PET data preprocessing and alternative image
reconstruction strategies
Niccolò Camarlinghi, Dipartimento di Fisica dell’università di Pisa
Contact: [email protected]
The DoPET Scanner
PET scanner for hadron therapy monitoring:• Made by 8 Modules (4 vs 4) • Each module is PMT H8500
coupled to 23x23 crystals LYSO matrix
• MLEM reconstruction• Reconstructed FOV is
100x100x100 mm3
• Voxel size 1x1x1 mm3
• 4 Million of Line Of Responses
Sketch of the DoPET scanner
The IRIS PET scanner
Pre-clinical PET scanner:• Made of two octagonal rings (16
modules) • Each module is a made of a PMT
H8500 coupled to 27x26 crystals LYSO matrix
• Reconstructed FOV is 86x86x102 mm3
• Two multi-ray System Response Matrices (SRM) available: • Small pixel: 0.42x0.42x0.855
mm3
• Big pixel: 0.855x0.855x0.855 mm3
• Offers the unique capability of performing both rotational and static acquisitions
• 23 Million of Lines Of Responses
Sketch of the IRIS PET scanner
PET Calibration/Preprocessing
Calibration :– Pixel identification– Energy calibration
Preprocessing tasks: – Estimation of random events distribution – Dead time correction– Decay correction…– LOR data for the reconstruction
PET Reconstruction software • We develop a framework for LOR based MLEM/OSEM
reconstruction • Insight ToolKit (ITK) based software (Image
processing/handling)• C++ implementation • CMake build system• Cross platform (Linux, Windows and OSX)• Based on a pre-computed System Response Matrix (SRM) • The SRMs is implemented using a Siddon based multi-ray
approach • Multi-thread implemented with Threading Build Block
(TBB) Intel • Massive use of symmetries• Component based normalization
Automatic symmetries exploitation from a pre-computed SRM
System Response Matrix representation (SRM)
• The SRM P(j,i), gives the probability that a photon pair emitted in the voxel i of the FOV is detected in the LOR j
• The SRM is the fundamental part of the MLEM/OSEM reconstruction
• In our representation each LOR of the SRM is represented as a list of entries
• The i-th entry is composed by the voxel Cartesian coordinates (x,y,z) vxi and a probability pi
Automatic exploitation of SRM symmetries
• For modern scanners, the SRM can easily exceed tents of GB
• One way of reducing the memory footprint of the SRM is by implementing symmetries
• Symmetries are an error prone task• Typically implemented manually • Our idea:– Would be good not to rely on any
scanner/geometry based assumption – Find an algorithm to extract symmetries from a
pre-computed SRM
Definition of symmetry
The algorithm considers two LORs L,M to be symmetric within a threshold T if the following three conditions are met:A. L and M intersect the same number of
voxelsB. The probabilities of the entries L,M are
the same within the percentage threshold T
C. It is possible to find a voxel space transformation to transform L into M
Implementation in practice• Loop over all the SRM LORs • Conditions A and B are easy to be verified• Condition C can be verified by restricting the
transformations to reflections and translations• Given two LORs L,M we can express translation and
reflection in voxel space:
• Translation (A=-1) : difference of coordinates is a constant
• Reflection (A=1) : sum of the coordinates is a constant
• 8 (23) Symmetry types are allowed• This relation cannot be used directly as depends on
the way L and M are arranged ( voxel ordering)
To solve this issue…
• If a symmetry exists between L,M then the following relation holds
• This gives 8 k values that can be tested directly
• If at least of component of k is not integer you can discard the k value
Long story short…
• Given two symmetric LORs L,M and (A,k) it is possible to recover L from M within the max percentage diff T by using
• This transformation can be performed before projections and retro-projections operations
• No tolerance in coordinates mismatch• Percentage tolerance T is allowed in the
recovered probability values
Results on the IRIS/PET scanner
• For this study an SRM reconstruction a FOV of 101x101x60 pixel was implemented
• The size of the SRM is 52 Gb • SRM is computed using 4x4x8 rays per
crystal, i.e. 16384 rays per LOR• Uncompressed SRM computation time 2-3
hrs • SRM compressed with
T=1%,2%,3%,4%,5%
Two slices of the same NEMA IQ phantom reconstructed with original SRM, SRM1%,SRM5%
Dividing voxel wise images obtained with SRM compressed at different threshold
T vs SRM size• The IRIS SRM was reduced by a
factor 29 using T=5%
Normalization
Stepwise normalization procedure with a planar source (IRIS PET) 1/2
In this position the coincidences involving this head are not used for normalization
• 18FDG filled planar source
• This procedure is repeated twice over 360 degrees
• View duration 10 min• Total normalization
duration is 160 min (16 views)
• Decay correction is needed
Stepwise normalization procedure with a planar source (IRIS PET) 2/2
• The same procedure is “simulated” using the SRM, i.e. by evaluating the forward projection of all the LORs onto a uniform planar phantom posed in different positions
• With this data a Defrise like [1] component based normalization is evaluated
[1] “A normalization technique for 3D PET data”,Defrise et al., Phys. Med. Biol 1991
Conclusions
• The Pisa group can take care of:– Implementing PET data preprocessing – Providing the data in the format needed by
the reconstruction – Implementing (if needed) some of the tools
contained in our Preprocessing and Reconstruction framework