Skip to main page content
U.S. flag

An official website of the United States government

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2022 Dec;56(6):1691-1704.
doi: 10.1002/jmri.28199. Epub 2022 Apr 22.

Accelerating Cardiac Diffusion Tensor Imaging With a U-Net Based Model: Toward Single Breath-Hold

Pedro F Ferreira  1   2 , Arjun Banerjee  3 , Andrew D Scott  1   2 , Zohya Khalique  1   2 , Guang Yang  1   2 , Ramyah Rajakulasingam  1   2 , Maria Dwornik  1   2 , Ranil De Silva  1   2 , Dudley J Pennell  1   2 , David N Firmin  1   2 , Sonia Nielles-Vallespin  1   2
Affiliations

Accelerating Cardiac Diffusion Tensor Imaging With a U-Net Based Model: Toward Single Breath-Hold

Pedro F Ferreira et al. J Magn Reson Imaging. 2022 Dec.

Abstract

Background: In vivo cardiac diffusion tensor imaging (cDTI) characterizes myocardial microstructure. Despite its potential clinical impact, considerable technical challenges exist due to the inherent low signal-to-noise ratio.

Purpose: To reduce scan time toward one breath-hold by reconstructing diffusion tensors for in vivo cDTI with a fitting-free deep learning approach.

Study type: Retrospective.

Population: A total of 197 healthy controls, 547 cardiac patients.

Field strength/sequence: A 3 T, diffusion-weighted stimulated echo acquisition mode single-shot echo-planar imaging sequence.

Assessment: A U-Net was trained to reconstruct the diffusion tensor elements of the reference results from reduced datasets that could be acquired in 5, 3 or 1 breath-hold(s) (BH) per slice. Fractional anisotropy (FA), mean diffusivity (MD), helix angle (HA), and sheetlet angle (E2A) were calculated and compared to the same measures when using a conventional linear-least-square (LLS) tensor fit with the same reduced datasets. A conventional LLS tensor fit with all available data (12 ± 2.0 [mean ± sd] breath-holds) was used as the reference baseline.

Statistical tests: Wilcoxon signed rank/rank sum and Kruskal-Wallis tests. Statistical significance threshold was set at P = 0.05. Intersubject measures are quoted as median [interquartile range].

Results: For global mean or median results, both the LLS and U-Net methods with reduced datasets present a bias for some of the results. For both LLS and U-Net, there is a small but significant difference from the reference results except for LLS: MD 5BH (P = 0.38) and MD 3BH (P = 0.09). When considering direct pixel-wise errors the U-Net model outperformed significantly the LLS tensor fit for reduced datasets that can be acquired in three or just one breath-hold for all parameters.

Data conclusion: Diffusion tensor prediction with a trained U-Net is a promising approach to minimize the number of breath-holds needed in clinical cDTI studies.

Evidence level: 4 TECHNICAL EFFICACY: Stage 1.

Keywords: CNN; U-Net; cardiac; deep learning; diffusion tensor imaging.

PubMed Disclaimer

Conflict of interest statement

Professor Dudley Pennell receives research support from Siemens and is a stockholder and director of Cardiovascular Imaging Solutions. The RBH CMR group receives research support from Siemens Healthineers

Figures

FIGURE 1
FIGURE 1
(a) Conventional linear least square tensor calculation. (b) U‐Net tensor prediction training workflow. The U‐Net is trained to predict the reference tensor results from reduced datasets
FIGURE 2
FIGURE 2
Loss function progress for the validation dataset for all three different reduced datasets
FIGURE 3
FIGURE 3
cDTI parameter maps for an example from the test dataset (healthy subject) where the tensor was calculated with an LLS algorithm for four different datasets, from left‐to‐right the reference gold‐standard (13 breath‐holds), the 5BH, 3BH, and the 1BH. The difference to the respective reference map is also shown (error map). Units: FA unitless; MD 10−3 mm2 sec−1; HA and E2A degrees.
FIGURE 4
FIGURE 4
cDTI parameter maps for the same subject from Fig. 3. The reference results (left) are compared to the tensor parameter maps from the U‐Net for the three different datasets: 5BH, 3BH, and 1BH. The difference to the respective reference map is also shown (error map). Units: FA unitless; MD 10−3 mm2 sec−1; HA and E2A degrees.
FIGURE 5
FIGURE 5
Bland–Altman plots showing healthy and patient cohorts for the LLS (left) and U‐Net (right) results for global mean FA, MD, HA gradient, and global median |E2A|. Each dot represents one subject (dots: healthy controls; hollow circles: patients). For each dataset, it is also shown the median and 5 and 95% quantiles (solid and dashed lines respectively). The 95% confidence interval is also shown with a faint rectangular area around the median. Units: FA unitless; MD 10−3 mm2 sec−1; HA gradient degrees/mm; and E2A degrees.
FIGURE 6
FIGURE 6
Mean absolute error for FA and MD and mean absolute angular error for HA and E2A. Each dot represents one subject (dot: healthy controls; hollow circle: patients). For each dataset, it is also shown the median and interquartile range (gray rectangles). Units: FA unitless; MD 10−3mm2 sec−1; HA and E2A degrees.
FIGURE 7
FIGURE 7
Top: Scatter plot with mean MD values for the infarcted and remote myocardial regions in 16 patients with an acute myocardial infarction. MD results are shown for the reference, LLS and U‐Net algorithms. The intersubject median and interquartile range (black rectangles) are also shown. Bottom: Example showing the MD maps and error maps for one patient. The red arrows indicate the infarcted region and the green arrows the remote region as given by gadolinium presence in a late‐gd enhancement scan. Units: MD 10−3 mm2 sec−1.
FIGURE 8
FIGURE 8
Mean absolute error for FA and MD and mean absolute angular error for HA and E2A for the prospective scans in five healthy volunteers at five equidistant slices in the left ventricle from base to apex. Each dot represents one subject at a color‐coded slice position. For each dataset, it is also shown the median and interquartile range (gray rectangles). Units: FA unitless; MD 10−3 mm2 sec−1; HA and E2A degrees.
See this image and copyright information in PMC

Comment in

Similar articles

See all similar articles

Cited by

References

    1. Kung GL, Nguyen TC, Itoh A, et al. The presence of two local myocardial sheet populations confirmed by diffusion tensor MRI and histological validation. J Magn Reson Imaging 2011;34(5):1080‐1091. - PMC - PubMed
    1. Helm P, Beg MF, Miller MI, Winslow RL. Measuring and mapping cardiac fiber and laminar architecture using diffusion tensor MR imaging. Ann N Y Acad Sci 2005;1047:296‐307. - PubMed
    1. Scollan DF, Holmes A, Winslow R, Forder J. Histological validation of myocardial microstructure obtained from diffusion tensor magnetic resonance imaging. Am J Physiol 1998;275(6 Pt 2):H2308‐H2318. - PubMed
    1. Ferreira PF, Kilner PJ, McGill L‐A, et al. In vivo cardiovascular magnetic resonance diffusion tensor imaging shows evidence of abnormal myocardial laminar orientations and mobility in hypertrophic cardiomyopathy. J Cardiovasc Magn Reson 2014;16:87. - PMC - PubMed
    1. von Deuster C, Sammut E, Asner L, et al. Studying dynamic myofiber aggregate reorientation in dilated cardiomyopathy using in vivo magnetic resonance diffusion tensor imaging. Circ Cardiovasc Imaging 2016;9:e005018. - PMC - PubMed

Publication types