Improved Vascular Transport Function Characterization in DSC-MRI via Deconvolution with Dispersion-Compliant Bases
Marco Pizzolato1, Rutger Fick1, Timothé Boutelier2, and Rachid Deriche1

1Athena Project-Team, Inria Sophia Antipolis - Méditerranée, Sophia Antipolis, France, 2Olea Medical, La Ciotat, France

Synopsis

Bolus dispersion phenomena affect the residue function computed via deconvolution of DSC-MRI data. Indeed the obtained effective residue function can be expressed as the convolution of the true one with a Vascular Transport Function (VTF) that characterizes the dispersion. The state-of-the-art technique CPI+VTF allows to estimate the actual residue function by assuming a model for the VTF. We propose to perform deconvolution representing the effective residue function with Dispersion-Compliant Bases (DCB) without assumptions on the VTF, and then apply the CPI+VTF on DCB results. We show that DCB improve robustness to noise and allow to better characterize the VTF.

Purpose

To improve robustness of dispersion kernel characterization in DSC-MRI by means of Dispersion-Compliant Bases.

Introduction

The residual amount of tracer, i.e. the residue function $$$R(t)$$$ computed from deconvolution of the measured arterial $$$C_a(t)$$$ and tissular $$$C_{ts}(t)$$$ concentrations, characterizes the tissue perfusion. However the actual arterial concentration may undergo dispersion. This causes the effective residue function to reflect additional vascular properties mathematically described by the convolution $$$R^{eff}(t)=R(t)\otimes{VTF(t)}$$$ where VTF is the Vascular Transport Function1,2. This severely affects the estimation of hemodynamic parameters3 such as the blood flow $$$BF$$$, corresponding to the peak of $$$R(t)$$$, and the mean transit time $$$MTT=BV/BF$$$ with $$$BV$$$ the blood volume. Indeed only effective parameters $$$BF^{eff}$$$ (peak of $$$R^{eff}(t)$$$) and $$$MTT^{eff}$$$ are computed.

A recent state-of-the-art technique4 based on control point interpolation, CPI+VTF, allows to recover the actual $$$R(t)$$$ assuming that it is convolved with a VTF described by a Gamma Dispersion Kernel (GDK) $$VTF(t,s,p)=GDK(t,s,p)=\frac{s^{1+sp}}{\Gamma(1+sp)}t^{sp}e^{-st}$$ where $$$s,p$$$ are unknown. This allows the estimation of the actual $$$BF$$$4. The estimation of $$$s,p,BF$$$ is not an easy task and requires a non-linear optimization routine which results are sensitive to noise.

We propose to improve robustness and precision of some of the estimates by performing deconvolution with Dispersion-Compliant Bases5 (DCB), and subsesquently fit the CPI+VTF4 model to the obtaned effective residue function.

Methods

We perform DCB deconvolution representing $$$R^{eff}(t)$$$ on a sampling grid $$$t_1,t_2,..,t_M$$$ as5 $$R_{DCB}(t) = \Theta(t-\tau) \sum_{n=1}^{N} [a_n + b_n (t-\tau)] e^{-\alpha_n (t-\tau)}$$ with order $$$N$$$ ($$$6$$$ here), $$$\tau,a_n,b_n$$$ unknown an $$$\alpha_n$$$ predefined5. The solution was constrained via quadratic programming to $$$R(t_m)\ge0\forall{t_m\in[t_1,t_{M-1}]}$$$ and $$$R(t_M)=0$$$.

The CPI+VTF deconvolution technique was implemented as in literature4 with 12 control points and initial parameters $$$p,s$$$ for the optimization routine $$$log2\pm2$$$ ($$$mean\pm{SD}$$$). In order to decouple the influence of the estimation framework from the model the estimation was performed non-linearly bounding parameters to $$$mean\pm3SD$$$. The CPI+VTF model was also fitted on the effective residue function computed with DCB by minimizing $$$||R_{DCB}(t)-{[CPI+VTF]}_{model}||^2$$$ over the control points, time-instants separations, and parameters $$$BF,s,p$$$4.

We perform synthetic experiments generating $$$C_a(t)$$$ in $$$[0:1:90]s$$$ as a gamma-variate function1.The tissular concentration $$$C_{ts}(t)$$$ was generated as $$$C_{ts}(t)=C_a\otimes[R\otimes{VTF}(t)](t)$$$ with bi-exponential5 $$$R(t)$$$. Three ground-truth VTF models were used4: gamma (GDK), exponetial and log-normal. For each, three dispersion levels were tested: low, medium, high4. Number 100 repetitions were generated for each combination of dispersion kernel, level, $$$BF\in[5:10:65]ml/100g/min$$$, $$$MTT\in[2:4:18]s$$$, and $$$delay\in[0,5]s$$$4 with noise added5 with $$$SNR=50$$$4. For each repetition DCB and CPI+VTF deconvolutions were performed, as well as the CPI+VTF model fitting on $$$R_{DCB}(t)$$$, henceforth DCB+VTF.

