[Paper Review] MRI Reconstruction via Cascaded Channel-wise Attention Network

Qiaoying Huang, Dong Yang+4 more|arXiv (Cornell University)|Oct 18, 2018

Sparse and Compressive Sensing Techniques12 references5 citations

TL;DR

This paper proposes MICCAN, a cascaded U-Net with channel-wise attention for MRI reconstruction from highly undersampled k-space data. By integrating channel-wise attention to emphasize salient features, long skip connections to preserve low-frequency content, and a combined loss for perceptual fidelity, MICCAN achieves state-of-the-art performance, reducing NRMSE to 0.0385 and PSNR to 28.45 at 12.5% undersampling rate.

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.

Motivation & Objective

• To address the challenge of reconstructing high-quality MRI from very low undersampled k-space data, where traditional deep learning methods underperform due to equal treatment of channel-wise features.
• To improve representation capability of deep networks in MRI reconstruction by introducing channel-wise attention that filters irrelevant and noisy features.
• To enhance reconstruction quality by preserving low-frequency information through a long skip connection and focusing the network on learning high-frequency details.
• To develop a combined loss function that improves perceptual quality and reduces blurriness compared to standard ℓ₂ loss.

Proposed method

• The model uses a modified U-Net architecture with a Channel-wise Attention (UCA) module that adaptively reweights feature channels based on their importance to the final output.
• The UCA module computes channel-wise attention weights via global average pooling and fully connected layers with a squeeze-and-excitation mechanism, enhancing feature representation.
• A long skip connection is introduced to bypass low-frequency components directly to the final output, ensuring structural integrity and allowing the network to focus on high-frequency details.
• A combined loss function is employed, integrating ℓ₂ loss and perceptual loss from VGG-16 features to improve visual quality and reduce blurriness.
• The network is trained end-to-end using PyTorch with a learning rate of 1e-4, decaying by 0.5 every 15 epochs, over 50 epochs with a batch size of 8.
• Data consistency layers are applied after each reconstruction block to enforce fidelity to the acquired k-space data.

Experimental results

Research questions

• RQ1 Can channel-wise attention mechanisms improve feature representation in deep learning-based MRI reconstruction for very low undersampled k-space data?
• RQ2 Does a long skip connection that preserves low-frequency components enhance reconstruction quality by allowing the network to focus on high-frequency details?
• RQ3 Can a combined loss function incorporating both ℓ₂ and perceptual losses reduce image blurriness and improve perceptual quality compared to ℓ₂ loss alone?
• RQ4 How does the proposed MICCAN model compare to state-of-the-art methods in terms of NRMSE, PSNR, and SSIM at 12.5% undersampling rate?

Key findings

• MICCAN with combined loss (MICCAN-B) achieves the best SSIM score of 0.8214, significantly outperforming all baselines and indicating superior structural preservation.
• The MICCAN-C variant achieves the lowest NRMSE of 0.0385 and highest PSNR of 28.4489, demonstrating the effectiveness of combining both long skip connection and combined loss.
• Even MICCAN-A, trained with only ℓ₂ loss, outperforms the state-of-the-art MRN5 model, achieving NRMSE of 0.0402 and PSNR of 28.0664.
• Qualitative analysis shows that MICCAN models produce sharper images with fewer blurring artifacts and better recovery of fine anatomical details compared to traditional methods and prior deep learning models.
• The ablation study confirms that both the channel-wise attention mechanism and the combined loss contribute significantly to performance gains, with the long skip connection further improving results.