Author: Cheng-Chieh Cheng, Jeffrey P Guenette, Yajun Li, Bruno Madore, Lei Qin 👨🔬
Affiliation: Brigham and Women's Hospital, National Sun Yat-sen University, Dana-Farber Cancer Institute 🌍
Purpose: Pilot tone (PT), a compact RF sensor, has been integrated into clinical practice for motion detection. Prior studies proposed mapping PT signals to head positions using a calibration step that requires acquiring extra training images. While effective, integrating the training stage into clinical workflows is challenging. Image metric-based correction (i.e. autocorrection), such as gradient entropy-based approaches have effectively corrected motion artifacts. However, they struggle with rotation angles exceeding 3˚, and often fail to converge in high dimensional phase encoding scenarios. We propose a calibration-free method for motion correction that eliminates the need for training images. Instead, gradient entropy minimization is used to estimate a linear transformation from PT signals to rigid-body motion parameters.
Methods: Data were obtained from four healthy volunteers on a 3T scanner using a 20-channel head/neck coil. For each Subject, up to 5 MPRAGE scans were run, with increasing motion complexity. The PT signals are extracted from the oversampled region of each readout. The rigid motion parameters z(r;t) is calculated from the PT signals p through a linear transformation C, as z(r;t)=Cp. The image x is reconstructed based on the motion parameters:
x=SHFH(r)TH(t)y Where SH, TH(t) and y represent the coil sensitivities, translation-related phase ramps, and k-space data, respectively. F is the inverse non-uniform Fourier operator, with coordinates rotated by r. The linear coefficients C are updated iteratively toward minimizing the gradient entropy E on x.
Results: We achieved significant improvement in image quality across various motion levels from 4 volunteers, even for large motion ranges with rotations of ±23° and translations of ±15 mm. Structural Similarity Index and Dice score improved ~30% when comparing uncorrected and corrected images to references.
Conclusion: This proposed method works without the need for a calibration scan. Effective estimation of motion information from the non-intrusive PT makes this approach practical.