Deep-DSP²: Cross-Domain Deep Learning Direct MR Signal Prediction for RF Shielding-Free MRI

To develop a deep learning approach to electromagnetic interference (EMI) elimination in the presence of dynamically varying electromagnetic coupling relationships (i.e., spectral domain transfer functions) between MRI receive and EMI sensing coils for RF shielding-free ultra-low-field (ULF) MRI.

Following the recently developed active EMI sensing and deep learning direct MR signal prediction (Deep-DSP), the proposed approach, named Deep-DSP², utilizes both time and spectral domains to form a cross-domain framework to enhance the exploitation of signal characteristics, thus providing better differentiation of MR and EMI signals. Deep-DSP² also utilizes a transformer-based U-Net architecture that integrates transformer blocks for global modeling and convolution blocks for local feature extraction. Deep-DSP² was quantitatively evaluated using both simulated and experimental data sets from a shielding-free 0.055T brain MRI scanner where the transfer functions were varied during scanning due to subject motion or arbitrarily changing EMI source location.

The Deep-DSP² consistently outperformed Deep-DSP, CNN, and analytical transfer function methods for retrospective elimination of strong narrowband or broadband EMI signals. More importantly, Deep-DSP² removed EMI signals most effectively when the subject motioned or EMI source location changed dynamically during scanning.

Deep-DSP² provides more robust EMI elimination than the original Deep-DSP and other existing methods, especially when the transfer functions dynamically vary. In point-of-care MRI, patient motion and EMI source location change during scanning can alter the intrinsic transfer functions between MRI receive and EMI sensing coils. Therefore, Deep-DSP² is expected to be more suitable for portable ULF MRI applications in diverse settings.