Yan Dai1, Xun Jia2, Yen-peng Liao1, and Jie Deng1
1University of Texas Southwestern Medical Center, Dallas, TX, United States, 2Johns Hopkins University, Baltimore, MD, United States
Synopsis
Keywords: Diffusion Modeling, Modelling, Quantitative Imaging, Diffusion
Motivation: Estimating IVIM diffusion MRI (dMRI) parameters through non-linear fitting is challenging due to the inherent ill-posed mathematical problem, causing large uncertainty and poor reproducibility.
Goal(s): Understand IVIM-dMRI model properties and derive a more reliable metric than individual IVIM parameters.
Approach: Given a dMRI protocol, we investigated diffusion decay signal distributions for various IVIM parameters via numerical simulations. Employing dimension reduction, we defined a new parameter that captured the largest decay signal variation and is hence most robust against noise.
Results: A new metric was proposed to have the best achievable robustness for IVIM model parameter fitting.
Impact: We proposed
a new metric that attains the best achievable robustness for IVIM model parameter
fitting. This study addressed the large
uncertainty and poor reproducibility issue in IVIM fitting.
Introduction
In the Intravoxel Incoherent Motion(IVIM) model of diffusion MRI(dMRI), signal of each pixel is fitted in a bi-exponential
decay form as a function of diffusion weighting $$$b$$$ to separate cellular diffusion component $$$(D_t)$$$ and pseudo-perfusion component $$$(F_p,D_p)$$$ as defined in Eq.(1)1. $$S=F_p*exp(-b*D_p)+(1-F_p)*exp(-b*D_t)\;\;\;\text{(1)}$$
For clinical applications as reliable
imaging biomarkers, it is crucial that the derived model parameters $$$(F_p, D_t, D_p )$$$ are robust against noise and sensitive2. However, it
is known that the parameters obtained are often
subject to a large uncertainty, particularly in the context of low
signal-to-noise ratios(SNRs)3. This uncertainty arises from the
ill-conditioned mathematical problem of bi-exponential model fitting4,5. To overcome this problem, we
conducted an analysis to understand inherent mathematical properties of the
IVIM-dMRI model. We also devised a novel metric by recombining individual
fitted IVIM model parameters. The new metric attains the best achievable
robustness in the presence of measurement noise. This metric can potentially
serve as a biomarker, enhancing the utility of IVIM model in clinical
applications. Methods
We considered IVIM model parameters in the
range of clinical relevance $$$(F_p\in (0,0.3],D_t\in (0,3\times 10^{-3}]mm^2/s,D_p\in (3\times 10^{-3},10\times 10^{-3}]mm^2/s)$$$. Each parmeter was equally sampled through the
range, generating 2250 tissue types defined by $$$(F_p,D_t,D_p)$$$.
For a dMRI acquisition protocol with eight $$$b$$$-values $$$(b=0,5,50,100,200,500,800,1000 s/mm^2)$$$, we computed the dMRI signal of each tissue,
with the signal $$$S_0$$$ at $$$b=0s/mm^2$$$ set to unity. These generated 2250 dMRI signals each represented as a vector in
the 8-D signal space.
To
understand the data distribution of dMRI signal, we performed principal
component analysis(PCA) to explore the feasibility of approximating the data
by a low-dimensional linear space, and if so, define the direction
corresponding to the largest variation of signal(first principal component).
We further projected the dMRI signal variation to this principal component and
mapped this value back to the 3D parameter space $$$(F_p,D_t,D_p)$$$. A linear and a quadratic function was used
to capture
the variation in decay signal.
For application, three IVIM parameters were calculated via
non-linear regression with the boundary set as $$$F_p\in (0,1),D_t\in (0,3\times 10^{-3}]mm^2/s,D_p\in (3\times 10^{-3},+\infty)mm^2/s$$$, which were then combined through the
derived linear or quadratic function to form a new metric. This
metric is expected to have the most achievable robustness against noise perturbation to dMRI signal and highest sensitivity to signal variations among all possible linear or quardric combinations
of IVIM model parameters, as variation of this metric is directly linked to
the largest variation
of the dMRI signal and conversely, small noise perturbation in dMRI signal does
not cause large variation of this metric.
