Removal of Cofrequency Harmonic Embedded in Magnetic Resonance Sounding Signal With an Attention-Based Convolutional Neural Network

The fidelity of magnetic resonance sounding (MRS) signals is severely affected by the multi-source noise. Among them, the powerline harmonic noise is particularly serious. A series of traditional methods have been proposed to suppress the harmonic noise and perform well if the optimum parameters are designed or the reference loop has been introduced. However, the performance gets worse when one of the harmonics has a frequency close to or coincident with the Larmor frequency, referred to as a co-frequency harmonic. Several intelligent algorithms have been proposed to recover the MRS signal, but cannot remove the co-frequency harmonic. To solve the problem, the manuscript proposes a novel co-frequency harmonic suppression algorithm utilizing an attention-based convolutional neural network. An attention mechanism is first introduced to suppress harmonic noise components outside the Larmor frequency of 2.5 Hz through adaptive weight allocation. Numerical simulations demonstrate that compared with the commonly used methods, the proposed method can remove the co-frequency harmonic from the noisy signal recorded by one coil without distorting the ideal MRS signal. Furthermore, the effectiveness of the proposed method in actual data processing is proved by field experiments.