Biomedical Engineering Reference
In-Depth Information
Nevertheless, segmentation-based AC methods are robust, computationally fast and
the associated quantification error for bone lesions is repeatable and well understood.
This has lead to the use of such AC approach in commercial PET/MRI hybrid scan-
ners: the Siemens Biograph mMR uses an adaptation of the algorithm implemented
by Martinez-Möller et al [
46
], and the Philips Ingenuity TF PET/MR [
55
]usesan
adaptation of the algorithm proposed by Schulz et al. [
49
,
51
].
An alternative method for MRI-guided AC is based on the registration or deformation
of an attenuation atlas template to patients' MR images in order to obtain an atten-
uation map adapted for the specific patient. Machine learning techniques have been
used in order to implement such methods and in particular to learn how to convert
MR tissue proton density data to CT tissue density Hounsfield unit (HU) values from
alignedMR and CT attenuation images. With the advent of hybrid PET/MRI devices,
several groups have proposed
template-based AC
approaches for PET images correc-
tion [
43
,
56
]. The advantage of such methods is that generally they do not need any
additional MR bone imaging in order to predict bone tissue, with a drawback being
a time-consuming map calculation. The sensitivity of these methods to anatomical
variability has still to be evaluated [
52
]. In fact, they require an accurate non-rigid
registration of the template to the acquired MR image. This is feasible at the level of
brain imaging, where template-based methods have shown great potential, but they
have been reported to be error-prone when applied for a whole-body analysis [
57
].
Therefore, Hofmann et al. [
47
] proposed to complement a template-basedACmethod
with the information available from the segmented MR and pattern recognition, and
the feasibility of such method was demonstrated also when applied to whole-body
images. Detailed reviews of MR-guided attenuation techniques have been published,
e.g. by Bezrukov et al. [
52
], Wagenknecht et al. [
58
] Martinez-Möller et al. [
51
].
Non-uniformities in the efficiencies of the individual detectors, due to variations
in detector geometry or electronics, may cause variations in coincidence detection
efficiency between different LORs. The
normalization correction
compensates for
these non-uniformities. To compute a normalization correction, detector efficien-
cies, geometrical efficiency variations and variations in plane-to-plane efficiency are
measured in order to account for the relative variation in coincidence detection effi-
ciencies between system LORs (e.g. by collecting data from a uniform plane source
of activity, positioned at 6-8 equally spaced projection angles). The normalization
correction factors can also be computed using the component-based method. In this
case, the normalization correction factors [
46
]are:
1
ʵ
i
×
ʵ
j
×
n
i
,
j
=
g
i
,
j
where, the coincidence detection efficiency of a pair of detectors
i
and
j
is assumed
to be composed of the product of the detector efficiencies,
ˉ
j
, and geometrical
factors,
g
i
,
j
, that include corrections for the angle of incidence of the annihilation
photons, systematic variation in crystal efficiency dependent on the position of the
crystal in a detector block module, and relative plane efficiency.
ˉ
i
and
