Logo of the Physikalisch-Technische Bundesanstalt
symbolic picture: "magazines"

Artificial intelligence for heart imaging

Improved medical image reconstruction by combining a physical model with a neural network

PTBnews 1.2022
Especially interesting for

manufacturers of MRI equipment

image reconstruction

medical engineering

In patients with a heart condition, magnetic resonance imaging allows parameters such as the pumping capacity of the heart to be determined. This method, however, has a disadvantage: data acquisition is time-consuming. At PTB, a physical model was combined with a neural network to develop a procedure that requires only a small amount of measurement data (and thus only a short measuring time) to obtain high-quality images of the phases of the cardiac cycle.

MR image of the heart in one phase of the cardiac cycle. Left: image reconstructed directly from the measured data; center: result of the recently developed method; right: reference image. The reference image required a measurement time that was more than six times longer than that of the new method.

A magnetic resonance imaging (MRI) scanner excites water molecules in the body cells via high-frequency radiation and then measures the fading magnetization. Due to spatial encoding, the measurement data is acquired in the Fourier space, from which the diagnostic images are then computed. This image reconstruction process requires a physical model which describes the imaging process. Iterative reconstruction procedures are used to obtain as much diagnostic information as possible within as short a measuring time as possible. Based on previous knowledge (e.g., spatial smoothness), it is already possible to compute high-quality images from only a few measuring points. Neural networks can learn this previous knowledge from the data and adapt it for the reconstruction problem.

At PTB, an iterative network has been developed specifically to reconstruct dynamic heart images. During the training phase, this procedure combines the physical model for imaging and the previous knowledge on the structure of the image data learned by the network. MR images of cardiac function then show the heart during different phases of the cardiac cycle. It was possible to use this temporal component of the data to ensure the best possible efficiency in training.

This method was applied to patients with a heart condition and then assessed. It was also compared to conventional iterative reconstruction procedures as well as to other machine learning methods. The network thus developed yields better results than conventional methods (by up to 6 dB in terms of signal-to-noise ratio and by up to –47 % in terms of relative errors). In addition, this network yielded results similar to those obtained by another method which was also based on neural networks. PTB’s procedure was additionally able to do this with only 10 % of the trainable parameters, which proves this new method’s robustness and efficiency.


Andreas Kofler
Department 8.1
Biomedical Magnetic Resonance
Phone: +49 30 3848-7749
Opens local program for sending emailandreas.kofler(at)ptb.de

Scientific publication:

A. Kofler, M. Haltmeier, T. Schäffter, C. Kolbitsch: An end‐to‐end‐trainable iterative network architecture for accelerated radial multi‐coil 2D cine MR image reconstruction. Medical Physics 48, 2412–2425 (2021)