Marc Vornehm1,2, Maximilian Fenski3, Elisabeth Preuhs4, Andreas Maier4, Jeanette Schulz-Menger3, Jens Wetzl1, and Daniel Giese1,5
1Magnetic Resonance, Siemens Healthcare GmbH, Erlangen, Germany, 2Artificial Intelligence in Biomedical Engineering, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany, 3Working Group Cardiovascular Magnetic Resonance, Experimental and Clinical Research Center, Charité Medical Faculty, Max Delbrück Center for Molecular Medicine, HELIOS Klinikum Berlin Buch, Department of Cardiology and Nephrology, Berlin, Germany, 4Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany, 5Institute of Radiology, University Hospital Erlangen, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany
Synopsis
Automatic segmentation of left
ventricular myocardium on LGE images of patients with chronic myocardial
infarction is challenging due to inhomogeneous signal intensities in the
presence of myocardial scar. However, the inclusion of scar in automatically generated
myocardium segmentations is critical for further LGE analysis. We propose a
Deep Learning-based method for myocardial segmentation on cardiac LGE images,
achieving average Dice scores of 0.78 and 0.80 on test images with and without scar,
respectively. We furthermore evaluate the network’s sensitivity to scar and
show that it can be improved by incorporating synthetic images with diverse
enhancement patterns in the training data.
Introduction
Contouring
of the left ventricular (LV) myocardium in late gadolinium enhancement (LGE)
images is an essential step in quantification of myocardial scar and is usually
performed manually by the reader1. Automatic segmentation of the LV
could facilitate this process and could furthermore serve as starting point for
a subsequent fully automatic quantification of myocardial scar2.
However, automatic segmentation on LGE images is challenging due to highly
variable enhancement patterns, inhomogeneous image intensities within the
scarred region, and ambiguous image boundaries, particularly in patients with
subendocardial scar. Inclusion of scarred tissue within the myocardial boundaries,
however, is critical for an accurate scar quantification in the follow-up. Multiple
Deep Learning-based methods for myocardial segmentation have been proposed
previously, achieving Dice similarity coefficients (DSC) of around 0.78-0.823-7.
However, they were not quantitatively evaluated for their sensitivity to
myocardial scar, i.e., the percentage of scar included in the myocardial
segmentation. In this study, we describe a U-Net-based approach for myocardial
segmentation and evaluate its performance in scarred regions. We investigate
the effect of incorporating simulated images, hypothesizing that an increased
variability in enhancement patterns in the training data improves the network’s
sensitivity to scar.Methods
Fifty-one
patients with chronic myocardial infarction underwent motion corrected PSIR-LGE
imaging on a 1.5T scanner (MAGNETOM AvantoFit, Siemens Healthcare, Erlangen, Germany)
~15-20 minutes after injection of 0.15mmol/kg Gadoteridol (slice thickness 6mm,
resolution 1.56mm²). Each acquisition was reconstructed as magnitude inversion
recovery (MagIR) and phase-sensitive inversion recovery (PSIR)8.
Myocardial borders were contoured by a reader with three years of experience
(consensus read with an SCMR Level III certified reader in case of inconclusive
findings). Corresponding cine images were used as reference to support the
contouring process. Enhanced regions were observed in 44 patients (LGE-positive)
and the scar area was determined using the full width at half maximum (FWHM) method9,10.
The data
set was supplemented with synthetic images created using a Bloch simulation
framework and the XCAT phantom11. Scars were simulated by altering the
tissue parameters within the myocardium. The simulations included variations in
anatomy, scar location, size, and transmurality, and parameters influencing the
image contrast. 300 acquisitions were simulated and reconstructed as MagIR and
PSIR images. Exemplary real and simulated images are given in Figure 1.
Images were
cropped to 256x256 pixels with the heart centered. Of the 44 LGE-positive
patients, 15 were used for training and five for validation. The remaining 24
LGE-positive patients and seven LGE-negative patients (i.e., without myocardial
scar) were set aside and formed two separate data sets for testing. MagIR and
PSIR images were used concurrently, resulting in 30 individual images for
training. The number of synthetic images in the training set was varied between
zero and 600. Images were augmented using affine transformations, blurring, and
local image deformation at a random location within the myocardium. The real
images were presented to the network twenty times per training epoch in
randomly augmented ways, such that real images were presented to the network at
least as often as synthetic images. The network was based on the U-Net architecture12
with residual connections, BatchNormalization and ReLU activation functions (Figure 2). It was optimized using the Dice
loss function13 and Adam optimizer14 with a learning rate
of 10-3. The obtained segmentation masks were post-processed by
removing all but the largest connected component.Results
The trained
networks were evaluated in terms of average DSC and sensitivity to myocardial
scar. Average DSC was higher on the LGE-negative (0.79-0.80) than on the LGE-positive
test set (0.76-0.78) in all experiments. Adding synthetic data did not significantly
affect the DSC. The average sensitivity to myocardial scar was evaluated on the
LGE-positive test set and was observed to increase when more synthetic images
were added to the training set. Detailed quantitative results and exemplary
segmentation results of images with myocardial scar are given in Figures 3 and 4, respectively.Discussion
The
reported DSC values are comparable to those of previously published methods and
were achieved using a standard network architecture and very few training samples.
