This chapter demonstrates a practical methodology to assess MRSI spatial integrity after retrospective reconstruction of both phantom and in-vivo datasets. This allowed investigation into the limits of scan time reduction using CS. Spatial resolution in MRSI is linked to inter-voxel contamination (section 2.3.1). Reduction of inter-voxel contamination is of paramount importance in correctly distinguishing between focal tumours and those which have infiltrated into healthy tissue. The phantom provided a practical means to assess
adjacent voxel contamination by means of the maximum gradient of the perpendicular metabolite edge. The phantom also provided a means of assessment of distant inter-voxel contamination by assessment of the Gibbs ringing at the edge.
The two-compartment gel edge phantom indicates good agreement with the theoretical values Figure 5-1 b) implying that the phantom methodology for measuring the spatial resolution is valid and directly proportional to the Edge Spread function (ERF). It also shows that there is a reduction in spatial resolution with increasing acceleration factor as expected. This figure demonstrates that the spatial resolution of CS-MRSI is consistently above that of the LRR however, and is maintained at 87% ( ± 3%) of the resolution of the FSSR with approximately 70% time saving (x3.3 acceleration), Figure 5-1 b). This implies that inter- voxel contamination of the adjacent voxels is reduced in CS-MRSI. The results of Figure 5-1 c- g) indicate reduced ringing artefact from CS-MRSI as compared to the LRR. This implies less inter-voxel contamination from distant voxels.
The assessment of the in-vivo contrast provided a means of assessment of the inter-voxel contamination in a more realistic clinical scenario. The results indicated that the contrast remains consistent up to a factor of x3.3 but degrades with increasing acceleration factor after this; Figure 5-5 a). The in-vivo methodology appears to only be sensitive to losses in spatial resolution of more than 50%.
The signal to noise ratios indicated on Figure 5-5 b) show that the CS-MRSI maintains a consistent SNR above that of the FSSR at all acceleration factors, and an SNR higher than the LRR up to a factor of x10. The LRR has a higher SNR at x10 acceleration factor but Figure 5-7 f) indicates that there is more inter-voxel contamination into the unsaturated region at this factor. The consistency in the SNR for the CS-MRSI is a reflection of the smoothing property of compressed sensing (Lustig, Donoho et al. 2007; Geethanath 2012).
The Bland Altman plots for the in-vivo dataset Figure 5-6 show that the CS-MRSI has a consistently reduced standard deviation as compared to the equivalent LRR. Some of the plotted points for Choline are above the level of one standard deviation for CS-MRSI at acceleration factors of x5 and x10. This may have implication for the use of CS-MRSI for the investigation of metabolites with lower concentrations such as my-inositol. This reduction in perceived concentration may be as a result of the applied weighting of the TV transform which warrants further investigation.
The reduction in inter-voxel contamination (section 2.3.1) as a result of CS-MRSI for the in- vivo datasets as compared to LRR at x10 acceleration is visible on the typical spectra obtained from the region highlighted in red on Figure 5-7.
In this work, both phantom and in-vivo results indicate an increase in standard deviation for repeated reconstructions as acceleration is increased beyond x3.3. This is likely to be due to the very high under-sampling requirement at these acceleration factors and the resulting lack of true measured data (Barker and Ernst 2013). The maximum gradient measurements were also only based on two data points in the edge response profile. The performance and reliability of CS-MRSI at higher acceleration factors is therefore likely to be improved for higher spatial resolution datasets with larger matrix sizes.
6 Development of digital phantoms for simulating high-resolution
CS-MRSI
The objective of this chapter was to provide a means to produce datasets which overcome the high sampling requirements at high acceleration factors (section 5.3), to provide a more realistic scenario, better statistics and to overcome the problems incurred during real-time MRSI acquisition. The measured ERF in the previous chapter only contained 2 data points, therefore increasing this will aid in the accuracy of the assessment.
The objective can most easily be achieved by creating software simulated, higher MRSI matrix size datasets (digital phantom). A typical MRSI scan takes between 5-10 minutes to acquire. Given that a patient will undergo several MRI scans during one appointment; this is prohibitive in terms of time and cost. Even with an adult volunteer there is still potential for movement artefacts in that time frame. The creation of software simulations of MRSI data can also be designed to provide datasets at a desired resolution, and permit controlled analysis of the effects of different reconstruction techniques in MRSI under very controlled conditions. Such conditions would include magnetic field strength, SNR, lipid contamination, and the addition of features in the MRSI dataset such as tumour spectra and anatomical features of particular shape and dimension. Software simulated MRSI data is not subject to the challenges often encountered during acquisition (section 2.3). An MRSI software simulation system was therefore developed in Matlab version 2012b.