TwinTree Insert

06-09 Parallel Imaging

n advanced approach for image acqui­sition uses multiple RF receiver coils with independent RF pathways, known as phased array or synergy surface coil ar­rays (see Chapter 3).

In these arrays, each individual coil pro­duces a signal of its own. Several in­de­pen­­dent image data sets can be acquired at a time and, after post­pro­ces­sing, syn­the­tiz­ed into a single image.

Commonly, a combination of such over­lapping multiple receiver coil elements is uti­liz­ed to improve the signal-to-noise ratio (Figure 06-23).

Figure 06-23:
Parallel imaging with an array of two coils: each coil provides half of the field-of-view of the final image; thus, only half of the phase-encoding steps have to be ac­quired.

In the sensitivity profiles of the individual­ receiver coil elements there is spatial in­­for­ma­tion con­­tain­­ed, and multiple phase-encoded data can be derived at the same time. Thus, the number of gra­dient-based spatial encoding steps can be decreased and conventional Fourier encoding reduced.

This procedure is based upon dedicated adjuvant reconstruction algorithms, among them SENSE (Table 06-01) [⇒ Pruessmann 1999], and SMASH. These al­go­rithms must not be confused with pulse sequences although their acronyms sound similar. In principle, they can be applied to any imag­ing sequence; con­trast behavior does not change.

Table 06-01:
Adjuvant parallel imaging reconstruc­tion algorithms.
The names are different, but the approach is the same.

Both SMASH and SENSE algorithms reconstruct missing data to obtain an image with­out backfolding artifacts. The way that this is actually done re­pre­sents the main dif­fe­ren­ce between the two techniques. SMASH will perform this calculation on the raw data, before the Fourier trans­form, while SENSE will work with the images ob­tain­ed from the respective coils. SENSE and SMASH are based on knowledge of the sen­si­ti­vi­ty pro­fi­les of the individual coil ele­ments.

This knowledge is acquired either through a very low resolution 3D volume scan (64×64×64 matrix) which then can be used for all consequent scans independent of ori­en­ta­tion, or through adding some ad­ditional phase encoding steps to each re­­du­ced ac­qui­si­tion. Therefore, from a practi­cal point of view, the first solution will be time effective if more scans are per­for­med, the latter is to be preferred if only one scan is wanted.

SENSE, PILS, and ASSET reconstruct the final image from the sub-images pro­duced by each coil after the Fourier trans­formation in the image domain, whereas GRAPPA re­con­structs the Fourier plane of the image from the fre­quen­cy signals of each coil be­fore the Fourier transformation.

SENSE and its relatives work with most pulse sequences and clinical ap­pli­ca­tions. More­over, they also allow the user to choose between either increased spa­tial or tem­po­ral re­so­lu­tion [⇒ Blaimer 2004, ⇒ Larman 2007].

The downside of parallel imaging is that the signal-to-noise ratio is reduced com­­par­ed to phased-array imaging. The reduc­tion in signal-to-noise does no longer fol­­low the common square-root dependence because of the non-Cartesian sampling and the noise correlation between pixels.

spaceholder redCritical Remarks. Since one can acquire an image with a full matrix in 2-3 mi­nu­tes using GRE and TSE sequences, there is often no real need for parallel imag­ing and therefore a loss of signal-to-noise.

It is often used in dynamic imaging and 3D/4D imaging where saving time is im­por­tant and one can afford to trade off some signal.

spaceholder redCompressed Sensing. To accelerate imaging beyond the gain achieved in pa­ral­lel imag­ing, additional processing of image data by Compressed Sensing was in­tro­duc­ed as a concept already used in a number of non-medical applications. It pro­du­ces an image based on as few variables as possible: sparse data [⇒ Donoho 2006, ⇒ Lus­tig 2007, ⇒ Mairal 2012. An image is considered as 'sparse' when its in­for­ma­ti­o­nal con­tent is re­pre­sent­ed by only a few pixels, while the contribution of the remaining ma­jo­ri­ty of pixels is close to zero; the image is reconstructed from un­der­sampl­ed mea­sure­ments to a "nearly perfect state" [⇒ Geethanath 2013].

There are no studies yet whether relevant diagnostic data are lost by such pro­ce­du­res.