MR Image Reconstruction Using Deep Density Priors.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Magnetic resonance (MR) image reconstruction from undersampled data requires prior information to compensate for missing k-space data.
• Deep learning (DL) methods have been used to learn implicit priors from paired undersampled and fully sampled MR images.
• Existing DL approaches are sensitive to variations in undersampling patterns and coil configurations.

Purpose of the Study:

• To develop an unsupervised deep learning approach for MR image reconstruction using variational autoencoders (VAE).
• To create a reconstruction algorithm that utilizes an explicit image prior, decoupling it from the encoding process.
• To eliminate the need for paired training datasets and mitigate sensitivities associated with current DL methods.

Main Methods:

• Proposed an unsupervised deep learning framework utilizing variational autoencoders (VAE) to learn the probability distribution of fully sampled MR images.
• Integrated the learned VAE prior as an explicit term within the MR image reconstruction process.
• Evaluated the method on T1-weighted images, multi-coil complex MR data from healthy volunteers, and images containing white matter lesions.

Main Results:

• The VAE-based prior produced visually high-quality MR image reconstructions with low Root Mean Square Error (RMSE) values.
• The algorithm outperformed most alternative reconstruction methods on the evaluated datasets.
• Accurate magnitude and phase reconstruction results were achieved for multi-coil complex data, and lesions were faithfully reconstructed.

Conclusions:

• The proposed unsupervised DL method effectively reconstructs MR images using an explicit VAE prior, compensating for undersampled k-space data.
• This approach offers a robust alternative to paired-data DL methods, avoiding associated sensitivities and improving reconstruction quality.
• The VAE prior demonstrates significant potential for enhancing MR image reconstruction across various imaging scenarios.