TL;DR
This paper introduces a novel deep learning approach for Magnetic Resonance (MR) image reconstruction that leverages Variational Autoencoders (VAEs) to learn an explicit prior distribution of MR images. This method achieves high-quality reconstructions from undersampled data without requiring paired training datasets, outperforming existing techniques in accuracy and robustness.
What Problem Does This Paper Address?
MR imaging is a powerful diagnostic tool, but acquiring high-quality images requires a significant amount of data, leading to long scan times. To reduce scan times, researchers undersample the k-space data (the raw data acquired by the MRI scanner) and then use reconstruction algorithms to infer the missing information. Traditional methods rely on hand-crafted priors (like sparsity or total variation) to guide this reconstruction. More recently, deep learning (DL) has shown promise by learning mappings from undersampled to fully sampled images. However, these DL methods often require paired datasets (undersampled and fully sampled images) for training and can be sensitive to variations in acquisition parameters between training and testing. This paper addresses the challenge of developing a robust and accurate MR image reconstruction method that can compensate for undersampling without relying on paired training data or being susceptible to variations in acquisition parameters.
Methods
The core of this paper is the Deep Density Prior (DDP) reconstruction method. It works by:
- Learning an Explicit Image Prior with VAEs: Instead of learning a direct mapping, the authors train a Variational Autoencoder (VAE) in an unsupervised manner on fully sampled MR images. The VAE learns to approximate the probability distribution of MR image patches. This learned distribution acts as an explicit prior, encoding the expected structure and appearance of MR images.
- Bayesian Reconstruction Framework: The DDP method formulates MR reconstruction as a Maximum-A-Posteriori (MAP) estimation problem within a Bayesian framework. The objective function combines a data consistency term (ensuring the reconstructed image matches the acquired k-space data) and a regularization term derived from the VAE's learned prior.
- Decoupling Prior and Data Consistency: A key innovation is the complete separation of the prior learning (VAE) from the data consistency enforcement. This decoupling is crucial for the method's robustness.
- Projection Onto Convex Sets (POCS): The optimization problem is solved using a POCS algorithm, which iteratively projects the solution onto convex sets defined by the data consistency and the learned prior. The prior projection involves gradient ascent steps on the VAE's ELBO (Evidence Lower Bound).
Datasets
- Training Data: T1-weighted structural images from the Human Connectome Project (HCP) dataset (158 subjects, 790 images). These were used to train the VAE prior.
- Evaluation Data:
- T1-weighted images from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset (subjects with Alzheimer's disease and white matter lesions).
- Newly acquired multi-coil complex MR images from 8 healthy volunteers.
- Additional images from the HCP dataset (17 subjects, separate from training).
The paper does not introduce new benchmark datasets but evaluates on publicly available and newly acquired data to demonstrate the method's versatility.
Key Insights and Ideas
- Unsupervised Prior Learning is Powerful: Using VAEs to learn an explicit, unsupervised prior distribution of MR images offers a powerful way to regularize reconstruction, outperforming methods that rely on hand-crafted priors or implicit DL mappings.
- Decoupling is Key to Robustness: Separating the prior learning from the data consistency term makes the reconstruction method insensitive to variations in undersampling patterns, coil settings, and other acquisition parameters between training and testing. This is a significant advantage over many existing DL methods.
- Generalization to Unseen Structures: The VAE prior, trained on healthy brain images, was able to reconstruct images with white matter lesions, demonstrating that the learned prior captures fundamental structural information that extends to variations like lesions.
- Magnitude and Phase Reconstruction: The DDP method successfully reconstructs both the magnitude and phase information of MR images, which is crucial for many advanced MR techniques.
Key Results
- Quantitative Accuracy: The DDP method achieved the lowest Root-Mean-Squared-Error (RMSE) for full Field of View (FOV) reconstructions on HCP data, and was second best for cropped FOV reconstructions. It also performed equally or better than other methods in terms of Contrast-to-Noise Ratio (CNR).
- Visual Quality: Reconstructions were visually high quality, faithfully reconstructing structures and lesions, with significantly fewer artifacts compared to many competing methods.
- Robustness to Acquisition Variations: Experiments showed that ADMM-Net (a DL method) was sensitive to differences in undersampling patterns between training and testing, while DDP demonstrated robustness.
- Lesion Reconstruction: The method successfully reconstructed white matter lesions in ADNI images, even though the training data did not contain lesions.
Limitations and Future Work
- Training Data Requirement: Like all DL methods, DDP requires high-quality training data. For sequences like functional or diffusion MRI, constructing appropriate training sets is an area for future research.
- Computational Cost: While efficient, the current implementation involves iterative computations and network evaluations, which can be time-consuming. Optimizing the code and network architecture is suggested.
- Phase Prior: The current method uses a theoretical or no explicit prior for the phase component of MR images. Developing better priors for phase reconstruction is a promising direction.
- Domain Adaptation: While DDP shows good generalization, further research into domain adaptation methods could improve performance when contrast or resolution differ substantially from training images.
- 3D Reconstruction: The current work focuses on 2D slices. Extending the method to 3D reconstruction is a natural next step.
- Complex Patches: The VAE was trained on magnitude image patches. Training a VAE for complex image patches could potentially improve reconstruction quality.
Summary updated