Optimization of Gradient Pulse Shape Prediction Using Temporal Convolutional Networks.

PURPOSE: To improve TCN-based gradient-waveform prediction with respect to acquisition efficiency and inference speed, and to evaluate its use for radial imaging.

METHODS

Gradient waveforms were measured using a thin-slice method on a preclinical MRI system. Separate TCN models were trained for each physical gradient axis to predict both gradient and B 0 terms. The framework was evaluated using phantom and in vivo radial acquisitions across a range of receiver bandwidths and compared with measured trajectories and gradient impulse response function (GIRF)-based trajectory predictions.

RESULTS

The proposed workflow reduced model complexity and substantially improved prediction speed. In phantom and in vivo radial imaging, model-based trajectories provided better image quality over a wider range of receiver bandwidths than GIRF-based trajectories. B 0 term estimation provided limited additional visual benefit after scanner preemphasis adjustment.

CONCLUSION

TCN-based prediction of gradient-waveform distortions can be made more efficient while remaining effective for radial MRI trajectory estimation on the studied preclinical system, particularly in demanding high-bandwidth acquisitions.