Abstract for this paper: High-resolution magnetic resonance imaging (MRI) scans are important to provide precise information about imaged tissues. Thus, we need to find the method for image super-resolution. Traditional methods based on end-to-end deep learning have to be retrained when the distribution of input shifts. This paper proposed a new method to increase the resolution of clinical MRI scans by using the latent diffusion model (LDM) trained on UK BioBank. This model has two novel strategies for SR with different sparsity, finding optimal latent code to map the given LR MRI into HR. Finally, the authors validate the method on over 100 brain T1w MRIs from the IXI dataset.

Methods

Latent Diffusion Models

This paper uses the Latent diffusion model which contains two different components:

  • Autoencoder
    • Encoder maps each high-resolution T1w brain MRI into a latent vector of size
    • Decoder: maps latent vectors back into the MRI image domain .
    • Dataset: 31730 T1w MRIs from the UK Biobank.
    • Loss function: a combination of an L1 loss, a perceptual loss, a patch-based adversarial loss, and a KL regularization term in the latent space
  • Diffusion Model
    • Conditional variables : age, gender, ventricular volume, brain volume
    • DDIM Sampling: use DDIM to predict when inference.
Brain Latent Diffusion Models

InverseSR

Algorithm

Corruption function: The corruption function is used to corrupt the HR image , and compare to the input image I. In clinical practice, a prevalent method for acquiring MR images is prioritizing high in-plane resolution. Thus, this paper uses a corruption function that generates masks for non-acquired slices.

This paper proposes two strategies for different conditions.

  • InverseSR(LDM)

    This strategy is used for SR with high sparsity. They will optimize the noise latent code and conditional variables .

    InverseSR(LDM)
  • InverseSR(Decoder)

    This strategy is used for low sparsity MRI SR, which directly find the optimal latent code using the decoder .

    InverseSR(Decoder)
InverseSR Structure

Results

Qualitative results
Evaluation results

Reference

[1] J. Wang, J. Levman, W. H. L. Pinaya, P.-D. Tudosiu, M. J. Cardoso, and R. Marinescu, ‘InverseSR: 3D Brain MRI Super-Resolution Using a Latent Diffusion Model’, arXiv [eess.IV]. 2023.