Results
The first three principal components(PC) account
for 98.0%,1.9%,0.1% of the total signal variation. Projecting dMRI
signals to the first PC and mapping the value to the 3D IVIM
parameter space yielded a colormap representing the signal variation in the parameter space(Fig.1a). The signal variation was fitted by a linear
function as $$$H_{1D}=-0.235\frac{F_p}{0.3}-1.243\frac{D_t}{0.003}-0.084\frac{D_p}{0.01}+0.794$$$, with a relative mean squared error(RMSE) of
0.016. Considering the small coefficient for $$$\frac{D_p}{0.01}$$$ compared to that for $$$\frac{F_p}{0.3}$$$ and $$$\frac{D_t}{0.003}$$$ and the unstable fitting of $$$D_p$$$ that frequently
reaches the upper bound, $$$D_p$$$ was
omitted, resulting in $$$H_{1D}=0.235\frac{F_p}{0.3}+1.243\frac{D_t}{0.003}$$$. Similarily, the quadratic metric was generated as $$$H_{2D}=-1.256(\frac{D_t}{0.003})^2+0.235\frac{F_p}{0.3}+2.504\frac{D_t}{0.003}$$$ with RMSE 0.003. Both the linear and quadratic functions
represented the signal variations well(Fig.1b and 1c), while the quadratic
function is more accurate numerically.
In a brain tumor patient(Fig.2) and a cervix
cancer patient(Fig.3), the tissue contrast in $$$H_{1D}$$$ and $$$H_{2D}$$$ maps provided better image SNR and constrast and
enhanced tissue delineation compared to individual IVIM parameters. Specifically,
intratumoral heterogeity and tumor boundary can be better visualized on the new
metric maps compared with $$$D_t$$$ map and apparent diffusion coefficient(ADC) map. No significant
difference was observed between $$$H_{1D}$$$ and $$$H_{2D}$$$ maps. Discussion and Conclusion
To further explore the impact of the range of $$$D_p$$$ on the $$$H_{1D}$$$, we increased the sampling upper bound of $$$D_p$$$ from $$$10\times 10^{-3}mm^2/s$$$ to $$$1mm^2/s$$$ in signal generation, and found the coefficient for $$$D_p$$$ decreased
as the upper bound increased. This indicates $$$D_p$$$ does not have much impact on signal differentiation, further
supporting ignoring the $$$D_p$$$ term in $$$H_{1D}$$$.
Since $$$H_{1D}$$$ and $$$H_{2D}$$$ maps provided similar contrast, we prefer the linear metric $$$H_{1D}$$$ for computational simplicity. In this study, we
explored the collinearity of IVIM parameters based on the data distribution of
diffusion decay signals with various ground truth through numerical simulation.
We developed a novel metric, derived as a linear or quadratic combination of
individual IVIM parameters, which attains the best achievable robustness among
possible combinations. The correlation of the new metric with underlying biological
properties requires futher investigations. Acknowledgements
No acknowledgement found.References
- Bihan, D. L.
et al. MR imaging of intravoxel incoherent motions: application to
diffusion and perfusion in neurologic disorders. Radiology 161, 401-407,
doi:10.1148/radiology.161.2.3763909 (1986).
- Novikov, D. S., Kiselev, V. G. & Jespersen,
S. N. On modeling. Magnet Reson Med 79, 3172-3193 (2018).
- Iima, M. & Le Bihan, D. Clinical intravoxel
incoherent motion and diffusion MR imaging: past, present, and future. Radiology 278, 13-32 (2016).
- Charles, C. N. Reliability and Uncertainty in Diffusion MRI Modelling, (2016).
- Landaw, E. & DiStefano 3rd, J.
Multiexponential, multicompartmental, and noncompartmental modeling. II. Data
analysis and statistical considerations. American
Journal of Physiology-Regulatory, Integrative and Comparative Physiology 246, R665-R677 (1984).