Radovan Jiřík^{1}, Michal Bartoš^{2}, Karel Souček^{3,4,5}, Eva Dražanová^{1}, Lucie Krátká^{1}, Lenka Dvořáková^{1}, Torfinn Taxt^{6}, and Zenon Starčuk, jr.^{1}

A method to improve reliability of DCE-MRI is presented. It is based on the assumption that neighboring voxels have similar physiological properties. It is formulated as a total variation regularization of the perfusion parameter maps. The method is applied to the adiabatic approximation of the tissue homogeneity model (ATH). Evaluation was done on data of five tumor bearing mice recorded using two contrast agents with different molecular weights. The resulting perfusion parameter maps were more spatially consistent with preserved edges.

Dynamic Contrast-Enhanced (DCE) MRI is used for assessment of microvascular tissue parameters by approximating the tissue contrast agent (CA) concentration curves with a pharmacokinetic model. Compared to standard pharmacokinetic models (Tofts, extended Tofts), use of advanced pharmacokinetic models provides also estimates of blood plasma flow (*F _{p}*) and permeability surface area product (

Here, the assumption that neighboring voxels have similar physiological properties is applied. Such spatial prior has been rarely published^{2-4}. It has been formulated using the Gaussian Markov random field^{2,3} or L_{2} regularization of the perfusion parameter maps^{4}. It has been applied to the Tofts model^{2,3} and its multicompartment variation^{4}. These models do not provide separate estimates of *F _{p}* and

The tissue concentration curves are modeled as a convolution of the arterial input function (AIF) and the impulse residue function (IRF). The AIF is estimated using blind multichannel deconvolution^{6}. The IRF is modeled using the ATH model, based on the comparison of the advanced pharmacokinetic models ATH and 2CXM evaluated on the same tumor tissue type as used here^{7}. The perfusion parameters (parameters of the IRF) are estimated in an iterative algorithm (5 iterations used). In the first iteration, the standard approximation of the voxel-specific tissue concentration curves by the convolutional model is performed as minimization of the sum of squared differences (SSD) between the tissue concentration curve and the convolutional model. In the subsequent iterations, the SSD criterion of the approximation task is extended by adding five TV regularization terms, each for one parameter of the IRF: *F _{p}*, extraction fraction, interstitial volume, capillary mean transit time and the bolus arrival time. Each TV term is a sum of absolute value differences between the sought perfusion parameter of the processed voxel and the perfusion parameters estimated in the neighboring voxels in the previous iteration. The TV regularization is known to impose partial smoothness while preserving edges. The weight of all TV terms is the same (set experimentally) as the perfusion parameters are scaled to approximately the same range.

The method was evaluated with five BALB/C mice after subcutaneous implantation of tumor cells (CT26), approved by the National Animal Research Authority. Two CA boluses were administered within one examination (delay of 30 min. in between): high-MW CA (GadoSpin P, Miltenyi Biotec, Germany) and low-MW CA (Magnevist, Bayer HealthCare Pharmaceuticals, Germany).

A 9.4T BioSpin (Bruker Biospin MRI, Germany) scanner was used; dynamic recording: 2D FLASH, TR/TE 14/2.5ms, flip angle 25deg., matrix 128x96, axial slice, thickness 1mm, sampling interval 1.05s, acquisition time 13min; 15 calibration images before bolus administrations: TR=14,30,50,100,250,500ms. The dynamic recordings were converted to CA concentration using calibration images^{8} (multiple-TR modification used here instead of the multiple-flip-angle method^{8} as more reliable).

This work was supported by GACR (1613830S) and MEYS (LO1212) projects.

1. Sourbron SP, Buckley DL. Tracer kinetic modelling in MRI:Estimating perfusion and capillary permeability. Phys Med Biol. 2012;57(2):R1–33.

2. Kelm BM, Menze BH, Nix O et al. Estimating Kinetic Parameter Maps From Dynamic Contrast-Enhanced MRI Using Spatial Prior Knowledge. IEEE T Med Imag. 2009;28(10):1534–1547.

3. Schmid VJ, Whitcher B, Padhani AR et al. Bayesian Methods for Pharmacokinetic Models in Dynamic Contrast-Enhanced Magnetic Resonance Imaging,” IEEE T Med Imag. 2006;25(12):1627–1636.

4. Sommer JC, Gertheiss J, Schmid VJ. Spatially regularized estimation for the analysis of dynamic contrast-enhanced magnetic resonance imaging data. Stat Med. 2014;33(6):1029–1041.

5. St. Lawrence KS, Lee T-Y. An Adiabatic Approximation to the Tissue Homogeneity Model for Water Exchange in the Brain: I. Theoretical Derivation. J Cerebr Blood F Met. 1998;18(12):1365–1377.

6. Jirík R, Soucek K, Mézl M et al. Blind Deconvolution in Dynamic Contrast-Enhanced MRI and Ultrasound. Conf. Proc. IEEE Eng Med Biol Soc. 2014:4276–9.

7. Jirík R, Taxt T, Soucek K et al. Comparison of the ATH and 2CXM Models UsingLow- and High-Molecular-Weight Contrast Agents in DCE-MRI. Abstracts of ESMRMB Conf. 2016;S447-448.

8. Li KL, Zhu XP, Waterton J, Jackson A. Improved 3D Quantitative Mapping of Blood Volume and Endothelial Permeability in Brain Tumors. J Magn Reson Imag. JMRI. 2000;12(2):347–57.

Figure 1. Anatomy image of mouse 1, T2-weighted RARE sequence, TR/TE=3500/36 ms, FOV 2.3x3.5 mm, image matrix 256x256, axial slice, slices thickness 0.7 mm. Dotted area – tumor, dashed area – erector spinae muscles.

Figure 2. Example of perfusion parameter maps of mouse 1 obtained with no regularization (left) and TV regularization (right).

Figure 3. Example of scatter plots for mouse 1 obtained with no regularization (left) and TV regularization (right). Each point of the scatter plot corresponds to a voxel in the tumor area (based on manual delineation) outside nercosis (deduced from the rate of the efflux from the interstitial space to blood, *k*_{ep}>0.2min-1) and reasonable signal to noise ration (SNR>1) and its perfusion parameter values estimated from the low-MW (horizontal axis) and high-MW (vertical axis) contrast agents.

Table I. Mean and standard deviation (SD) of correlation and regression coefficients of 5 mice relating the perfusion parameters estimated from the low-MW a high-MW recordings.