A novel algorithm for correcting motion artifacts in cardiac cine MRI data is more accurate than standard gradient techniques, researchers in Finland reported in Academic Radiology.
The spatial and temporal resolution of MRI scanners continue to improve, leading to a need for more automated tools to analyze the images quickly and accurately, wrote Dr. Jyrki Lötjönen, Mika Pollari, Sari Kivisto, and Kirsi Lauerma from VTT Information Technology, Helsinki University of Technology, and Helsinki University in Finland.
"MR slice images from different spatial locations are acquired for the assessment of 3D volume," they explained. "The 3D volume can be formed from separate slices because their locations and orientations relative to the imaging device are defined in the image headers" (Academic Radiology, October 2005, Vol. 12:10, pp. 1273-1284).
Several MR image series are typically acquired in an imaging session. But when the subject moves due to breathing, coughing, or discomfort, the relationship between the image series (derived from the headers) is lost, making 3D accurate reconstruction impossible. ECG gating does a good job of eliminating heart motion artifacts. However, breathing is less regular than heartbeat, making it the most common and perplexing source of image blurring.
There is no need for correction when each MR slice is analyzed separately, but correction is needed when performing direct 3D analysis with model-driven techniques, the authors noted. And although standard reconstruction techniques work fine when two volumes contain corresponding structures, "the objective of this study is to develop a method to align images within a volume where exactly corresponding structures do not exist between neighboring slices," they stated.
A handful of research papers have addressed this problem. For their research, Lötjönen and colleagues expanded on the work of Moore et al (Ellis R, Peters T eds, "Lecture notes in computer science 2878: MICCAI 2003," Berlin, Germany). Moore's team built a high-resolution dynamic heart model in cardiac MRI images of a volunteer, correcting breath-hold alignment by registering a 3D volume with axial and scout images.
Lötjönen's team began by optimizing the location of each slice in an image stack by maximizing a similarity measure of the slice with another image slice stack, an iterative process that corrected both image stacks at the same time. The group then tested two procedures to optimize the similarity, including standard gradient optimization and stochastic optimization, "in which one slice is chosen randomly from the image stacks, and its location is optimized," Lötjönen and his colleagues wrote.
The first study examined nine healthy volunteers who underwent MRI using a 1.5-tesla Magnetom Vision scanner (Siemens Medical Solutions, Erlangen, Germany). Breath-hold gradient-echo turbo flash cine series were acquired in both the left ventricle (LV) short-axis (SA) planes using 10-mm slices and a 5-mm gap. Cine scans in the atria were acquired with contiguous 10-mm slice thicknesses.
A second study assessed 16 volunteers. Breath-hold true-FISP cine data were acquired in the LV short-axis planes with 7-mm slice thicknesses and an 8-mm gap from the LV base to the apex. Cine images in the four-chamber planes were acquired using slice thicknesses of 6-7 mm and a gap of 3-4 mm for the atria. Imaging parameters included 1.5 msec echo time, flip angle 60º, and time resolution of 34 msec. In all, four to six long-axis (LA) images and four to eight SA images were acquired, depending on the size of the heart.
In the first step of the process, voxel-by-voxel correspondence, "the coordinates of a voxel in the source volume, denoted by (X, Y, Z), are defined in the coordinate system of the destination volume, denoted by (X1, Y1, Z1)," the authors wrote. As the corrected SA slices are transformed to the LA coordinates system by means of six equations, the validity of the transformation was evaluated by means of a chessboard visualization technique.
"Data from every second box are from the transformed SA data, and the data in every other box are from the original LA data," they explained. "The borders between the boxes appear continuous when the data are well aligned…. In other words, data from one imaging direction are not enough to make the motion correction properly, but data from the other imaging directions should be used to guarantee the correct anatomy."
This movement correction is performed automatically via the proposed algorithm, which uses the highly versatile normalized mutual information (NMI) method to evaluate the similarity of data in image registration.
"The NMI measures how well the Gray value of a pixel (x, y, z) in image A can be predicted by knowing Gray value of the same pixel (x, y, z) in image B," they explained. When images A and B have been registered, each pixel corresponds to the same anatomic object in both images. In this case, a relation exists between the gray values. For example, a pixel sampled from bone appears bright in a (CT) image and dark in an MR image, and the prediction of gray values is possible." One reason for NMI's popularity is that it works equally well across imaging modalities, they stated.
The algorithm's next task is to maximize NMI between the SA and LA data, and apply the resulting values to probability distribution of Gray values. "In this work, the probability distributions are estimated from joint grayscale histograms as proposed by Maes et al from Katholieke Universiteit Leuven in Belgium" (IEEE Transactions on Medical Imaging, April 1997, Vol. 16:2, pp. 187-198), the authors wrote.
All time points related to one spatial location are acquired in a single breath-hold, so there is no respiratory motion between different cardiac values phases belonging to the same slice. Also, time resolution is equal in both SA and LA cine series, so data from all time points can be used to compute the similarity measure, they wrote.
NMI is computed using only voxels from the heart region, and the optimization algorithm moves each slice in the direction that maximizes NMI. Each image series (slices from various time points but only one spatial location) is assumed to be independent from displacements of the other image series.
The accuracy of the algorithm was evaluated visually, by inspecting results of real MRI data, and quantitatively, by simulating the movements in four real MR datasets, the authors wrote. The mean error and standard deviation were defined for 50 simulated movements as slices were randomly displaced, and the statistical difference between the optimization methods was evaluated using a paired t test.
According to the results, the algorithm was successfully applied to both real and simulated image data. When typical motion artifacts were simulated, the error rate was approximately 3% and 13% with displacements simulating typical and extreme conditions, respectively. Subvoxel registration was also accurate.
In a comparison of different optimization methods, the registration accuracy of the stochastic approach was superior to that of the standard gradient technique (p < 10-9).
"A good indication of successful motion correction with real data was that as we segmented the images after the motion correction, the segmentation results (a triangulated 3D surface) fitted well to both SA and LA volumes," the group wrote.
"Motion correction is a prerequisite for some 3D measures, such as shape descriptors or wall thickness, even if 2D techniques are used to segment slices," they wrote. "In addition, motion correction is needed if information on more than one imaging direction is used in segmentation."
The algorithm corrected motion artifacts using only SA and four-chamber LA volumes; however, data from other imaging volumes such as parasternal, horizontal, and vertical LA cine series could be substituted easily, they wrote. And because the NMI was to measure similarity, volumes acquired with different imaging sequences, or even different modalities such as ultrasound or CT, could also be used.
By Eric Barnes
AuntMinnie.com staff writer
November 1, 2005
Related Reading
Novel real-time 3D echo quantifies LV dyssynchrony, August 18, 2005
4D MRI enables blood-velocity measurement in mitral valve insufficiency, June 24, 2005
Coronary MS CTA quantifies LV function adequately, March 23, 2004
Copyright © 2005 AuntMinnie.com
