[Paper Review] Adaptive-CS-Net: FastMRI with Adaptive Intelligence

Nicola Pezzotti, Elwin de Weerdt+10 more|arXiv (Cornell University)|Dec 13, 2019

Sparse and Compressive Sensing TechniquesEngineering7 references17 citations

TL;DR

Adaptive-CS-Net proposes a deep learning-based MR image reconstruction method that integrates domain-specific prior knowledge—such as data consistency, phase behavior, and background anatomy—into an iterative compressed sensing framework using UNet-shaped networks. By refining reconstruction at each step with learnable, adaptive components, it achieves state-of-the-art performance, winning the multi-coil 8x accelerated track in the FastMRI challenge.

ABSTRACT

Adaptive intelligence aims at empowering machine learning techniques with the extensive 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. Our results show that an adaptive intelligence approach performs better than traditional methods as well as deep learning methods that do not take prior knowledge into account.

Motivation & Objective

• To improve MR image reconstruction speed and quality by integrating domain-specific prior knowledge into deep learning models.
• To address the limitations of traditional compressed sensing and standard deep learning methods that lack explicit modeling of physical MRI priors.
• To develop an adaptive, iterative reconstruction framework that dynamically corrects reconstruction errors using learned priors.
• To achieve high acceleration factors (up to 10x) while preserving image structural and contrast fidelity.
• To demonstrate superior performance on real-world clinical MRI data, particularly in challenging high-acceleration scenarios.

Proposed method

• The method uses an iterative learning-based reconstruction inspired by compressed sensing, with each step refined by a UNet-shaped deep neural network.
• At each iteration, the network applies a learnable soft thresholding function to feature maps, enabling adaptive correction of reconstruction errors.
• The network inputs include the current image estimate and domain-specific priors: data consistency (eb,k), phase behavior (eϕ,k), and background location (ebg,k).
• A 2.5D learning strategy reconstructs the target slice and its neighbors simultaneously to exploit inter-slice correlations.
• The loss function combines Multiscale-SSIM (α=0.84) and L1 norm to preserve structural details and luminance, with supervision applied only on the center slice.
• The model is trained using RAdam optimizer with a decaying learning rate of 10−4 and batch size of 6 on two RTX 8000 GPUs.

Experimental results

Research questions

• RQ1 Can explicit integration of domain-specific priors such as data consistency and phase behavior improve deep learning-based MR reconstruction?
• RQ2 How does adaptive, iterative refinement using learned priors compare to standard end-to-end deep learning in high-acceleration MRI?
• RQ3 To what extent does incorporating inter-slice correlation via 2.5D learning enhance reconstruction quality at high acceleration factors?
• RQ4 Can a learnable, iterative framework with structured priors outperform traditional compressed sensing and non-adaptive deep learning methods?
• RQ5 What is the impact of combining Multiscale-SSIM and L1 loss on preserving structural and intensity fidelity in reconstructed MR images?

Key findings

• Adaptive-CS-Net achieved the best performance in the multi-coil 8x accelerated track of the FastMRI challenge, outperforming all other methods.
• The model was trained and tested on both single- and multi-coil datasets with acceleration factors ranging from 2x to 10x, demonstrating robustness across scales.
• The use of domain priors—data consistency, phase behavior, and background location—enabled a self-correcting mechanism that improved reconstruction accuracy.
• The 2.5D learning approach effectively leveraged inter-slice correlations, enhancing reconstruction quality without increasing computational cost per slice.
• The combination of Multiscale-SSIM and L1 loss (α=0.84) effectively preserved structural details and luminance, reducing uniform biases in reconstructed images.
• The model achieved high-quality reconstructions at 8x and 10x acceleration, confirming its capability for ultra-fast MRI with diagnostic image quality.