -to- ∆C)= - 0.91. Conclusion: The subject-specific LSTM models are accurate and stable (C=0.89) over 3-10min, and — by adapting to changes in breathing patterns that occur over longer timeframes — the hybrid LSTM-TCC method is stable (C=0.90) beyond 30min without retraining the LSTM network. Hyperparameters for LSTM modeling can be informed based on the ∆RMSE -to- ∆C correlation. The hybrid approach provides a potential online, adaptive solution for respiratory gated radiotherapy.

Keywords: deep learning, adaptive model, motion prediction

References: 1.

Milewski AR, Olek D, Deasy JO, Rimner A, Li G. Enhancement of Long-Term External-Internal Correlation by Phase-Shift Detection and Correction Based on Concurrent External Bellows and Internal Navigator Signals [published online ahead of print 2019/04/24]. Adv Radiat Oncol. 2019;4(2):377-389.