Proposal: Integrating Image Reconstruction Algorithms

[mersad95zd]

This is a proposal from the MONAI developers and the DLMED group.

Overview

Goal
A very active line of research that is currently missing in MONAI is medical image reconstruction. As an interesting subfield that has gained a lot of attention recently, MRI reconstruction is proposed to be added. A major problem in this field is compressed sensing MRI whose purpose is to enable accelerating the MRI process by recording few measurements, and then recovering the high-quality MRI scan from those under-sampled measurements.

Proposed solution
Add baseline and state-of-the-art accelerated MRI reconstruction methods with all their dependencies to MONAI.

Data details

Dataset
fastMRI dataset which contains single- and multi-coil measurements of brain and knee anatomies.

Data location
fastmri.org

Data description
fastMRI contains 1398 knee as well as 7002 brain volumes that are fully sampled. The dataset also comes with DICOM images. fastMRI was created to advance the state-of-the-art for accelerated MRI reconstruction. In order to use fastMRI for accelerated
MRI reconstruction, under-sampling masks should be applied to fastMRI samples and the result will be fed to reconstruction models. The model is then expected to learn to generated the fully-sampled data (either in the measurement or the image domain).

Data Access
MONAI users should request access through fastmri.org.

Model specifics

We plan to add the following reconstruction models to MONAI:

• Baseline U-Net which takes in the inverse Fourier transform of the under-sampled data and learns to output the inverse Fourier
  transform of the fully-sampled data (ref: https://ieeexplore.ieee.org/iel7/83/4358840/07949028.pdf).
• State-of-the-art e2e VarNet which takes in the under-sampled data and learns to output the fully-sampled data (thus, e2e VarNet, unlike U-Net,
  operates in the k-space domain) (ref:https://arxiv.org/abs/2004.06688).
• SSDU which is a self-supervised training framework for accelerated MRI reconstruction (note that U-Net and e2e VarNet were initially
  proposed to be trained in a fully supervised manner) (ref: https://arxiv.org/abs/1910.09116).

Major components

To add this feature, several major components need to be added:

• Fourier utilities
• MRI utilities
• fastMRI data loader and transforms
• Reconstruction loss functions (a common choice is the SSIM loss function)
• Model implementations (e.g., the e2e VarNet is not available in MONAI)
• Training demos