Deep Image Prior with Structured Sparsity (Discus) for Dynamic MRI Reconstruction

Muhammad Ahmad Sultan, Chong Chen, Yingmin Liu, Xuan Lei, Rizwan AhmadISBI 2024 · DOI

TL;DR

This paper introduces DISCUS, a self-supervised deep learning method for reconstructing dynamic MRI images, which are often hampered by limited training data. DISCUS leverages a "deep image prior" combined with a novel "structured sparsity" constraint to automatically discover the underlying low-dimensional manifold of temporal variations, outperforming existing methods in image quality.

What Problem Does This Paper Address?

Dynamic MRI, crucial for capturing physiological processes like heartbeats or breathing, faces a significant challenge: acquiring fully sampled, high-quality data for training deep learning models is often impractical due to long scan times. This leads to a scarcity of training data, especially for dynamic scans. While methods like Compressed Sensing (CS) and Deep Image Prior (DIP) exist, they have limitations. CS relies on hand-crafted priors and can struggle with complex dynamics, while standard DIP requires precise knowledge of the underlying data manifold's dimensionality, which is often unknown or variable in dynamic MRI. This paper aims to bridge this gap by proposing a self-supervised method that can reconstruct dynamic MRI without requiring fully sampled training data and, crucially, can automatically infer the manifold's dimensionality.

Methods

The core of DISCUS builds upon the Deep Image Prior (DIP) framework. DIP uses a convolutional neural network (CNN) as an implicit prior, where the network's structure itself guides the reconstruction by mapping a random input code vector to an image. For dynamic MRI, DISCUS extends this by using a CNN generator that takes both a static code vector (common across all frames) and a dynamic code vector (specific to each frame) to reconstruct a series of images.

The key innovation in DISCUS is the introduction of structured sparsity applied to the dynamic code vectors. This is achieved using a hybrid $l_2$-$l_1$ norm. This norm encourages group sparsity, meaning that the set of non-zero entries in the dynamic code vectors should be similar across all frames. This similarity in sparsity patterns effectively forces the model to discover the underlying low-dimensional manifold that describes the temporal variations. Importantly, unlike previous manifold learning approaches, DISCUS does not require the user to pre-specify the manifold's dimensionality; it is learned directly from the data through this sparsity constraint.

Datasets

The paper evaluates DISCUS using:

  • Simulated 2D Shepp-Logan phantom: Used to demonstrate the method's ability to discover manifold dimensionality under random rotations and translations.
  • Simulated LGE phantom (MRXCAT): A realistic phantom simulating late gadolinium enhancement (LGE) scans, used to compare DISCUS with CS, L+S, and DIP. This dataset incorporated breathing motion and simulated myocardial scars.
  • Measured LGE patient data: Free-breathing, single-shot LGE data from five patients, acquired at different acceleration rates (R=2, 3, 4). This dataset was used to evaluate DISCUS in a more clinically relevant scenario.

Key Insights and Ideas

  • Automatic Manifold Dimensionality Discovery: The structured sparsity constraint is a powerful mechanism for automatically inferring the dimensionality of the underlying temporal manifold without prior knowledge. This is a significant advantage over methods that require manual specification.
  • Robustness to Unknown Dynamics: DISCUS's approach of not assuming temporal closeness implies image similarity makes it suitable for applications like free-breathing single-shot imaging, where consecutive frames might not be perfectly correlated.
  • Group Sparsity as a Prior: The paper highlights that enforcing group sparsity on code vectors is crucial for discovering the true manifold dimensionality and leads to performance gains, as demonstrated by comparing DISCUS with DISCUS-GS (without group sparsity).
  • Self-Supervised Learning for Dynamic MRI: DISCUS offers a viable self-supervised solution for dynamic MRI reconstruction, alleviating the need for large, fully sampled training datasets.

Key Results

  • Phantom Studies: DISCUS successfully discovered the correct manifold dimensionality (1 for rotations only, 1 for translations only, and 2 for combined rotations and translations) in the Shepp-Logan phantom study. It significantly outperformed CS, L+S, and DIP in terms of Normalized Mean Squared Error (NMSE) and Structural Similarity Index (SSIM).
  • Simulated LGE Data: DISCUS achieved superior NMSE and SSIM compared to CS, L+S, and DIP. Removing the group sparsity constraint (DISCUS-GS) resulted in a noticeable performance drop.
  • Patient LGE Data: At acceleration rates R=2, 3, and 4, DISCUS consistently offered a significant advantage in NMSE over CS and L+S. While L+S marginally outperformed DISCUS in SSIM at higher acceleration rates, DISCUS provided better preservation of fine details and reduced artifacts, particularly around myocardial scars.

Limitations and Future Work

  • Computational Cost: While trained on a single GPU, the optimization process for DISCUS can still be computationally intensive, especially for longer dynamic series.
  • Future Work: The authors suggest extending DISCUS to prospectively undersampled data and applying it to dynamic MRI applications beyond LGE. Further exploration into optimizing the network architecture and training procedures could also be beneficial.

Summary updated

BibTeX
@inproceedings{sultan2024deep,
  author = {Muhammad Ahmad Sultan and Chong Chen and Yingmin Liu and Xuan Lei and Rizwan Ahmad},
  title = {{Deep Image Prior with Structured Sparsity (Discus) for Dynamic MRI Reconstruction}},
  booktitle = {ISBI},
  year = {2024},
  doi = {10.1109/isbi56570.2024.10635579},
}
Figure