NoSENSE: Learned unrolled cardiac MRI reconstruction without explicit sensitivity maps

TL;DR

A novel learned image reconstruction method for accelerated cardiac MRI with multiple receiver coils based on deep convolutional neural networks (CNNs) and algorithm unrolling that implicitly captures and learns to exploit the inter-coil relationships of the images.

Abstract

We present a novel learned image reconstruction method for accelerated cardiac MRI with multiple receiver coils based on deep convolutional neural networks (CNNs) and algorithm unrolling. In contrast to many existing learned MR image reconstruction techniques that necessitate coil-sensitivity map (CSM) estimation as a distinct network component, our proposed approach avoids explicit CSM estimation. Instead, it implicitly captures and learns to exploit the inter-coil relationships of the images. Our method consists of a series of novel learned image and k-space blocks with shared latent information and adaptation to the acquisition parameters by feature-wise modulation (FiLM), as well as coil-wise data-consistency (DC) blocks. Our method achieved PSNR values of 34.89 and 35.56 and SSIM values of 0.920 and 0.942 in the cine track and mapping track validation leaderboard of the MICCAI STACOM CMRxRecon Challenge, respectively, ranking 4th among different teams at the time of writing. Code will be made available at https://github.com/fzimmermann89/CMRxRecon

Figures (3)

• Figure 1: Schematic overview of the proposed method. It applies U-Net blocks to the multi-coil image and $k$-space data and uses intermediate (coil-wise) data-consistency blocks and a final RSS reconstruction. Note the information sharing between different iterations via the hidden latent variables $h_k$ ($k$-space) and $\tilde{h}_k$ (image-space). Both U-Nets are conditioned on $\mathbf{c}_k \in \mathbb R^{192}$, in which we encode information about the axis, slice, iteration, and acceleration factor via a MLP (details in the main text). In each iteration, the DC block performs \ref{['eq:minsolution']} with $\lambda^c_k$ obtained from $\mathbf{c}_k$ by a learned affine mapping.
• Figure 2: Detailed view of one of the U-Nets used within our reconstruction network as $X_\theta$ (at 4 resolution scales) and $Y_\theta$ (3 resolution scales). The residual blocks employed as encoder/decoder apply FiLM conditioning film and separate spatial/temporal convolutions. We introduced MGU-inspired mguLatentGU blocks in the skip connections at different resolutions for sharing information between different iterations.
• Figure 3: An example of an LAX image reconstructed with the reported methods of comparison and our proposed approach at $R=8$ and corresponding amplified point-wise absolute error images (bottom) as well as the fully-sampled ground truth and the zero-filled RSS reconstruction. The example is in the official train split of the CMRxRecon challenge, but has not been used for training.