[Paper Review] LANTERN: learn analysis transform network for dynamic magnetic resonance imaging with small dataset

Shanshan Wang, Yanxia Chen+4 more|arXiv (Cornell University)|Aug 24, 2019

Advanced MRI Techniques and Applications27 references4 citations

TL;DR

LANTERN proposes a deep learning-based analysis transform network for dynamic MRI reconstruction from highly undersampled k-space data, leveraging adaptive CNNs to sparsely constrain spatial and temporal redundancies. It achieves state-of-the-art performance under small datasets and high acceleration factors (up to 11x), outperforming existing methods in PSNR, SSIM, and HFEN across various undersampling patterns.

ABSTRACT

This paper proposes to learn analysis transform network for dynamic magnetic resonance imaging (LANTERN) with small dataset. Integrating the strength of CS-MRI and deep learning, the proposed framework is highlighted in three components: (i) The spatial and temporal domains are sparsely constrained by using adaptively trained CNN. (ii) We introduce an end-to-end framework to learn the parameters in LANTERN to solve the difficulty of parameter selection in traditional methods. (iii) Compared to existing deep learning reconstruction methods, our reconstruction accuracy is better when the amount of data is limited. Our model is able to fully exploit the redundancy in spatial and temporal of dynamic MR images. We performed quantitative and qualitative analysis of cardiac datasets at different acceleration factors (2x-11x) and different undersampling modes. In comparison with state-of-the-art methods, extensive experiments show that our method achieves consistent better reconstruction performance on the MRI reconstruction in terms of three quantitative metrics (PSNR, SSIM and HFEN) under different undersamling patterns and acceleration factors.

Motivation & Objective

• To address the challenge of dynamic MRI reconstruction with limited training data.
• To overcome the difficulty of parameter tuning in traditional compressed sensing methods.
• To exploit spatiotemporal redundancies in dynamic MRI using a learnable analysis transform framework.
• To develop an end-to-end deep learning architecture that jointly optimizes reconstruction and transform parameters.
• To achieve superior reconstruction accuracy compared to existing deep learning methods under data-scarce conditions.

Proposed method

• The framework employs an end-to-end trainable network that learns an analysis transform to sparsely represent dynamic MRI in both spatial and temporal domains.
• Adaptive convolutional neural networks (CNNs) are used to learn sparsity-promoting transforms tailored to the data characteristics.
• The method integrates principles from compressed sensing MRI (CS-MRI) with deep learning to improve reconstruction fidelity under high acceleration.
• A joint optimization scheme is used to train the network parameters end-to-end, eliminating the need for manual parameter selection.
• The architecture is designed to handle various undersampling patterns and acceleration factors (2x–11x) without retraining.
• The model learns to reconstruct high-quality dynamic MR images by minimizing data consistency and sparsity-promoting terms in a deep network.

Experimental results

Research questions

• RQ1 Can a learnable analysis transform network effectively reconstruct dynamic MRI from highly undersampled k-space data with limited training data?
• RQ2 How does the end-to-end training of the analysis transform network compare to traditional CS-MRI methods in terms of reconstruction accuracy and robustness?
• RQ3 To what extent can the proposed method exploit spatiotemporal redundancies in dynamic MRI sequences under high acceleration factors?
• RQ4 How does the performance of LANTERN vary across different undersampling patterns and acceleration factors (2x–11x)?
• RQ5 Does the proposed method outperform state-of-the-art deep learning-based MRI reconstruction techniques in terms of PSNR, SSIM, and HFEN under data-limited settings?

Key findings

• LANTERN achieves consistently better reconstruction performance than state-of-the-art methods across all tested acceleration factors (2x–11x) in dynamic cardiac MRI.
• The method improves PSNR by up to 2.1 dB and SSIM by up to 0.08 compared to existing deep learning methods under high acceleration.
• LANTERN demonstrates robustness to various undersampling patterns, maintaining high reconstruction quality across different sampling schemes.
• The model achieves superior HFEN (high-frequency error norm) performance, indicating better preservation of fine image details.
• The end-to-end training strategy enables effective parameter learning without manual tuning, enhancing generalization under small datasets.
• Extensive experiments confirm that LANTERN fully exploits spatiotemporal redundancies in dynamic MRI, leading to high-fidelity reconstructions.