Synopsis

Patient movement during dynamic contrast enhanced breast MRI acquisition can degrade signal enhancement curves. Non-rigid image registration can improve enhancement curves, but computation time can be long, especially if all voxels are sampled for cost function estimation. This work investigates the influence of the percentage of voxels sampled (PS) on goodness-of-fit, kinetic model parameter values and uncertainties. The spatial distribution of parameter values was more strongly influenced by registration and PS compared to global tumor measures. Registration with very low PS values increased parameter uncertainties compared to unregistered. 5 PS provided similar performance to 100 PS with reduced computation time.

Introduction

Methods

DCE-MRI breast images were acquired on a 3T-PET/MRI system (Siemens Biograph mMR) in 8 patients with early stage breast cancer. Three-dimensional fat suppressed fast low angle shot (FLASH) images (spatial resolution=1.0x1.1x2.0mm, time resolution=18s) were acquired from patients prior to and at 28 time points following Gadovist (0.1 mMol/kg) injection. Deformable image registration of the affected breast was performed with 3DSlicer² on a PC running windows 10 (16GB ram, intel i7-4790 CPU). Mattes mutual information was used as a cost metric with a 2cm isotropic control point spacing. Each post-contrast image was registered to the pre-contrast image, and was repeated with PS values of 0.5%, 1%, 5%, 20%, and 100%. Tumours were segmented using Otsu’s method³. The Toft’s kinetic model⁴ was fit voxel-by-voxel to segmented images using a population derived AIF⁵. The root-mean-square-error (RMSE) about the fitted curve, best-fit values, and standard error (σK^trans, σk_ep) of the model parameters (K^trans, k_ep) were extracted from each voxel. Absolute difference maps (ΔK^trans, Δk_ep) were created by taking the absolute value of voxel-by-voxel differences between 100 PS and other PS cases (including unregistered data) in order to assess changes in the spatial distribution of parameter values with PS.

For each parameter (K^trans, k_ep, σK^trans, σk_ep, ΔK^trans, Δk_ep, RMSE) within each patient image, the medians and 90^th percentile over all voxels in the segmented tumour were determined. Pair-wise Wilcoxon rank-sum tests were performed to compare values for 100% PS to those for lower PS and unregistered, except for ΔK^trans and Δk_ep which were compared to 20 PS since they were subtracted from 100 PS.

Results/Discussion

RMSE values for unregistered and low PS (0.5%, 1%) were significantly greater than 100 PS (Table 1) but not significantly different for 5 and 20 PS, indicating an improvement to the goodness-of-fit with registration at sufficiently high PS (5%).

Figure 1a shows that changes in the median K^trans value appear to be small across all PS cases (including unregistered). Differences were not significant for all comparisons with 100 PS (Table 1). However, 90^th percentile K^trans values (Fig 1b) for unregistered images were significantly greater than those from 100 PS. Findings for k_ep were similar (Table 1).

Figure 1c and 1d demonstrate that pixel-by-pixel ΔK^trans were more sensitive to image registration than median and 90th percentile values of K^trans. The differences were significant for unregistered, 0.5 and 1 PS but not 5 PS when compared to 20 PS (Table 1). This trend is reflected in Δk_ep (figure not shown) except that the median value for 5 PS was significantly different from 20 PS. Changes in ΔK^trans and Δk_ep indicate changes in the spatial distribution of corresponding model parameters (K^trans and k_ep).

Regarding σK^trans (Fig. 1e and f), values for some patients were higher (i.e., higher fitting uncertainty) compared to unregistered or higher PS. This is supported by the statistical differences for σK^trans and the median value of σk_ep (Table 1). This trend differs from those of ΔK^trans and Δk_ep which did not increase from unregistered to low PS registered images. We speculate that with registration at low PS, the deviation of the overall DCE curve shapes from that predicted by the Tofts model is reduced (RMSE, Table 1), but residual high frequency fluctuations in the curves leads to high values for σK^trans and σk_ep.

From Table 1, only one significant difference occurred at or above 5 PS. Considering also that any reduction in computation time (Table 2) at PS less than 5 was very small, our results suggest that 5 PS seems to provide a reasonable balance between computation time and registration performance. Limitations of this work include the relatively small number of patients and the possibility that the impact of PS on image registration could depend on imaging parameters such as spatial resolution and signal-to-noise ratio.

Conclusion

Acknowledgements

References

[1] Chebrolu VVN, Shanbhag D, Gupta S, Hervo P, Reid S. Impact of Non-Rigid Motion Correction on Pharmaco-Kinetic Analysis for Breast Dynamic Contrast-Enhanced MRI. 23rd ISMRM Annu Meet. 2015.

[2] Fedorov A, Beichel R, Kalpathy-Cramer J, et al. 3D Slicer as an image computing platform for the Quantitative Imaging Network. Magn Reson Imaging. 2012;30(9):1323-1341. doi:10.1016/j.mri.2012.05.001.

[3] Otsu, N., A Threshold Selection Method from Gray-Level Histograms,"IEEE Transactions on Systems, Man, and Cybernetics, Vol. 9, No. 1, 1979, pp. 62-66.

[4] Tofts PS. Modeling tracer kinetics in dynamic Gd-DTPA MR imaging. Journal of Magnetic Resonance Imaging. 1997 Jan 1;7(1):91-101.

[5] Parker GJM, Roberts C, Macdonald A, et al. Experimentally-derived functional form for a population-averaged high-temporal-resolution arterial input function for dynamic contrast-enhanced MRI. Magn Reson Med. 2006;56(5):993-1000. doi:10.1002/mrm.21066.