CMR Deep learning reconstruction

The repository is part of the ISMRM member-initiated tutorial Cardiovascular MR: From Theory to Practice.

Tutorials

The repository hosts some example codes to perform MR image reconstruction with deep learning architectures. The code runs on the MNIST database. Since MNIST only contains real-valued images, phase data is simulated to provide complex-valued inputs. We showcase a comparison of denoising networks (denoising) and unrolled reconstruction networks (physics-based). Pure real-valued processing (real) is compared against complex-valued processing as complex-valued operations (complex) or 2-channel real-valued operations (2chreal). Respective complex-valued operations and data consistency layers are provided.

Please refer to the corresponding book chapter if you are using code of this tutorial:

@incollection{HammernikKuestnerRueckertMLRecon2021,
    title={Machine Learning For MRI Reconstruction},
    author={Kerstin Hammernik and Thomas K\"ustner and Daniel Rueckert},
    booktitle={MRI Reconstruction: Theory, Methods and Applications},
    year={2021 (in Preparation)},
}

• Image Denoising
  • Real-valued denoising (magnitude images)
  • Complex-valued denoising, where real and imaginary part are stored in separate channels.
  • Complex-valued denoising, with complex convolutions and complex activations
• Image Reconstruction with complex convolutions and complex activations
• Complex Activation Functions Workbook and solutions

Challenge: Deep Plug-and-Play Prior for cardiac CINE

Deep plug-and-play priors exploit the advantages of learning an advanced denoising scheme offline and plugging it into any optimization scheme to solve inverse problems, such as Magnetic Resonance Image Reconstruction.

The goal of this challenge is to find a way to reconstruct complex-valued MR images by using available denoisers. The plug-and-play prior challenge consists of two tasks:

1. Deploying a denoising model. If you plan to use available denoisers trained on real-valued images, you have to think about a way to apply them to the complex-valued images.
2. Definition of an optimization scheme. Inspiration can be found in the Suggested Reading section. The code parts that have to be changed are marked with TODO.

The dataset subject1.h5, containing fully sampled k-space, can be used for quantitative and qualitative evaluation in the deployment phase. The results have to be generated for subject2.h5, which contains the undersampled k-space. The challenge will be evaluated quantitatively using the normalized root mean squared error (NRMSE), and based on the creativity and complexity to generate the solution.

We are looking forward to your creative submissions! Happy coding!

Submission [CLOSED]

Deadline: Wednesday 20th of May 2021, 11:59 am UTC

Template: Please fill out the template slide with

• Provide a short abstract (max. 150 words) about your solution
• Figure: Provide an overview figure of your solution

Submission: Please fill out the submission form and upload in this form:

• Template: Containing your abstract and overview figure
• Results: Your Jupyter notebook as *.ipynb containing your solution

Solution

Exemplary solution of challenge code