Perfusion MRI: Non-Linear Stochastic Regularization for Deconvolution of Model-Free Arterial Spin Labelling Data
Abstract: Introduction: Arterial spin labelling (ASL) is a non-invasive magnetic resonance imaging (MRI) technique for assessment of perfusion. ASL uses magnetically labelled blood water as a diffusible endogenous tracer to measure the perfusion of brain tissue, i.e. the cerebral blood flow (CBF). Traditionally, ASL samples images at a specific point in time allowing the CBF to be calculated using a model of the signal. With time-resolved ASL, perfusionweighted images are sampled as a function of time, allowing the CBF to be calculated using a model-free approach by means of deconvolution. The QUASAR ASL implementation employs this approach and has previously been used to calculate CBF through deconvolution with block-circulant singular value decomposition (cSVD) resulting in slightly lower perfusion values than those achieved with gold standard methods (such as PET). The use of non-linear stochastic regularization (NSR) for deconvolution has previously been investigated, showing good potential to reproduce realistic tissue residue functions in contrast-agent-based perfusion studies by dynamic susceptibility contrast MRI (DSC-MRI). The NSR method gives the possibility to correct for arterial dispersion and therefore potentially improve the perfusion quantification. In this study, NSR was applied as a deconvolution method for absolute quantification of CBF using time-resolved ASL data obtained by the QUASAR pulse sequence. The aim was primarily to see if the implementation was feasible for ASL data, and secondly to assess the potential of improving the absolute CBF quantification and the retrieved residue functions in ASL experiments. Materials and Methods: Data originated from volunteers participating in a multi-centre reproducibility study of QUASAR. The imaging was carried out with a Philips Achieva 3T MRI unit using the following protocol: 7 slices, 6 mm slice-thickness, 2 mm slice gap, 64×64 matrix, 240×240 mm2 FOV, 35°/11.7° flip angles, TR/TE/ΔTI/TI1 = 4000/23/300/40 ms, 13 inversion times, 84 series, 2.5 SENSE factor, 150 mm labelling thickness, 3.75×3.75 mm2 in-plane resolution and a total scantime of 6 min. To obtain the data required for the deconvolution, the acquired ASL image data were post-processed using in-house-developed software, with some additions and modifications compared to previous post-processing software from the QUASAR studies. The most important addition was the implementation of a fractional segmentation and a subsequent novel method to calculate the equilibrium magnetization of arterial blood, which is an essential parameter that directly scales the CBF. To make use of NSR with ASL data, some modifications compared to the available NSR software for DSCMRI data was needed. One of the main modifications was to model the high measurement noise generally immanent to ASL. Results and Discussion: The implementation resulted almost consistently in smooth, physiologically realistic residue functions. Especially compared to the most common model-free deconvolution method, cSVD, the resolved residue functions had significantly more physiologically realistic characteristics, primarily regarding monotonic decrease, non-negative values and, consequently, no oscillations. Visual inspection of reconvolved tissue signals served as a qualitative validation, and indicated that the method was able to solve a variety of different kinetic signals. The absolute quantification of CBF in grey matter did not differ from previous methods and resulted in problems with under- and overestimations in certain voxels. However, the new implementations of fractional segmentation and estimation of the equilibrium magnetization of arterial blood proved to have the potential to reduce CBF quantification errors due to partial volume effects and the presence of non perfused volumes (e.g., CSF). Conclusion: The implementation of NSR on QUASAR data was successful and encouraging results were obtained. Resolved residue functions showed more realistic characteristics than what is normally achievable with common deconvolution methods. However, the resulting CBF values did not appear to be more realistic than those obtained using previous deconvolution methods. This could partially be explained by the fact that NSR tends to fail if the measured signal is too noisy in a voxel, whereas, for example, cSVD produces a potentially false but non-zero result for those voxels. NSR as a deconvolution method for ASL is likely to require substantial verification and optimization to be considered a complement to the already existing methods. However, ASL sequences are continuously improved, reducing the impact of measurement noise, and computer performance increased, and hopefully this will allow the computationally expensive NSR algorithm to be clinically applicable in the future.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)