Explicitly encoding the cyclic nature of breathing signal allows for accurate breathing motion prediction in radiotherapy with minimal training data

Abstract

Background and purpose: Active breathing motion management in radiotherapy consists of motion monitoring, quantification and mitigation. It is impacted by associated latencies of a few 100 ms. Artificial neural networks can successfully predict breathing motion and eliminate latencies. However, they require usually a large dataset for training. The objective of this work was to demonstrate that explicitly encoding the cyclic nature of the breathing signal into the training data enables significant reduction of training datasets which can be obtained from healthy volunteers.

Material and methods: Seventy surface scanner breathing signals from 25 healthy volunteers in anterior-posterior direction were used for training and validation (ratio 4:1) of long short-term memory models. The model performance was compared to a model using decomposition into phase, amplitude and a time-dependent baseline. Testing of the models was performed on 55 independent breathing signals in anterior-posterior direction from surface scanner (35 lung, 20 liver) of 30 patients with a mean breathing amplitude of (5.9 ± 6.7) mm.

Results: Using the decomposed breathing signal allowed for a reduction of the absolute root-mean square error (RMSE) from 0.34 mm to 0.12 mm during validation. Testing using patient data yielded an average absolute RMSE of the breathing signal of (0.16 ± 0.11) mm with a prediction horizon of 500 ms.

Conclusion: It was demonstrated that a motion prediction model can be trained with less than 100 datasets of healthy volunteers if breathing cycle parameters are considered. Applied to 55 patients, the model predicted breathing motion with a high accuracy.

Keywords: 4D image guidance; Intrafractional motion; Long short-term memory network; Motion prediction; Real-time tumour motion monitoring.

Fig. 1

Demonstration of data pre-processing for Model 2. a) the interpolated and filtered dataset after peak-detection including the dynamically scaled baseline (dashed red line) and the centre line at 50 % inhale (dashed blue line) representing y_50% for dynamic scaling. The peaks were the basis for calculation of the deflection d shown in b) representing one-half of a breathing amplitude and the breathing phase φ(t) shown in c). Each prediction was based on the breathing signal of an input window x. The window length was a parameter of the LSTM model (illustrated with 8 s corresponding to the best results, see Section 3.2). The prediction horizon $\widehat{x}$ was set at 500 ms.

Fig. 2

Distribution of all breathing amplitudes presented in the datasets. For both datasets (healthy volunteers on the left side and patients on the right side) the distribution of breathing amplitudes was separated into two groups: a group with a large breathing amplitude and a group with a small breathing amplitude where a presence of an amplitude > 12 mm was used for separation. The dashed and dotted vertical lines show the median and the mean of each group for each dataset, respectively.

Fig. 3

Examples of prediction accuracy for different breathing scenarios. For each example, the whole acquisition is shown in the left column, while the zoom-in is presented in the right column. The original breathing signal is the blue solid line and the predicted breathing signal dashed orange line. Below each breathing signal, the prediction error is given together with the absolute RMSE of the whole breathing data acquisition. Note that the scale of the breathing signal and error can be different due to different amplitudes of an acquisition.