Dominick Romano1,2, Qihao Zhang1,2, Ilhami Kovanlikaya2, Pascal Spincemaille2, and Yi Wang2,3
1Biomedical Engineering, Cornell University, New York, NY, United States, 2Radiology, Weill Cornell Medical College, New York, NY, United States, 3Biomedical Engineering, Cornell University, Ithaca, NY, United States
Synopsis
Keywords: Quantitative Imaging, Perfusion, Regularization
This study compared L1 and L2 regularized Quantitative
Transport Mapping (QTM
1-3) of dynamic contrast
enhanced (DCE) MRI in breast and neck tumor using image quality
scoring. Improved consistent soft tissue and lesion characterization was
observed when using the L1 norm.
Introduction
Advanced stage tumors are known to alter the surrounding vasculature.
This causes increased uptake and blood flow, which can be visualized with
dynamic contrast enhanced MRI imaging (DCE-MRI), which encodes physiological
transport parameters. DCE-MRI can be modeled using the convection-diffusion
equation, in a method called Quantitative Transport Mapping (QTM1-3). QTM is ill-posed
and requires regularization to achieve consistent results. In this study, we
compare $$$L_{1}$$$ and $$$L_{2}$$$ regularized QTM on breast and
neck tumor data. The images were evaluated using a scoring system by an
experienced radiologist.Theory
The DCE-MRI signal can be modeled by the simplified
convection-diffusion equation3: $$\partial_{t}c(\textbf{r},t)=-\nabla
\cdot (c(\textbf{r},t)\textbf{u}(\textbf{r}))$$
The $$$L_{1}$$$ regularized QTM inverse problem is defined as
follows:
$$\textbf{u}^{*}=\underset{u}{\operatorname{argmin}}\sum_{k=1}^{N-1}\sum_{k=1}^{N-1}\lVert
\partial_{t}c(\textbf{u},t) +\nabla \cdot
(c(\textbf{r},t)\textbf{u}(\textbf{r}))\rVert_{2}^{2}+\lambda \lVert \nabla
\textbf{u}\rVert_{1}$$
And $$$L_{2}$$$ QTM is defined as follows:
$$\textbf{u}^{*}=\underset{u}{\operatorname{argmin}}\sum_{k=1}^{N-1}\sum_{k=1}^{N-1}\lVert
\partial_{t}c(\textbf{u},t) +\nabla \cdot
(c(\textbf{r},t)\textbf{u}(\textbf{r}))\rVert_{2}^{2}+\lambda \lVert \nabla
\textbf{u}\rVert_{2}^{2}$$
Once $$$\textbf{u}$$$ is obtained from the QTM reconstruction,
we then report the speed map defined by $$u^{*}=\sqrt{\textbf{u}\cdot
\textbf{u}}$$
Regularization weights were selected from the L-curve method.Methods
For the neck tumor dataset, 88 subjects diagnosed with
Nasopharyngeal carcinoma (NPC) with a DCE-MRI acquisition were selected for QTM
processing. The scanning parameters at 3T (Siemens, Erlangen, Germany):
FLASH/vibe sequence; time points = 50; dt = 4.9s; TR=4.9ms; TE=1.47ms; flip
angle = 9 degrees; image orientation: axial/transversal; Phase FOV = 75%;
bandwidth = 400 Hz; dz = 4mm; slice spacing = 0mm; in-plane FOV = 180x180 mm2;
in-plane matrix = 192x144; Acquisition Time = 245s
The breast
tumor cohort consisted of 26 subjects. The subjects were scanned at 3T with an
8-channel breast coil with acquisition parameters: time points = 5; dt = 15.4s;
TR=3.95 ms; TE=1.7 ms; flip angle = 10; image orientation: axial/transversal;
in-plane resolution = 0.71mm; dz = 1.8mm.
The resulting
speed maps from $$$L_{1}$$$ and $$$L_{2}$$$ are compared with DCE-MRI as a
reference. An experienced radiologist (I.K.) scored the images on the following
scale: $$$L_{1}$$$ preferred, $$$L_{2}$$$ preferred, and equal preference. The rationale
for each case preference was based on the following: 1) Lesion intensity should
not be exaggerated, 2) Soft tissue contrast should be considered, 3) QTM
reconstructed lesion should not appear artifactual, 4) QTM reconstructed
structures should not be exaggerated.Results
In NPC cases, $$$L_{1}$$$ regularized maps were preferred in
81.8% of the cases, with $$$L_{2}$$$ regularized maps were preferred in 10.2$
of the cases, and 8.0% were equally preferred (Table 1). This may be explained
by the fact that $$$L_{1}$$$ tissue maps present consistent characterization of
the normal appearing tissue, while the $$$L_{2}$$$ speed maps provide
artifactual normal appearing tissue (Figure 1). Furthermore, $$$L_{1}$$$ speed
maps are consistently preserved at and within the lesion boundary when compared
with $$$L_{2}$$$ maps.