We then proceed with the following experiments:

1. we compare DCB5, CPI+VTF4 and oSVD6 deconvolutions and calculate the relative errors of the recovered effective parameters $$$BF^{eff}$$$ (Fig. 1), $$$MTT^{eff}=BV/BF^{eff}$$$ (Fig. 2), and time-to-maximum $$$T2MAX$$$ of the $$$R^{eff}$$$ (Fig. 3); comparisons are performed on all of the ground-truth dispersion kernels (left images) and just on the GDK (right images);

2. we compare estimates of $$$p,s,BF$$$ obtained with CPI+VTF and DCB+VTF in case of GDK (Fig. 4);

3. we apply DCB+VTF on stroke MRI data and show maps of $$$p,s,BF$$$ (Fig.5).

Results

Results in Fig. 1,2,3-left show that DCB-based estimates of $$$BF^{eff},MTT^{eff},T2MAX$$$ have sensibly lower relative error than those obtained with CPI+VTF and oSVD. When the gound-truth kernel is GDK (right columns) CPI+VTF sensibly improve. Still, DCB perform comparably or better. DCB generally reduce errors and their variability. In addition DCB results appears more stable than with CPI+VTF with respect to the ground-truth kernel. Results in Fig. 4 show that CPI+VTF and DCB+VTF render similar results for $$$BF,p$$$ but DCB+VTF relevantly improves $$$s$$$ estimates at medium and high dispersion levels. Maps in Fig. 5 well depict the infarcted area which is specially highlighted in $$$s$$$-map.

Discussion

The use of DCB deconvolution renders a better estimation of the effective hemodynamic parameters. The DCB do not assume any model for the dispersion kernel (VTF) and can handle the exponential and log-normal kernels better than CPI+VTF. The use of these functional bases improves robustness to noise and reduces variability in the results (Figs. 1-3). This leads to a relevant improvement also when the CPI+VTF technique is applied directly on the DCB results (DCB+VTF), specifically in the estimation of the $$$s$$$ parameter of the gamma kernel (Fig. 4). The same parameter reveals to be very effective in delimiting the infarcted area in in-vivo results (Fig. 5).

Conclusion

Perfusion deconvolution of DSC-MRI data by means of Dispersion-Compliant Bases (DCB) provides more robust results in quantifying the effective residue function and improves subsequent VTF assessments. This allows a better characterization of dispersion phenomena and a consequent deeper understanding of the vascular dynamic.

Acknowledgements

We thank Olea Medical and the PACA Regional Council for providing grant and support.

References

1. Calamante et al., “Delay and dispersion effects in dynamic susceptibility contrast mri: simulations using singular value decomposition,” Magn Reson Med, vol. 44(3), pp. 466–473, 2000

2. Calamante et al., “Estimation of bolus dispersion effects in perfusion mri using image-based computational fluid dynamics,” NeuroImage, vol. 19, pp. 341–353, 2003.

3. Willats et al., “Improved deconvolution of perfusion mri data in the presence of bolus delay and dispersion,” Magn Reson Med, vol. 56, pp. 146156, 2006.

4. Mehndiratta et al., “Modeling and correction of bolus dispersion effects in dynamic susceptibility contrast mri: Dispersion correction with cpi in dsc-mri,” Magn Reson Med, vol. 72, pp. 17621774, 2013.

5. Pizzolato et al., “Perfusion mri deconvolution with delay estimation and non-negativity constraints,” in 12th International Symposium on Biomedical Imaging (ISBI). IEEE, 2015, pp. 1073–1076.

6. Wu et al., “Tracer arrival timinginsensitive technique for estimating flow in mr perfusionweighted imaging using singular value decomposition with a blockcirculant deconvolution matrix,” Magn Reson Med, vol. 50(1), pp. 164–174, 2003.

Figures

Relative absolute errors boxplots for the effective blood flow $$$BF^{eff}$$$ (peak of the effective residue function) computed with oSVD6, CPI+VTF4 and DCB5. Left. combined results when gamma (GDK), exponential and log-normal dispersion kernels are used as ground-truth for the Vascular Transfer Function (VTF). Right. only the gamma kernel is used.

Relative absolute errors boxplots for the effective mean transit time $$$MTT^{eff}=BV/BF^{eff}$$$ (with $$$BV$$$ the blood volume) computed with oSVD6, CPI+VTF4 and DCB5. Left. combined results when gamma (GDK), exponential and log-normal dispersion kernels are used as ground-truth for the Vascular Transfer Function (VTF). Right. only the gamma kernel is used.

Relative absolute errors boxplots for the time-to-maximum $$$T2MAX$$$ of the effective residue function computed with oSVD6, CPI+VTF4 and DCB5. Left. combined results when gamma (GDK), exponential and log-normal dispersion kernels are used as ground-truth for the Vascular Transfer Function (VTF). Right. only the gamma kernel is used.

Estimation of the $$$BF,p,s,$$$ parameters with CPI+VTF deconvolution4 (purple) and with the fitting of the CPI+VTF model on the the effective residue function obtained with DCB deconvolution5 (DCB+VTF). The ground-truth data was generated with the gamma dispersion kernel (GDK). Results shown for low, medium, high and all dispersion levels4.

Maps of the actual blood flow $$$BF$$$ and of the parameters $$$p,s$$$ characterizing a gamma Vascular Transport Function obtained after fitting of the CPI+VTF model on the the effective residue function estimated with DCB deconvolution5 (DCB+VTF). Maps reveal an infarcted region in the right hemisphere particularly evident in the $$$s$$$-map.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
1474