[Paper Review] XPDNet for MRI Reconstruction: an application to the 2020 fastMRI challenge

Zaccharie Ramzi, Philippe Ciuciu+1 more|arXiv (Cornell University)|Oct 15, 2020

Medical Imaging Techniques and ApplicationsMedicine19 citations

TL;DR

XPDNet is a deep learning-based MRI reconstruction network that unrolls the Primal-Dual Hybrid Gradient (PDHG) algorithm, integrating a multi-scale wavelet CNN for image denoising and a U-Net for sensitivity map refinement. It achieved second place in the 2020 fastMRI challenge for both 4× and 8× acceleration factors, demonstrating state-of-the-art performance in PSNR and NMSE across brain and knee MRI datasets.

ABSTRACT

We present a new neural network, the XPDNet, for MRI reconstruction from periodically under-sampled multi-coil data. We inform the design of this network by taking best practices from MRI reconstruction and computer vision. We show that this network can achieve state-of-the-art reconstruction results, as shown by its ranking of second in the fastMRI 2020 challenge.

Motivation & Objective

• To develop a learnable, efficient, and accurate MRI reconstruction method that improves upon traditional compressed sensing and parallel imaging techniques.
• To address the limitations of handcrafted components in MRI reconstruction, such as regularization, optimization parameters, and sensitivity map estimation.
• To design a modular neural network architecture that enables end-to-end learning while maintaining interpretability and computational efficiency.
• To achieve high reconstruction quality at high acceleration factors (4× and 8×) using multi-coil undersampled k-space data.

Proposed method

• XPDNet is a cross-domain unrolled network that alternates between image space and k-space measurements, inspired by the PDHG optimization algorithm.
• It uses a learnable buffer of 5 previous iterates to adaptively tune optimization parameters such as step sizes and momentum.
• The image correction network is a lightweight Multi-scale Wavelet CNN (MWCNN) with reduced filters to allow more unrolled iterations within memory constraints.
• A U-Net-based sensitivity map refiner is introduced to refine coil sensitivity maps initially estimated from low-frequency k-space data.
• No k-space correction network is used to reduce memory footprint, relying instead on the image-domain denoiser and data consistency in the image space.
• The network is trained with a compound loss combining L1 and multi-scale SSIM, using RAdam optimizer over 100 epochs per acceleration factor, with fine-tuning on contrast-specific data.

Experimental results

Research questions

• RQ1 Can a learned unrolling of the PDHG algorithm outperform traditional optimization-based MRI reconstruction in terms of speed and accuracy?
• RQ2 How does the integration of a multi-scale wavelet CNN improve image reconstruction quality compared to standard CNNs in the context of MRI?
• RQ3 To what extent can sensitivity map refinement via a lightweight U-Net improve reconstruction fidelity without increasing computational cost?
• RQ4 Does the absence of a k-space correction network compromise performance, and if so, how is image quality preserved through other components?
• RQ5 Can a modular, unrolled architecture generalize across different MRI contrasts and acceleration factors while maintaining state-of-the-art performance?

Key findings

• XPDNet achieved second place in the fastMRI 2020 challenge for both 4× and 8× acceleration factors, ranking second in PSNR across all contrasts.
• For 4× acceleration, the average PSNR was 41.56 (T1), 40.68 (T2), 39.60 (FLAIR), and 42.53 (T1POST), with SSIM values above 0.93.
• At 8× acceleration, the average PSNR was 38.57 (T1), 37.41 (T2), 36.81 (FLAIR), and 38.90 (T1POST), with SSIM values ranging from 0.9086 to 0.9517.
• The method demonstrated minimal visual differences from ground truth at 4× acceleration, with only subtle smoothing visible at 8×, indicating robustness to high undersampling.
• The network achieved state-of-the-art performance in PSNR and NMSE across both brain and knee MRI datasets at 4× and 8× acceleration factors.
• Despite strong PSNR results, the method underperformed on the SSIM metric, suggesting potential trade-offs in perceptual quality that warrant further investigation.