Classification-guided Neural Network-based Correction of Magnetic Resonance-related Gradient Artifact Residuals in Simultaneously Recorded Surface Electromyography.

Spontaneous muscular activities can be studied by simultaneous recordings of surface electromyography (sEMG) and diffusion-weighted magnetic resonance imaging (DW-MRI). For reliable assessment of the spontaneous activity rate in sEMG data during active MR imaging, it is necessary to have a decent gradient artifact (GA) correction algorithm enabling the detection of small spontaneous activities with an amplitude of few microvolts. In this work, a neural network with weak label annotations during the training process is utilized for enhanced correction of GA residuals in the sEMG recordings. Based on sEMG signal decomposition and class-activation maps from the neural network classification, the amount of GA residuals is iteratively decreased in the sEMG signal. This leads to a reduction of the false-positive rate in automated spontaneous activity detection. Quality of GA residual correction is therefore estimated by using a specialized second neural network model. Clinical relevance- This work establishes an improved GA residual correction for simultaneously recorded sEMG data during MRI to enhance the ability for small spontaneous activity detection.