Parallel MRI reconstruction by using complex convolution and attention mechanism
for the reconstruction of parallel magnetic resonance imaging,a deep complex attention network(DCANet)model is proposed.A data consistency layer is used to maintain acquired k-space data unchanged,and a cascade network is formed.Additionally,since magnetic resonance images of multiple coils differ,the proposed model also uses a channel-wise attention mechanism to focus on channels with more effective features.All of these techniques are used to replace the conventional method of convolving real and imaginary parts separately in complex convolution.Experiments are conducted on two different magnetic resonance imaging datasets with three different undersampling patterns.According to the experimental findings,the DCANet model performs better during reconstruction and achieves lower high-frequency error norm(HFEN)and greater peak signal-to-noise ratio(PSNR)and structural similarity(SSIM).The DCANet model achieves average PSNR improvements of 4.52 dB,2.30 dB and 1.21 dB over MRI cascaded channel-wise attention network(MICCAN),Deepcomplex and DONet models,respectively.
parallel magnetic resonance imagingimage reconstructiondeep learningcomplex convolutional networksattention mechanism
万方数据
维普
