Benjamin Weppner1,2, Qihao Zhang2, Dominick Romano1,2, Renjiu Hu2,3, Pascal Spincemaille2, Shun Zhang4,5, and Yi Wang1,2
1Biomedical Engineering, Cornell University, Ithaca, NY, United States, 2Weill Cornell Medical College, New York, NY, United States, 3Mechanical Engineering, Cornell University, Ithaca, NY, United States, 4Radiology, Tongji Hospital, Tongji Medical College, Wuhan, China, 5Huazhong University of Science and Technology, Wuhan, China
Synopsis
Keywords: Stroke, Perfusion, Stroke, DSC MRI, Ischemia, Blood Flow Quantification, Deep Learning
Motivation: To assess the ability of quantitative transport mapping (QTM) to estimate blood flow in stroke from DSC MRI through a deep learning model.
Goal(s): To use an automated deep learning based method to measure blood flow in stroke using DSC MRI.
Approach: A deep learning network (QTMnet) is trained on synthetic MR data generated using realistic vascular models to learn the mapping between DSC MR data and underlying tissue blood flow.
Results: QTMnet demonstrates decreased perfusion in ischemic lesion compared to contralateral healthy tissue (p=0.0006), similar to results using traditional modeling. QTMnet performed well without needing to select an appropriate AIF or regularization.
Impact: QTMnet may identify hypoperfused tissue following stroke in an automated
manner. Accurate blood flow estimation may assist in determining whether
reperfusion therapy is beneficial.
Background
Dynamic Susceptibility Contrast (DSC)
MRI is used to quantify perfusion in stroke and is part of the
perfusion-diffusion mismatch approach to assess the risk of progressive
ischemic stroke by determining the non-core hypoperfused area whose size may
indicate whether reperfusion therapy will be beneficial1. Traditional
perfusion analysis relies on the selection of a global input function (AIF)
which ignores the spatial aspect of flow and is therefore susceptible to delay,
dispersion, and dilution effects as the bolus travels through the brain
vasculature. Global AIF selection, both manual and automatic, can lead to large
variations in blood flow estimation2. Quantitative transport mapping
(QTM), a biophysical modeling approach to tracer kinetics, and its deep
learning extension, QTMnet3, have been developed to eliminate the
use of a global AIF. QTM has been shown to outperform traditional tracer
kinetics models in simulations and phantoms where ground truth is available and
in classification tasks4,5,6,7,8. Here, we apply QTM to DSC MRI,
particularly for blood flow quantification. We show that QTMnet demonstrates decreased
perfusion in ischemic lesions without needing to select a global AIF or tuning regularization
as is required for the traditional kinetics model.Methods
Following stroke, 11 patients underwent DSC
MRI (1.9x1.9x6.5 mm3 resolution, 1.5 s time resolution, TE=19 ms,
TR=15 ms) and DWI MRI (0.94 mm in-plane resolution, 6.5 mm slice resolution, 3
s time resolution, TE=70.9 ms, TR=3 s, b-values of 0 and 1000) on a 3T GE
Discovery. Post-processing included skull stripping, resampling to 1mm3
isotropic resolution, and co-registration of DWI to DSC and was performed using
FSL9. Contrast agent concentration was assumed to be given by:
$$C_t = -log(\frac{S_t}{S_0})$$ where $$$S_t$$$ was the DCE MRI
image at time $$$t$$$. ADC maps were computed from the DWI data.
In QTMnet, a Unet is trained to map the
concentration images to the underlying
blood flow. The training data consisted of synthetic data generated as follows.
Anatomically inspired synthetic vasculature with an approximate capillary
network in 32x32x32mm3 cubes was generated using the constrained
constructive optimization (CCO) algorithm10 based on synthetically
generated 3D maps of piecewise constant blood flow, $$$F$$$, and cerebral blood volume, $$$V_p$$$, and a local tracer boundary condition given by a randomly
generated gamma variate with second pass characteristics equation11.
Computational flow modeling is used to generate the simulated concentration
profile of tracer through the vasculature, as well as perfusion into the tissue
structure. These training data aim to provide a realistic depiction of tracer
transport in the discrete domain. We implemented a standard Unet design12
with 45 temporal input channels and 16 layers. We generated 420 32x32x32 mm3
cubes with an 80/20 training/validation split and trained for 500 epochs using
the SGD optimizer (learning rate = 0.001, momentum = 0.9) with $$$L_1$$$ loss.
Traditional perfusion analysis was done
by solving a one-compartment tracer kinetic model: $$V_p\partial_t C_p (t)=F(C_a (t)-C_p (t))$$ Lesions
and contralateral healthy tissue were segmented based on the DWI and ADC maps.
