Karl Kiser1, Jin Zhang1, and S. Gene Kim1,2
1Radiology, Weil Cornell Medical College, New York, NY, United States, 2Center for Biomedical Imaging (CBI), Center for Advanced Imaging Innovation and Research (CAI2R), New York University School of Medicine, New York, NY, United States
Synopsis
Quantitative analysis of
MRI image features for estimating tumor grading and treatment response is a
growing area of research, however a lack of reproducibility and validation
present a major challenge in the field. Our study investigates how spatial resolution affects texture features of DCE-MRI images by comparing features
generated from 3D isotropic high resolution kinetic parameter maps with typical
thick slice maps. We demonstrate that 3D textural features can differ by
several orders of magnitude when extracted from isotropic versus
thick slice images. These findings have potentially significant implications in
the predictive capabilities of texture features.
Introduction
Tumor heterogeneity
captured by image texture features derived from DCE-MRI images have been shown to estimate
tumor grading and response to treatment,1-5 however, typical 2D multi-slice acquisition
methods fall short on capturing the complexity of a tumor microenvironment. 3D
texture features have been shown to be more robust than 2D texture features but
are limited as 3D features require the up sampling of thick slice images to
isotropic resolution. Our study compares the DCE-MRI kinetic parameter maps
acquired with 3D isotropic high resolution to those with thick slices to
investigate the impact of resolution on image feature analysis.Methods
Six to eight-week-old C57BL6 mice (n=8) with GL261 mouse glioma xenograft
models were used in this study. MRI scans were performed on a Bruker 7T
micro-MRI system, with a 1H four-channel phased array receive-only
MRI CryoProbe with a volume transmit coil. The GRASP method6 was used to acquired dynamic contrast-enhance
data (TR/TE = 4 / 0.028
ms) with image matrix = 128 x 128 x
128, up sampled to 256x256x256, and field of view = 20 x 20 x 20 mm3.
In order to achieve contrast encoding for τi measurement, this sequence was continuously run to acquire 154,080 spokes with two flip angles (51,360
spokes for each flip angle segment 8o - 25o - 8o)
for 10 minutes and 13 seconds. Pre-contrast T1
mapping was obtained using the 3D-UTE-GRASP sequence with 38,328 spokes (12,776
spokes for each flip angle segment 8o - 2o - 12o),
for a total acquisition time of 153 s at the same resolution as the DCE scan.
The joint compressed sensing and parallel imaging reconstruction was
implemented based on the 3D-UTE-GRASP algorithm6 with temporal frame
resolution T = 5 s/frame. Arterial input function (AIF) was obtained
following the Principal Component Analysis (PCA) method used in our previous
study7 with the independently measured T1 map. The
T1 weighted images were also used to manually segment whole
tumors. Pharmacokinetic model analysis was carried out for the whole tumor with
the same Two Compartment Exchange Model (TCM)8 and Three Site Two
Exchange (3S2X)9 Model for τi estimation with the two-flip
angle approach.7 Five parameters, interstitial space volume fraction
(ve), vascular space volume fraction (vp),
blood flow (Fp),
permeability surface area product (PS),
and intracellular water lifetime (τi)
were estimated from the model fit. Transfer constant (Ktrans) was calculated from PS and Fp.
To compare parameter maps from our 3D isotropic high-resolution
images with traditional thick slice DCE-MRI images we subdivided the tumor ROI
into 13 slice sections (0.078 mm slice thickness) and took a vector average of
each slice section to emulate same in plane resolution but with a depth of 1 mm
(0.078 mm x 13 =1.016 mm). Pharmacokinetic modeling was then carried out for
the simulated thick slice images.
