An edge guided cascaded U-net approach for accelerated magnetic resonance imaging reconstruction

Magnetic resonance imaging, despite of its significant role in today's healthcare, suffers from long image acquisition time which leads to patient discomfort and cost increment. Compressive sensing magnetic resonance imaging, where clinically acceptable images are reconstructed using partially sampled k-space data, is one possible approach to mitigate this problem. The recent evolution in compressive sensing magnetic resonance imaging field is the model based deep learning approach, which is comprised of cascaded convolutional neural network based denoizer and data consistency layer. In this paper, we propose an edge guided model based deep learning approach employing U-net module as an artifact removal unit. The proposed model contains cascaded U-net architectures with interleaved data consistency layer. To effectively retain the fine details in the reconstructed output, along with the image, edge maps of the image were also applied at the input of each stage in the cascaded structure and the edge map loss was incorporated in the objective function along with the pixel loss. Experiments were performed on MR-PD and MR-T1 images with different sampling patterns. Qualitative and quantitative comparison of the results obtained with the proposed method with other model based and deep learning methods validates the superiority of the proposed method in reconstructing high quality magnetic resonance images.