ESTRO 2025 - Abstract Book

S3383

Physics - Machine learning models and clinical applications

ESTRO 2025

1546

Digital Poster Enhancing the accuracy of respiratory-gated radiotherapy (RGRT) by using a hybrid deep-learning model to

predict respiratory-induced organ motion Andrew R Milewski 1,2 , 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: By deploying long short-term memory (LSTM) neural networks to predict internal respiratory motion from an external surrogate, we aim to enhance the accuracy, efficiency, and quality of the respiratory gating for respiratory gated radiotherapy (RGRT). We explore the gating benefits by adaptively correcting the residual phase shifts between the predicted and actual waveforms by maximizing their time-domain cross-correlation (TCC). Material/Methods: Concurrent waveforms from an external bellow and an internal navigator were acquired over 3-10min (mid-term) and again ~20min later (long-term) for ten subjects under an IRB-approved protocol. Subject-specific models were built by training LSTM networks on the first 50% of the mid-term waveforms and then deployed to predict internal waveforms from the long-term, external waveforms. Residual phase shifts between the LSTM-predicted and actual waveforms were corrected using the TCC method, termed the hybrid approach. The entry and exit to the gating window (under 30% amplitude) were simulated using gating waveforms (native bellows, TCC-corrected bellows, LSTM prediction, or hybrid prediction), and the target was defined as the lower half of the actual internal waveform. The %harm — defined as the ratio of missing-the-target time over total beam-on time — and %effic — defined as the actual, valid beam-on time over all possible beam-on time — were calculated to assess gating quality 1 . The beam-on and beam-off times predicted by each method were evaluated against the ground truth residing in the internal waveforms. Statistical comparisons were performed via the Wilcoxon signed-rank test. Results: The gating predicted by the hybrid approach yielded less harm (%harm=5.2±5.7%) and higher efficiency (%effic=66.5±15.1%) compared with that obtained by the LSTM models (%harm=8.3±8.4%, %effic=64.5±15.9%), TCC method (%harm=29.2±31.3%, %effic=42.8±18.8%), and native waveforms (%harm=37.1±23.1%, %effic=37.7±13.8%). The hybrid approach and LSTM models yielded significantly better results than the TCC method and native waveforms ( p <0.006 for all comparisons). The average errors in predicting the beam-on and beam- off times were smaller for both the hybrid approach (∆t ON =0.47±0.26sec, ∆t OFF =0.54±0.52sec) and LSTM models (∆t ON =0.57±0.31sec, ∆t OFF =0.52±0.54sec) than for the TCC method (∆t ON =0.70±0.25sec, ∆t OFF =1.09±0.65sec) and native waveforms (∆t ON =0.77±0.51sec, ∆t OFF =1.46±0.72sec), though the differences were significant only for the beam-off times ( p <0.01 for all comparisons). Conclusion: The simulated respiratory gating achieved by the hybrid approach was superior to that obtained by either the LSTM or TCC methods over timeframes commensurate with radiotherapy treatments (20-30min), suggesting a strong potential for applying the hybrid model in RGRT.

Keywords: adaptive learning, respiratory-gated radiotherapy,

References: 1.

Milewski A, Li G. Stability and Reliability of Enhanced External-Internal Motion Correlation via Dynamic Phase-Shift Corrections Over 30-min Timeframe for Respiratory-Gated Radiotherapy. Technol Cancer Res Treat. 2022;21:15330338221111592.