Skip to main content
2025 SCMR Conference Main banner
Final Program


Schedule at a Glance
Back
On-Demand Video

CMR Innovations

CMR Innovations: Open-Source CMR Software 1

TensorFlow MRI: A TensorFlow Library for CMR Applications

Thursday, January 30, 2025
2:00 PM – 2:15 PM East Coast USA Time
Location: Empire Ballroom
Open-source Software Description:

Introduction





Open-source reconstruction software is vital for ongoing development of accelerated CMR. To enable rapid reconstructions this software must support GPU operations, and the increasing use of Deep Learning (DL) requires packages that also incorporate TensorFlow or PyTorch. Thus, we present TensorFlow MRI, an open-source library of TensorFlow operators for rapid computational MRI, focussed on reconstruction (DL and non-DL) and post-processing.



TensorFlow MRI enables fast prototyping of novel MRI applications. The library is primarily written in Python and runs on heterogenous and distributed systems. It contains optimised operators for:




  • Fast, non-uniform fast Fourier transform (see ‘TensorFlow NUFFT’)

  • Non-Cartesian trajectories and sampling density estimation

  • Unified MRI reconstruction for Cartesian and non-Cartesian parallel imaging (PI), compressed sensing (CS) and DL

  • Common linear and nonlinear operators, e.g. Fourier operators and regularizers

  • Multicoil imaging operators, e.g. coil combination, coil compression and estimation of sensitivity maps

  • Image processing operations, e.g. filtering, gradients and image quality assessment

  • Keras objects including models, layers, metrics

  • Loss functions and metrics for classification, segmentation and image restoration





TensorFlow MRI integrates seamlessly in PI, CS and DL, and enables interfacing between these techniques. It has been successfully deployed using Gadgetron for seamless on-line use.



TensorFlow NUFFT





TensorFlow NUFFT is a highly efficient non-uniform fast Fourier transform (NUFFT); often used repeatedly within non-Cartesian MRI reconstructions. We have adapted the FINUFFT library [6, 7] for seamless use within TensorFlow MRI.



Performance





We compared an iterative conjugate gradient SENSE reconstruction from TensorFlow MRI to BART [8], finding similar CPU performance, but 7x faster GPU performance of TensorFlow MRI (Figure 1a)



The performance of the TensorFlow NUFFT operator was compared to BART and TF KB-NUFFT [9]. TensorFlow NUFFT was 4x faster than TF KB-NUFFT (Figure 1b).



Applications for Cardiac MRI 





TensorFlow MRI has been used for many CMR applications, including DL reconstruction for real-time cardiac imaging [1][2][3] and sodium imaging [4][5]. See figure 2 and 3.



Conclusion





TensorFlow MRI is an efficient, easy-to-use, open-source library that contains common computational MRI operators and provides a unified computing framework for modern MRI processing.



TensorFlow MRI provides rapid reconstructions, which we have leveraged to provide low-latency online DL reconstruction. We believe that TensorFlow MRI has significant potential to improve CMR research by simplifying image reconstruction and post-processing of cardiac MRI data, both in children and adults.


Open-source software access instructions:
TensorFlow MRI releases can be easily downloaded and installed using “pip”:



$ pip install TensorFlow-MR



(source code: https://github.com/mrphys/TensorFlow MRI, documentation: https://mrphys.github.io/TensorFlow MRI). 



TensorFlow NUFFT is released as a separate package (source code: https://github.com/mrphys/TensorFlow NUFFT, documentation: https://mrphys.github.io/TensorFlow NUFFT).



$ pip install TensorFlow-NUFFT 



Open-Source Applications


[2]: https://github.com/mrphys/HyperSLICE.git
[3]: https://github.com/mrphys/Image_Reconstruction_Inter4k.git
[4]: https://github.com/mrphys/sodium_MRI_DnCNN.git