MR-based Attenuation Correction for PET/MR - Correction Techniques in Emission Tomography - page 228

FIGURE 11.7: Segmentation of 2-point Dixon sequence[23]. MRI water (A)
and fat (B) images are combined and segmented to generate a 4-class MRI-
based attenuation map that does not include bone (C). (D) CT-based atten-
uation map of the same patient.
sifying bone voxels as air. Only two patients underwent both PET/CT and
MRI examinations. Following MR-AC based on the Dixon method (ignoring
bone) one lesion in the rst patient showed an SUV underestimation of 4:6%
and six lesions in the second patient had an SUV dierence ranging from 6%
to +1%.
Although predominantly used to date for brain imaging, atlas-based meth-
ods can also be applied to whole body imaging, allowing in principle to gener-
ate attenuation maps that include bone attenuation values. However, anatomic
variability is high and it is unlikely that a general, spatial transformation cap-
tures all variables between a template and patient-specific anatomy. Hofmann
et al. [16] presented a machine learning approach that combines the infor-
mation from an atlas registration with local information that is drawn from
small image patches. The method thus reduces reliance on accurate template
to patient registration.
Even in cases where it is impossible to acquire an attenuation image
through a transmission scan (or from an MR image), it is, in principle, possible
to simultaneously estimate the transmission and emission image from emission
data only [6, 29]. This was shown already in 1979 by Censor et al. [6]. Among
others, Nuyts et al. [29] have further advanced this idea and incorporated it
into a maximum-a-posteriori (MAP) algorithm. This allowed them to include
additional prior knowledge about the attenuation coecients (which usually
fall into only a small number of classes) and about local smoothness. Despite
a significant effort this approach has not been adopted widely. One of the
reasons for this may be the artifacts that may arise from cross-talk between
emission and attenuation image.
In very recent work, Salomon [35, 34] and colleagues have presented an
approach that iteratively estimates the attenuation and emission images. The
approach is based on a segmentation into anatomical regions (which could for
example be derived from the MR image) and then uses PET emission data and
consistency conditions to estimate attenuation coecients for each segment.