[Paper Review] Deep Simultaneous Optimisation of Sampling and Reconstruction for Multi-contrast MRI

Xinwen Liu, Jing Wang+4 more|arXiv (Cornell University)|Mar 31, 2021

Medical Image Segmentation Techniques7 references4 citations

TL;DR

This paper proposes a deep learning framework, AR-net, that jointly optimizes k-space sampling patterns and image reconstruction for multi-contrast MRI by leveraging complementary information from one contrast (e.g., T1WI) to improve the sub-sampling and reconstruction of another (e.g., T2WI). The method achieves superior image quality, with average PSNR >35 dB and SSIM >0.98 at acceleration rates up to 10, outperforming conventional sampling patterns and single-contrast methods.

ABSTRACT

MRI images of the same subject in different contrasts contain shared information, such as the anatomical structure. Utilizing the redundant information amongst the contrasts to sub-sample and faithfully reconstruct multi-contrast images could greatly accelerate the imaging speed, improve image quality and shorten scanning protocols. We propose an algorithm that generates the optimised sampling pattern and reconstruction scheme of one contrast (e.g. T2-weighted image) when images with different contrast (e.g. T1-weighted image) have been acquired. The proposed algorithm achieves increased PSNR and SSIM with the resulting optimal sampling pattern compared to other acquisition patterns and single contrast methods.

Motivation & Objective

• To accelerate MRI scanning by jointly optimizing sampling patterns and reconstruction for multi-contrast images using shared anatomical information.
• To address the limitation of prior methods that focus only on reconstruction while neglecting optimal sampling pattern design.
• To improve image quality and reduce scan time by leveraging complementary contrast information (e.g., T1WI) to guide sub-sampling of the target contrast (e.g., T2WI).
• To develop an end-to-end deep learning framework that learns both the optimal sampling mask and the reconstruction model in a unified manner.
• To demonstrate that joint optimization of acquisition and reconstruction leads to better performance than sequential or single-contrast approaches.

Proposed method

• The proposed AR-net is an integrated network with two components: an acquisition network that learns an optimized sampling pattern in k-space and a reconstruction network based on U-Net for image recovery.
• The acquisition network uses a sigmoid activation with a large slope and ℓ₁-norm sparsity constraint to approximate discrete sampling, producing a probability map of k-space locations to sample.
• The sampling mask is derived via thresholding the learned probability map, resulting in a binary mask with ones at sampled locations and zeros elsewhere.
• The reconstruction network (Recon-net) takes the zero-padded, sub-sampled T2WI and concatenates the corresponding T1WI at the input layer to exploit cross-contrast information.
• The model is trained using mean absolute error (MAE) loss and the Adam optimizer with a learning rate of 0.0005 to minimize reconstruction error.
• The framework is trained on 2D brain MRI slices from the BraTS2019 dataset, with normalization to [0,1], and tested on unseen data at sub-sampling rates up to 10 (22 out of 240 k-space lines).

Experimental results

Research questions

• RQ1 Can leveraging complementary information from one MRI contrast (e.g., T1WI) improve the optimization of sampling patterns for another contrast (e.g., T2WI)?
• RQ2 Does joint optimization of sampling pattern and reconstruction network lead to better image quality than optimizing them separately or using conventional sampling patterns?
• RQ3 How does the inclusion of multi-contrast information in the reconstruction network affect performance at high acceleration rates?
• RQ4 Can the proposed deep learning framework achieve high-quality image reconstruction at acceleration rates up to 10 while maintaining diagnostic fidelity?
• RQ5 How does the proposed method compare quantitatively to standard sampling patterns (e.g., centred, equal-distance, Gaussian) and single-contrast reconstruction baselines?

Key findings

• The proposed method achieves an average PSNR of over 35 dB and an SSIM of over 0.98 at acceleration rates up to 10, demonstrating high-quality image reconstruction.
• The learned sampling pattern significantly outperforms conventional patterns such as low-resolution, equal-distance, and Gaussian masks in terms of PSNR and SSIM.
• Incorporating T1WI as a reference during reconstruction leads to improved performance compared to single-contrast T2WI reconstruction, even at high acceleration rates.
• The error maps for the proposed method show reduced artifacts and noise, indicating more accurate reconstruction with better structural preservation.
• The joint optimization of sampling and reconstruction via AR-net yields superior results compared to sequential optimization or standalone reconstruction methods.
• The method generalizes well to unseen data, with consistent performance across the test set of 2,165 slices from the BraTS2019 dataset.