Back to all papers

Assessment of simulated respiratory gating based on external-to-internal motion prediction using a hybrid, adaptive deep-learning network.

November 18, 2026pubmed logopapers

Authors

Milewski A,Nie X,Li G

Affiliations (3)

  • Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, New York, USA.
  • Department of Anesthesiology, Weill Cornell Medicine, New York, New York, USA.
  • Department of Radiology, Kentucky University College of Medicine, Lexington, Kentucky, USA.

Abstract

To assess the accuracy, safety, and efficiency of respiratory-gated radiotherapy (RGRT) based on internal respiratory waveforms predicted by a hybrid long short-term memory (LSTM) deep-learning network, which adapts to breathing irregularities through the application of time-domain cross-correlation (TCC) to check and correct any residual phase/time shift in the LSTM predictions. The adaptive LSTM-TCC model was built in two steps: (1) subject-specific LSTM models were trained to predict internal diaphragm motion from external bellows motion using each of 10 subjects' concurrent waveforms (≥20 Hz, 3-10 min) acquired from four-dimensional magnetic resonance imaging (4DMRI) protocol scans and (2) near real-time TCC was applied to the LSTM predictions to check and correct any residual phase/time shifts to safeguard against new breathing irregularities beyond the training data in the mid-term (3-10 min) and long-term (20-30 min) datasets. The testing datasets comprised both the mid-term and long-term concurrent motion waveforms, and the external-to-internal motion predictions and gating quality assessments for each subject were evaluated and validated with the actual internal motion as the ground truth. The hybrid LSTM-TCC models were compared with LSTM, TCC, and native external-to-internal correlation models. The temporal gating accuracy (beam-off and beam-on) and beam-gating quality (missing-target rate: %harm, and beam-on efficiency: %effic) of the hybrid LSTM-TCC model were evaluated and compared with LSTM, TCC, and native predictions. The hybrid LSTM-TCC models produced the highest long-term sustained correlations (C > 0.8, within 30 min), yielding the lowest %harm = 5 ± 6% and the highest %effic = 67 ± 15%, compared with the LSTM (%harm = 8 ± 8%, %effic = 65 ± 16%), the TCC (%harm = 29 ± 31%, %effic = 43 ± 19%), and the native (%harm = 37 ± 23%, %effic = 38 ± 14%). The hybrid LSTM-TCC models produced a high, sustained performance (C = 0.9 ± 0.1) without remodeling, regardless of new breathing irregularities, and the average temporal error of the predicted gating windows was ∆t < 0.3 s. This study assesses the simulated respiratory-gating qualities of four methods, demonstrating that the hybrid LSTM-TCC approach outperforms the native waveforms, the TCC method, and the LSTM models, yielding the highest correlation, gating accuracy, and %effic, but the least %harm. The hybrid LSTM-TCC model can safeguard LSTM predictions and gating accuracy against major breathing irregularities, demonstrating a strong potential for future clinical applications.

Topics

Respiratory-Gated Imaging TechniquesDeep LearningRadiotherapy Planning, Computer-AssistedLung NeoplasmsImage Processing, Computer-AssistedJournal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAISlice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.