Purpose: Respiratory motion estimation of the liver using A-mode ultrasound as a surrogate signal.
Methods: Two LSTM networks of differing complexity have been made to function as motion models. The performance of
these models was validated using a synthetic dataset. The best-performing model architecture was additionally validated on data recorded from three human subjects. The ground truth was acquired from simultaneously recorded B-mode ultrasound data.
Results: The synthetic dataset had a Real-time MAE (MAER) of 0.48 cm and 0.59 cm for the shallow and deep motion model respectively. Due to the better performance, the shallow model was further applied on the human subject data. The shallow model had an MAER of 0.83 cm, 0.18 cm and 0.54 cm for subject 1, 2 and 3 respectively.
Conclusion: Respiratory motion model performance differs significantly between subjects. The subjects with better model
performance also had better surrogate signal quality. If the surrogate signal is of sufficient quality, the current methodology
has the potential to outperform conventional biopsy protocol on tumours smaller than 1 cm.