MRI Reconstruction Via Cascaded Channel-Wise Attention Network

Abstract

We consider an MRI reconstruction problem with input of k-space data at a very low undersampled rate. This can practically benefit patient due to reduced time of MRI scan, but it is also challenging since quality of reconstruction may be compromised. Currently, deep learning based methods dominate MRI reconstruction over traditional approaches such as Compressed Sensing, but they rarely show satisfactory performance in the case of low undersampled k-space data. One explanation is that these methods treat channel-wise features equally, which results in degraded representation ability of the neural network. To solve this problem, we propose a new model called MRI Cascaded Channel-wise Attention Network (MICCAN), highlighted by three components: (i) a variant of U-net with Channel-wise Attention (UCA) module, (ii) a long skip connection and (iii) a combined loss. Our model is able to attend to salient information by filtering irrelevant features and also concentrate on high-frequency information by enforcing low-frequency information bypassed to the final output. We conduct both quantitative evaluation and qualitative analysis of our method on a cardiac dataset. The experiment shows that our method achieves very promising results in terms of three common metrics on the MRI reconstruction with low undersampled k-space data. Code is public available 1 .

A Very Deep Densely Connected Network for Compressed Sensing MRI

“…Schlemper et al [25] constructed a deep CS-MRI model by cascading multiple sub-networks and introduced data consistency (DC) layer to correct sub-networks' outputs by considering data fidelity in k-space. Since then, DC layer has become an indispensable component in latest networks [26]- [29] for CS-MRI.…”

“…The function of DC layer is summarized in Figure 1. However, deep networks [21]- [29] constituted of multiple sub-networks have two major limitations. First, the subnetworks update MR images successively using only the latest prediction, not all the former predictions which might be useful to improve the final reconstruction performance.…”

“…A number of CS-MRI methods have been proposed in recent years since the CS theory was developed [1], [2]. These methods can be categorized into two groups: model-based methods [3]- [13] and deep learning methods [14]- [29].…”

“…Recently, deep learning models [14]- [29] have also been adopted to CS-MRI, motivated by the great success of deep learning in many computer vision applications, such as object recognition [30], [31], image super-resolution [32]- [35], and image segmentation [36]- [38]. Generally, deep learning CS-MRI methods use deep convolutional neural networks (CNNs) to learn the relationship between the under-sampled MRI data and the corresponding fully-sampled MR images in training dataset.…”

“…Latest deep learning studies [25]- [29] for CS-MRI focus on learning a deep cascade of multiple sub-networks by introducing data consistency (DC) layer to enforce data fidelity in k-space and improve reconstruction performance. Each sub-network generates a prediction of the reconstructed MR image by using the predicted result of the preceding sub-network.…”