5061
To evaluate the effect of different initial guess selection approaches on quantitative analysis of DCE-MRI data of brain tumor patients1Indian Institute of Technology Delhi, New Delhi, India, 2KTH Royal Institute of Technology, Stockholm, Sweden, 3University of Pennsylvania, Philadelphia, PA, United States, 4Fortis memorial research institute, Gurugram, India, 5AIIMS, New Delhi, India
Synopsis
Quantitative analysis of dynamic-contrast-enhanced(DCE)-MRI data using various tracer kinetic models is widely used in cancer diagnosis and follow-up. In general, voxelwise model fitting using nonlinear-least-square method requires a long processing time depending upon image-resolution, data noise, choice of initial guess, model type and computer-platform. In this study, we proposed a tissue specific initial guess selection approach, for the voxel wise fitting using nonlinear–least-square method, which substantially reduced computation-time without compromising accuracy of parameters compared to regular global initial guess approach. It also performed better than recently proposed Image-Downsampling-Expedited-Adaptive-Least-squares fitting approach. Parallel-processing was also implemented to further reduce the time
Introduction
Dynamic-contrast-enhanced(DCE)-MRI data is analyzed quantitatively using various tracer-kinetic models2,4,7 to obtain clinically important physiological information. Voxel-wise quantitative analysis is time consuming process mainly due to nonlinear-least-square based fitting approach1,2. In general, a global initial guess is used for voxelwise fitting of a kinetic model. Since the accuracy of fitting and the time required for fitting depends on this initial guess parameter, its selection needs experience. Along with this, lower and upper bounds of the parameters also play role in fitting. Recently, an Image-Downsampling-Expedited-Adaptive-Least-Squares(IDEAL) fitting algorithm was implemented on CEST MRI quantification3, which would substantially downsample the image and use the values from previously downsampled image to fit the data, which will substantially create separate initial guess for each voxel. There are large variations in the shape of DCE-MRI data curves for different tissues. However, within a tissue variations are small. Therefore, we hypothesis that a tissue specific initial guess parameter can reduce processing time without affecting accuracy of parameters. In the current study, we have proposed a tissue specific fitting approach, where the initial fitting parameters and boundaries are constrained for each of segmented Gray Matter (GM), White Matter (WM), CSF and Tumor tissues. This method was also compared with IDEAL approach and with conventional fitting approach. In the current study, we used Generalized-Tracer-Kinetic-Model(GTKM)7 and Leaky-Tracer-Kinetic-Model(LTKM)4 for DCE-MRI analysis.Methods
In this IRB approved retrospective study, DCE-MRI data of five patients with brain tumor were acquired using a 15-channel coil at 3.0T (Ingenia, Philips, The Netherlands) in addition to conventional MRI data. The parameters used for image acquisition were: FOV=240×240mm2, number of slices=12, number of time points=32, slice thickness=6mm and acquisition matrix=256×256. Data analysis was performed using in-house developed programs in MATLAB based software tool. DCE-MRI signal intensity curves were converted to concentration time curves using T1 map4. Voxelwise concentration curves were analyzed using GTKM and LTKM models. Local AIF function was used during fitting. In the proposed approach, the brain was initially segmented into GM, WM, CSF and Tumor tissues. Each of four segmentation masks were considered as individual ROIs and average concentration time curve (Ct) was generated for each ROI. These average Ct curves were fitted with GTKM and LTKM model using trust region reflective algorithm, inbuilt in MATLAB. The parameters from these average Ct fitting were used as initial guesses for all the voxels in respective tissue. Moreover, lower and upper bounds were constraints based upon these tissue specific initial guess parameters. The IDEAL based approach was also implemented for performing GTKM and LTKM models. Along with execution time, goodness-of-fit(R2) and ROI based analysis were carried out for comparison purpose. Parallel processing toolbox of MATLAB was also used to further speedup the fittingResults
Figure-1 shows the Quantitative maps of GTKM(Vp, Ktrans, Ve) obtained using all three approaches, which visually appeared similar. The R2 values for all three approaches was ~0.98 in tumor tissues. Figure-2 shows histogram plots of Ktrans values in the tumor ROI, which clearly show better results for the proposed tissue specific initial guess approach compared to IDEAL approach. Our proposed approach resulted in reduction of scan time approximately by a factor of two times compared to conventional fitting approach. However, IDEAL approach has slight increase in processing time. Comparison of scan time for a single slice is shown in figure-3. Moreover, as shown by coefficient of variation (table-1) and histogram of Ktrans (figure-2) in tumor ROI, IDEAL approach shows slight overestimation compared to regular approach, while the results from proposed tissue specific approach stayed close.Discussion
In the estimation of initial fitting parameter for the quantitative analysis of pharmacokinetic parameters, the proposed tissue specific fitting approach produce a substantial reduction in execution time and without compromising accuracy of parameters. The maps obtained from proposed fitting approach and IDEAL approach visually appear to be smoother than the maps obtained from conventional fitting approach. However, IDEAL approach takes more processing time and it overestimates the tracer kinetic parameters compared to other two approaches. This could be because of large variations in the DCE-MRI data curve. Hence it is observed that the proposed approach is more suitable for quantitative DCE-MRI analysis of brain tumor compared to regular approach as well as IDEAL approach.Conclusion
By analyzing the above mentioned factors, the initial guess selection based on the proposed method was found to preserve accuracy by achieving reduction in processing time. This can substitute the regular fitting method based on a single initial guess for all the voxels in DCE-MRI analysis of human brain.Acknowledgements
The Authors acknowledge technical support of Philips India Limited and Fortis Memorial Research Institute Gurugram for MRI data acquisition. The authors thank Dr. Pradeep Kumar Gupta and Mamta Gupta for data handling. This work was supported by Science and Engineering Research Board (IN) (YSS/2014/000092).References
[1] Tofts PS, Kermode AG (1991) Measurement of the blood-brain barrier permeability and leakage space using dynamic MR imaging. 1. Fundamental concepts. Magnetic Resonance in Medicine 17:357-67.
[2] Tofts PS (1997) modeling tracer kinetics in dynamic Gd-DTPA MR imaging. Journal of Magnetic Resonance Imaging 7:91-101.
[3] Zhou, Iris Yuwen, et al. "Quantitative chemical exchange saturation transfer (CEST) MRI of glioma using Image Downsampling Expedited Adaptive Least-squares (IDEAL) fitting." Scientific Reports 7.1 (2017): 84.
[4] Sahoo P, Gupta PK, Awasthi A, Pandey CM, Patir R, Vaishya S, Saha I, Gupta RK. Comparison of actual with default hematocrit value in dynamic contrast enhanced MR perfusion quantification in grading of human glioma. Magn Reson Imaging. 2016; 34: 1071-7.
[5] Hsu, Yu-Han H., et al. "GPU-accelerated compartmental modeling analysis of DCE-MRI data from glioblastoma patients treated with bevacizumab." PloS one 10.3 (2015): e0118421.
[6] D'argenio, David Z., and Alan Schumitzky. "A program package for simulation and parameter estimation in pharmacokinetic systems." Computer programs in biomedicine 9.2 (1979): 115-134.
[7] Singh, Anup, et al. "Quantification of physiological and hemodynamic indices using T1 dynamic contrastâenhanced MRI in intracranial mass lesions." Journal of Magnetic Resonance Imaging: An Official Journal of the International Society for Magnetic Resonance in Medicine 26.4 (2007): 871-880.
Figures