A two-sample t-test was performed to determine differences in perfusion
parameters between these two regions.Results
Figure 1 shows the synthetic vasculature
generated using the CCO algorithm for a single 32x32x32 mm3 training
cube, as well as the corresponding flow map for a given slice. Figure 2 show
the flow maps for one patient, with QTMnet better defining the
hypoperfusion in the lesion compared to the contralateral healthy tissue compared
to the traditional method. This is reflected across subjects, as seen in Figure
3. Both QTMnet and the traditional perfusion analysis showed a decrease in flow
in the stroke lesion compared to healthy tissue, p<0.0006 and p<0.03,
respectively.Discussion
This study shows the feasibility of
using QTMnet for measuring the difference in blood flow in a stroke lesion
versus contralateral healthy tissue. While the traditional model performed
similarly to QTMnet, the traditional model required careful AIF selection for
each patient as well as regularization tuning, for the model to output a
reasonable result given the noisy data. For the patient with higher flow in the
lesion compared to contralateral healthy for QTMnet, it was discovered that a
vessel nearby the lesion may be contributing to signal within the lesion ROI. The
traditional kinetics model also estimated higher flow in the lesion for this
patient, as well as for an additional patient where QTMnet correctly estimated
lower flow in the lesion. Plans for future work include expanding QTMnet to the
two-compartment exchange model, which will allow for estimation of permeability
with the aim of predicting hemorrhagic transformation risk following stroke13.Acknowledgements
This work was funded by the NIH 1 R01 EB034755-01 and NIH 1 R01 AG080011-01A1.References
- Hsu, C. Y., Cheng, C.
Y., Tsai, Y. H., Lee, J. D., Yang, J. T., Weng, H. H., Lin, L. C., Huang, Y.
C., Lee, M., Lee, M. H., Wu, C. Y., Lin, Y. H., Hsu, H. L., Yang, H. T., Pan,
Y. T., & Huang, Y. C. (2015). Perfusion-diffusion Mismatch Predicts Early
Neurological Deterioration in Anterior Circulation Infarction without
Thrombolysis. Current neurovascular research, 12(3),
277–282. https://doi.org/10.2174/1567202612666150605122536
- Calamante, F. (2013). Arterial input
function in perfusion MRI: A comprehensive review. Progress in Nuclear Magnetic Resonance Spectroscopy, 74, 1-32. https://doi.org/10.1016/j.pnmrs.2013.04.002
- Zhang, Q. Estimating perfusion and
vascular properties from medical images: Quantitative transport mapping (QTM)
[Ph.D.]. United States – New York: Cornell University; 2023. 104p
- Zhou, L., Zhang, Q., Spincemaille, P., Nguyen, T. D., Morgan,
J., Dai, W., Li, Y., Gupta, A., Prince, M. R., & Wang, Y. (2021).
Quantitative transport mapping (QTM) of the kidney with an approximate
microvascular network. Magnetic
resonance in medicine, 85(4), 2247–2262. https://doi.org/10.1002/mrm.28584
- Huang, W., Zhang, Q., Wu, G., Chen, P.P., Li, J., Gillen, K.M., Spincemaille,
P., Chiang, G.C., Gupta, A., Wang, Y., & Chen, F. (2021). DCE-MRI
quantitative transport mapping for noninvasively detecting hypoxia inducible
factor-1α, epidermal growth factor receptor overexpression, and Ki-67 in nasopharyngeal
carcinoma patients. Radiotherapy and
Oncology, 164, 146-154. https://doi.org/10.1016/j.radonc.2021.09.016
- Zhang, Q., Spincemaille, P.,
Drotman, M., Chen, C., Eskreis-Winkler, S., Huang, W., Zhou, L., Morgan, J.,
Nguyen, T. D., Prince, M. R., & Wang, Y. (2022). Quantitative transport
mapping (QTM) for differentiating benign and malignant breast lesion:
Comparison with traditional kinetics modeling and semi-quantitative enhancement
curve characteristics. Magnetic
resonance imaging, 86, 86–93. https://doi.org/10.1016/j.mri.2021.10.039
- Zhang, Q., Luo, X., Zhou, L.,
Nguyen, T., Prince, M., Spincemaille, P., & Wang, Y. (2023) Fluid mechanics
approach to perfusion quantification: Vasculature computational fluid dynamics
simulation, quantitative transport mapping (QTM) analysis of dynamics contrast
enhanced MRI, and application in nonalcoholic fatty liver disease
classification. IEEE
Transactions on Biomedical Engineering,
70(3), 980-990. https://doi.org/10.1109/TBME.2022.3207057
-
Zhang, Q., Lee, K. S., Talenfeld, A.
D., Spincemaille, P., Prince, M. R., & Wang, Y. (2022). Prediction of Lung
Shunt Fraction for Yttrium-90 Treatment of Hepatic Tumors Using Dynamic
Contrast Enhanced MRI with Quantitative Perfusion Processing. Tomography (Ann Arbor, Mich.), 8(6), 2687–2697. https://doi.org/10.3390/tomography8060224
- Jenkinson,
M. and Smith, S. M. A Global Optimisation Method for Robust Affine Registration
of Brain Images. Medical Image Analysis, 5(2), 143-156, 2001.
- Karch, R., Neumann, F., Neumann, M.,
& Schreiner, W. (1999). A three-dimensional model for arterial tree
representation, generated by constrained constructive optimization. Computers
in biology and medicine, 29(1), 19–38.
https://doi.org/10.1016/s0010-4825(98)00045-6
- Patil, V., &
Johnson, G. (2011). An improved model for describing the contrast bolus in
perfusion MRI. Medical physics, 38(12), 6380–6383. https://doi.org/10.1118/1.3658570
- Ronneberger, O., Fischer,
P., & Brox, T. (2015). U-net: Convolutional networks for biomedical image
segmentation. https://doi.org/10.48550/arXiv.1505.04597
- Lin, K., Kazmi, K.S., Law, M., Babb, J., Peccerelli,
N., & Pramanik, B.K. (2007). Measuring elevated microvascular permeability
and predicting hemorrhagic transformation in acute ischemic stroke using
first-pass dynamic perfusion ct imaging. American Journal of Neuroradiology,
28(7), 1292-1298.