A Simulation Study of the Safety and Efficiency of Respiratory-Gated Radiotherapy Using a Hybrid Deep-Learning Network Adaptable to Breathing Irregularity to Predict Internal Motion
Abstract
Purpose
To evaluate the safety and efficiency of respiratory-gated radiotherapy by using a long-short-term memory (LSTM) model and its hybridized version with time-domain cross-correlation (TCC) to adapt to breathing motion irregularities without remodeling.
Methods
Subject-specific deep-learning LSTM networks were trained using concurrent external-internal motion waveforms (20Hz, 88s-300s) from 10 subjects, and the LSTM motion prediction (with a latency of 200ms) was safeguarded with a conventional TCC technique (LSTM-TCC) to check and correct (latency <100ms) any residual time/phase shifts in the LSTM predictions. The LSTM-TCC model yields an equal or better external-internal correlation by eliminating the residual time (phase) shifts. The LSTM model was trained with 1,760-60,000 time points (88-300s, 20Hz) of external-internal motion waveforms to predict internal motion over 20-30-minute timeframes. The 30% breathing amplitude from the simulation 4DCT was used as the gating threshold for the breamOn window. Using the native motions, TCC, LSTM, and LSTM-TCC predictions, the gating uncertainty (percentage of false beamOn time over all beamOn time, %harm) and gating efficiency (percentage of correct beamOn time over the available gating window, %efficiency) were used to quantify the gating quality.
Results
The LSTM and LSTM-TCC models produce low %harm=8.3±8.4% and 5.2±5.7% in 20-30 minutes, improved from %harm=17.1±18.2% by TCC and 45.1±23.8% for native data. High %efficiency of 64.5±15.9% and 66.5±15.1% are achieved by using LSTM and LSTM-TCC models, respectively, better than the conventional TCC method (%efficiency=48.2±18.8%). The adaptive LSTM-TCC demonstrates its advantage for subjects 4 and 7, who experience substantial breathing irregularities. A higher gating performance is observed within shorter timeframes (5-10 minutes) from the training dataset.
Conclusion
The simulation results demonstrate that the LSTM-TCC model is superior to LSTM alone, especially for patients with large breathing irregularities, as TCC safeguards against any residual time/phase shifts within the LSTM prediction caused by breathing motion irregularities.