Logo of the Physikalisch-Technische Bundesanstalt

Quantitative MRI

Research group 8.13

Advanced Image Reconstruction

MR raw data are acquired in so-called k-space which is a spatial frequency representation of the imaged object. In order to obtain diagnostic images from the raw data, a direct reconstruction using the Fast Fourier transform can be applied. K-space is often not fully sampled, to speed up MR data acquisition. This can lead to so-called undersampling artefacts. In recent years, several new concepts for image reconstruction of undersampled data have been introduced to MRI, such as parallel imaging, compressed sensing or machine learning, which greatly improve image quality and hence diagnostic accuracy.

Compressed Sensing

The principle of compressed sensing is, that signals can be accurately recovered from fewer acquired samples than required by the Shannon-Nyquist sampling theorem if certain conditions apply. For MRI, these conditions translate to the requirements that undersampling artefacts have to be incoherent, that the object or a transformation of the object is sparse and that a non-linear image reconstruction approach is used such as L1-minimization.

A common transformation to obtain a sparse representation of anatomical MR images is the Wavelet transform which has been used in a wide range of applications. Although this transform works very well for brain MR images, it is not necessarily optimal for cardiac MRI. Novel approaches such as shearlet transforms (www.shearlab.org) lead to a sparser representation of the object and allow for faster MR imaging.

3D high-resolution cardiac MRI. Reference images were reconstructed from 13 min of data acquisition using parallel imaging (non-Cartesian iterative sensitivity encoding). Images with different reconstruction techniques were obtained using only one third of this data.

Parallel imaging (SENSE) cannot fully remove the undersampling artefacts and which leads to blurring of the anatomy. Wavelet-based CS (Wavelet) only achieves a small improvement compared to SENSE. Total-variation regularised reconstruction (TV) can successfully remove undersampling artefacts but exhibits “patchy” regularisation artefacts. Shearlet-based CS (Shearlet) ensures accurate depiction of the anatomy while successfully minimizing undersampling artefacts.

Machine learning

Machine learning using Neural Networks is being used for a wide range of applications in image processing such as image segmentation, object detection or image classification. More recently, medical image reconstruction for CT and MRI has been carried out with Neural Networks.

New machine learning approaches can for example be used in order to improve image quality of functional cardiac MR images showing the changes of the heart during the cardiac cycle. These acquisitions are commonly acquired during a breathhold, hence improving image quality while reducing the required scan time, strongly increases the applicability of this technique.

Comparison of a 2D dynamic MR acquisition showing the heart during the cardiac cycle. The reference image uses data from a 10s breathhold and was reconstructed with a non-Cartesian iterative kt-SENSE approach. The other reconstruction only used one third of this data (i.e. 3s).

A direct reconstruction (i.e. gridding or NUFFT) shows significant undersampling artefacts. A compressed sensing approach (kt-FOCUSS) is able to remove some of these artefacts. A machine learning approach (XY-ML) applied in the spatial domain (x-y-space) can remove undersampling artefacts but suffers from strong blurring. A novel machine learning approach (XT-YT-ML) applied along the temporal dimension of this dynamic data set successfully minimizes undersampling artefacts while accurately preserving the anatomical and functional cardiac information.

Clinical applications

Advanced MR image reconstruction can improve image quality of MRI. In addition, it can help to reduce MR scan time without impairing diagnostic accuracy. This is especially important for cardiac MRI, where data acquisition has to take respiratory and cardiac motion of the heart into account, often leading to long scan times. We are collaborating closely with the Department of Mathematics, TU Berlin to utilize novel concepts of signal theory and with Charité Berlin to ensure developed approaches are evaluated in patients by clinical experts.

To top

Current projects

Machine learning for MR image reconstruction

Novel concepts of machine learning are utilised to improve the image quality of highly undersampled dynamic MR acquisition allowing for shorter scan times. This work is carried out as part of the DFG research training group BIOQIC (BIOphysical Quantitative Imaging Towards Clinical Diagnosis) and in collaboration with Charité Berlin.

Cooperation with:
Prof. Dr. Marc Dewey, Department of Radiology, Charité Berlin, Germany

To top

Selected references

A. Kofler, M. Haltmeier, C. Kolbitsch, M. Kachelriess, M. Dewey,
A U-Nets Cascade for Sparse View Computed Tomography
Machine Learning for Medical Image Reconstruction, 91-99 (2018).

J. Ma, M. Maerz, S. Funk, J. Schulz-Menger, G. Kutyniok, T. Schaeffter, C. Kolbitsch,
Shearlet-based compressed sensing for fast 3D cardiac MR imaging using iterative reweighting
Opens external link in new window Physics in Medicine & Biology 63, 235004 (2018).

J. Ma, F. Schnabel, Z. Chen, S. Funk, G. Kutyniok, J. Schulz-Menger, T. Schaeffter, C. Kolbitsch,
3D high-resolution LGE MRI using shearlet-based compressed-sensing image reconstruction
Proceedings of the 20th Annual Scientific Sessions of SCMR, Washington, United States, (2017).

To top



Physikalisch-Technische Bundesanstalt
Abbestraße 2–12
10587 Berlin