Using the open-source
python package PyRadiomics,10 parameter map intensity histogram
features (n=18) and texture features(n=68) were calculated for both isometric
and thick slice images. For 3D texture feature extraction, the thick slice
images were up sampled to isotropic resolution of 256x256x256. Discretization
of parameter maps was performed using a fixed bin width determined by: W
= 2 (IQR) N-1/3; where IQR is the interquartile range
and N is the number of pixels.11 The default PyRadiomics settings
were used for all other parameters. The difference of features between
isometric and thick slice images was defined as the absolute value of the
difference between the log transform of a given feature value.Results and Discussion
Figure 1 depicts volume renderings of the 3D isotropic high resolution
kinetic parameter maps of a mouse glioma tumor. Figure 2 demonstrates the isotropic parameter map slices
(Slice 1 - Slice 13, spatial
resolution of 0.078 x 0.078 x 0.078 mm3), and the thick slice parameter
maps (spatial resolution of 0.078 x 0.078 x 1.016 mm3) estimated from the vector
average of the isometric slices. Figure 3 and 4 are heatmaps of texture feature
differences from isotropic versus thick parameter maps, displaying differences of several
orders of magnitude. These include the texture features largely related
to image heterogeneity, found to be significant in recent cancer imaging
studies. One study demonstrated such features, from thick slice parameter maps, to be most predictive of breast cancer therapy response.1
Of the 9 most predictive features in this study, we found the Co-occurance
Matrix (GLCM) defined features of: difference entropy, contrast, difference variance
and the Run Length Matrix (GLRLM) defined features of: gray-level nonuniformity and long run
emphasis to be impacted by several orders of magnitude when extracted from
isometric maps versus thick slice maps. Furthermore, texture features
relating to heterogeneity correlating with Ki-67 expression in breast tumors,
such as GLCM contrast, correlation and GLRLM long run low grey level emphasis and long run high grey level emphasis are also impacted by being measured from isotropic versus thick slice images.3,4Conclusion
For
3D texture feature analysis, thick slice images must be up sampled for the
analysis. This could create an inference of information, potentially obfuscating
texture features of the image and impacting its predictive ability. To improve
reproducibility in texture features and best capture tumor heterogeneity,
DCE-MRI images would benefit from an isotropic resolution acquisition.Acknowledgements
NIH R01CA160620, NIH R01CA219964, UG3CA22869References
1.
Thibault G, Tudorica A, Afzal A,
Chui SY, Naik A, Troxell ML, Kemmer KA, Oh KY, Roy N, Jafarian N, Holtorf ML,
Huang W, Song X. DCE-MRI Texture Features for Early Prediction of Breast Cancer
Therapy Response. Tomography. 2017;3(1):23-32.
2. Niu Q, Jiang X, Li Q, et al.
Texture features and pharmacokinetic parameters in differentiating benign and
malignant breast lesions by dynamic contrast enhanced magnetic resonance
imaging. Oncol Lett. 2018;16(4):4607-4613.
3. Juan MW, Yu J, Peng GX, Jun LJ,
Feng SP, Fang LP. Correlation between DCE-MRI radiomics features and Ki-67
expression in invasive breast cancer. Oncol Lett. 2018;16(4):5084-5090.
4. Ma W, Ji Y, Qi L, Guo X, Jian X,
Liu P. Breast cancer Ki67 expression prediction by DCE-MRI radiomics features.
Clin Radiol. 2018;73(10):909.e1-909.e5.
5. Liu M, Mao N, Ma H, et al.
Pharmacokinetic parameters and radiomics model based on dynamic contrast
enhanced MRI for the preoperative prediction of sentinel lymph node metastasis
in breast cancer. Cancer Imaging. 2020;20(1):65.
6. Zhang, J., Feng, L., Otazo, R., & Kim, S. G. Rapid
dynamic contrast-enhanced MRI for small animals at 7T using 3D ultra-short echo
time and golden-angle radial sparse parallel MRI. Magnetic Resonance in
Medicine. 2019;81(1), 140–152.
7. Zhang, J., &
Kim, S. G. Estimation of cellular-interstitial water exchange in dynamic
contrast enhanced MRI using two flip angles. NMR in Biomedicine. 2019;32(11),
1–17.
8. Sourbron SP, Buckley DL. Tracer
kinetic modelling in MRI: estimating perfusion and capillary permeability. Phys
Med Biol. 2012;57(2):R1-33.
9. McConnell, H. M. Reaction rates by nuclear magnetic
resonance. The Journal of Chemical Physics. 1958; 28(3), 430–431.
10. van
Griethuysen, J. J. M., Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan,
V., Beets-Tan, R. G. H., Fillon-Robin, J. C., Pieper, S., Aerts, H. J. W. L.
Computational Radiomics System to Decode the Radiographic Phenotype. Cancer
Research. 2017; 77(21)
11. Izenman, A. J. Recent developments
in nonparametric density estimation.
Journal of the American Statistical Association. 1991; 86(413):205-224.
12. Lambin, P. Radiomics Digital Phantom. CancerData.
2016; 41(2), 366–373.