In the breast tumor dataset,
$$$L_{1}$$$ was preferred in 73.1% of the cases, $$$L_{2}$$$ was preferred in
3.8% of the cases, and the remaining 21% of cases were preferred equally (Table
2). It was found that the normal appearing tissue and lesion characterization
was exaggerated in the $$$L_{2}$$$ computed speed maps (Figure 2).Discussion
Previous studies have found $$$L_{1}$$$
regularization of perfusion models perform best in numerical simulations and
acquired CT data, especially in lower SNR cases4-7. Furthermore, these works found that
$$$L_{1}$$$ regularization localized perfusion quantities to tissue types, in
which our work falls into agreement. In MRI, many studies report piecewise
constant CBF between white and grey matter8-10, which supports the choice for the $$$L_{1}$$$
norm. Notably, our work fills a gap of comparing regularization methods for convection-diffusion
based modeling of time-resolved perfusion imaging signals. For the QTM inverse
problem, we find improved consistent soft tissue and lesion characterization
when using the L1 norm.Acknowledgements
No acknowledgement found.References
- Huang
W, Zhang Q, Wu G, Chen PP, Li J, McCabe Gillen K, Spincemaille P, Chiang GC,
Gupta A, Wang Y, Chen F. DCE-MRI quantitative transport mapping for noninvasively
detecting hypoxia inducible factor-1α, epidermal growth factor receptor
overexpression, and Ki-67 in nasopharyngeal carcinoma patients. Radiother Oncol
2021;164:146-154.
- Zhang
Q, Spincemaille P, Drotman M, Chen C, Eskreis-Winkler S, Huang W, Zhou L,
Morgan J, Nguyen TD, Prince MR, Wang Y. Quantitative transport mapping (QTM)
for differentiating benign and malignant breast lesion: Comparison with
traditional kinetics modeling and semi-quantitative enhancement curve
characteristics. Magn Reson Imaging 2022;86:86-93.
- Zhou
L, Zhang Q, Spincemaille P, Nguyen TD, Morgan J, Dai W, Li Y, Gupta A, Prince
MR, Wang Y. Quantitative transport mapping (QTM) of the kidney with an
approximate microvascular network. Magnetic Resonance in Medicine
2021;85(4):2247-2262.
- Fang
R, Chen T, Sanelli PC. Towards robust deconvolution of low-dose perfusion CT:
sparse perfusion deconvolution using online dictionary learning. Med Image Anal
2013;17(4):417-428.
- Fang
R, Chen T, Sanelli PC. Tissue-specific sparse deconvolution for low-dose CT
perfusion. Med Image Comput Comput Assist Interv 2013;16(Pt 1):114-121.
- Fang
R, Karlsson K, Chen T, Sanelli PC. Improving low-dose blood-brain barrier
permeability quantification using sparse high-dose induced prior for Patlak
model. Med Image Anal 2014;18(6):866-880.
- Ruogu
F, Shaoting Z, Tsuhan C, Sanelli PC. Robust Low-Dose CT Perfusion Deconvolution
via Tensor Total-Variation Regularization. IEEE Trans Med Imaging
2015;34(7):1533-1548.
- Brown
GG, Clark C, Liu TT. Measurement of cerebral perfusion with arterial spin
labeling: Part 2. Applications. J Int Neuropsychol Soc 2007;13(3):526-538.
- Buxton
RB. Quantifying CBF with arterial spin labeling. J Magn Reson Imaging
2005;22(6):723-726.
- Fan AP, Guo J, Khalighi MM, Gulaka PK,
Shen B, Park JH, Gandhi H, Holley D, Rutledge O, Singh P, Haywood T, Steinberg
GK, Chin FT, Zaharchuk G. Long-Delay Arterial Spin Labeling Provides More
Accurate Cerebral Blood Flow Measurements in Moyamoya Patients: A Simultaneous
Positron Emission Tomography/MRI Study. Stroke 2017;48(9):2441-2449.