A more accurate segmentation of the myocardium on LGE images is mainly hindered
by the presence of enhanced regions, i.e., scarred tissue. This is manifested
in the decreased DSC observed in LGE-positive images compared to LGE-negative images.
The reported sensitivity values for myocardial scar indicate that considerable
portions of the scars were not included in the myocardial borders. Subendocardial
scars, exhibiting similar signal intensities to the adjacent blood pool, are particularly
prone to exclusion from automatic segmentations (Figure 4d). Increasing the number of
synthetic images in the training set led to improved sensitivity values,
possibly due to an increased variability of enhancement patterns in the
training data.Conclusion
We
presented an automatic approach for myocardial segmentation on LGE images,
achieving similar results as previous methods in terms of DSC. The clinical
value of such algorithms, however, is considerably diminished if the sensitivity
to myocardial scar is low. We showed that this issue can be addressed by increasing
the variability of enhancement patterns in the training data, specifically by
incorporating simulated images. Other approaches like, for instance, style
transfer methods for image synthesis6 or innovative augmentation
techniques5 could similarly be of interest in this regard.Acknowledgements
No acknowledgement found.References
- Schulz-Menger J, Bluemke D, Bremerich J, et al. Standardized image interpretation
and post-processing in cardiovascular magnetic resonance - 2020 update: Society
for Cardiovascular Magnetic Resonance (SCMR): Board of Trustees Task Force on
Standardized Post-Processing. J Cardiovasc Magn Reson. 2020;22:19.
- Moccia
S, Banali R, Martini C, et al. Development and testing of a deep learning-based
strategy for scar segmentation on CMR-LGE images. Magn Reson Mater Phy. 2019;32(2):187-195.
- Fahmy
AS, Rausch J, Neisius U, et al. Automated cardiac MR scar quantification in hypertrophic
cardiomyopathy using deep convolutional neural networks. JACC-Cardiovasc Imag. 2018;11(12):1917-1918.
- Yue Q,
Luo X, Ye Q, et al. Cardiac Segmentation from LGE MRI Using Deep Neural Network
Incorporating Shape and Spatial Priors. In: Shen D, et al. (eds) Medical Image
Computing and Computer Assisted Intervention – MICCAI 2019. MICCAI 2019. 2019;559-567.
- Campello
VM, Martín-Isla C, Izquierdo C, et al. Combining Multi-Sequence and Synthetic
Images for Improved Segmentation of Late Gadolinium Enhancement Cardiac MRI.
In: Pop M, et al. (eds) Statistical Atlases and Computational Models of the
Heart. Multi-Sequence CMR Segmentation, CRT-EPiggy and LV Full Quantification
Challenges. STACOM 2019. 2020;290-299.
- Chen C,
Ouyang C, Tarroni C, et al. Unsupervised Multi-modal Style Transfer for Cardiac
MR Segmentation. In: Pop M, et al. (eds) Statistical Atlases and Computational
Models of the Heart. Multi-Sequence CMR Segmentation, CRT-EPiggy and LV Full
Quantification Challenges. STACOM 2019. 2020;209-219.
- Roth H,
Zhu W, Yang D, et al. Cardiac Segmentation of LGE MRI with Noisy Labels. In:
Pop M, et al. (eds) Statistical Atlases and Computational Models of the Heart.
Multi-Sequence CMR Segmentation, CRT-EPiggy and LV Full Quantification
Challenges. STACOM 2019. 2020;228-236.
- Kellman
P, Arai A, McVeigh E, et al. Phase-sensitive inversion recovery for detecting
myocardial infarction using gadolinium-delayed hyperenhancement. Magn Reson Med.
2002;47(2):372-383.
- Amado L,
Gerber B, Gupta S, et al. Accurate and objective infarct sizing by
contrast-enhanced magnetic resonance imaging in a canine myocardial infarction
model. J Am Coll Cardiol. 2004;44(12):2383-2389.
- Flett A,
Hasleton J, Cook C, et al. Evaluation of Techniques for the Quantification of
Myocardial Scar of Differing Etiology Using Cardiac Magnetic Resonance. JACC-Cardiovasc
Imag. 2011;4(2):150-156.
- Segars
WP, Sturgeon G, Mendonca S, et al. 4D XCAT phantom for multimodality imaging
research. Med Phys. 2010;37(9):4902–4915.
- Ronneberger O, Fischer P, Brox T, et al. U-Net: Convolutional Networks for
Biomedical Image Segmentation. In: Navab N, et al. (eds) Medical Image
Computing and Computer-Assisted Intervention – MICCAI 2015. MICCAI 2015.
2015;234-241.
- Milletari F, Nassir N, Seyed-Ahmad A. V-Net: Fully Convolutional Neural
Networks for Volumetric Medical Image Segmentation. 2016 Fourth International
Conference on 3D Vision (3DV). 2016;565-571.
- Kingma DP,
Ba J. Adam: A Method for Stochastic Optimization. arXiv:1412.6980 [cs.LG]. 2017.