[Paper Review] An Adaptive Intelligence Algorithm for Undersampled Knee MRI Reconstruction

Nicola Pezzotti, Sahar Yousefi+10 more|arXiv (Cornell University)|Apr 15, 2020

Advanced MRI Techniques and Applications4 citations

TL;DR

This paper presents Adaptive-CS-Net, a deep learning-based MRI reconstruction method that integrates compressed sensing principles with a multi-scale, multi-block neural network to reconstruct undersampled knee MRI data. Trained on the 2019 fastMRI challenge dataset, it achieved top performance—ranking #1 (shared) on 8x and 4x multi-coil, and #3 on 4x single-coil reconstruction—demonstrating superior image quality and robustness across acceleration levels.

ABSTRACT

Adaptive intelligence aims at empowering machine learning techniques with the additional use of domain knowledge. In this work, we present the application of adaptive intelligence to accelerate MR acquisition. Starting from undersampled k-space data, an iterative learning-based reconstruction scheme inspired by compressed sensing theory is used to reconstruct the images. We adopt deep neural networks to refine and correct prior reconstruction assumptions given the training data. The network was trained and tested on a knee MRI dataset from the 2019 fastMRI challenge organized by Facebook AI Research and NYU Langone Health. All submissions to the challenge were initially ranked based on similarity with a known groundtruth, after which the top 4 submissions were evaluated radiologically. Our method was evaluated by the fastMRI organizers on an independent challenge dataset. It ranked #1, shared #1, and #3 on respectively the 8x accelerated multi-coil, the 4x multi-coil, and the 4x single-coil track. This demonstrates the superior performance and wide applicability of the method.

Motivation & Objective

• To accelerate MRI acquisition by reconstructing undersampled k-space data while preserving diagnostic image quality.
• To address the challenge of aliasing and blurriness in accelerated MRI by combining deep learning with compressed sensing principles.
• To improve reconstruction accuracy and robustness across varying acceleration factors (4x, 8x) and coil configurations (single- and multi-coil).
• To develop a method that learns from data while incorporating domain-specific priors such as data consistency and phase behavior.
• To achieve state-of-the-art performance in the 2019 fastMRI challenge, particularly in clinically relevant multi-coil settings.

Proposed method

• The method employs an iterative learning-based reconstruction framework inspired by compressed sensing, using a deep neural network to refine initial reconstructions.
• A multi-scale, multi-block architecture with residual connections enables the network to learn complex image patterns across multiple resolution levels.
• The network is trained using a loss function combining mean structural similarity (MSSIM) and ℓ₁ norm to balance perceptual quality and sparsity.
• A 2.5D training scheme reconstructs three adjacent slices, with loss calculated only on the central slice to improve feature learning while maintaining efficiency.
• The model incorporates data consistency constraints and known phase behavior as inductive biases, though these contributed only marginally to performance.
• Optimization is performed using the RAdam optimizer, selected for its superior convergence and stability during training on large-scale MRI data.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on undersampled k-space data outperform traditional compressed sensing and existing deep learning baselines in MRI reconstruction?
• RQ2 How does the inclusion of neighboring slices in a 2.5D training scheme affect reconstruction quality compared to 2D or full 3D training?
• RQ3 To what extent do domain-specific priors—such as data consistency and phase behavior—improve reconstruction performance when integrated into the network?
• RQ4 What is the optimal network depth (number of blocks) and architecture for balancing performance, memory usage, and training time on MRI data?
• RQ5 How well does the model generalize across different acceleration factors (4x, 8x) and coil configurations (single- vs. multi-coil) in clinical MRI settings?

Key findings

• Adaptive-CS-Net achieved the highest rank in the 2019 fastMRI challenge, securing #1 (shared) on the 8x and 4x multi-coil reconstruction tracks.
• On the 4x single-coil track, the method ranked #3, demonstrating strong generalization across different acquisition protocols.
• The 2.5D training scheme, which uses neighboring slices for feature learning while computing loss only on the central slice, outperformed both 2D and full 3D loss-based approaches.
• The model's performance was maximized by using the maximum number of blocks feasible within GPU memory, with no significant overfitting observed despite extensive training.
• The combination of MSSIM and ℓ₁ loss yielded the best results, outperforming alternative loss functions including SSIM-only or MSE-based losses.
• Despite incorporating priors such as data consistency and phase behavior, their impact on final performance was minimal, indicating the network's strong capacity to learn from data alone.