1. 4-Class Multi-level Otsu Thresholding
- Produces classes with
well-discriminated intensity values
- High computational cost with
increasing classes, high
sensitivity to noise
Thresholding + Joint Mask Generation (AND, XOR)
Problem Statement
PET based photon
attenuation values
PET/CT:
GoldStandardPET/MRI
! Combined PET/MRI systems is an emerging technology that enables probing diseases in novel ways.
! However, these systems suffer from PET image artifacts due to MRI’s inability to provide accurate PET
photon attenuation values.
[1,2]
PET reconstructed with
CTAC (Gold Standard)
PET/MR systems yield artificially increased uptake
in the vicinity of the kidneys (f) and the lungs (d) ,
relative to the gold standard PET/CT fusion image
(c, e), which yields a lower PET signal.
These artifacts results from an inaccurate MRI-
based photon attenuation correction image, where
the kidney for example, appears larger than its
accurate size as obtained with CT-based
attenuation correction (b vs. a).
MRI based photon
attenuation
gradients
True emission
activity map
Proposed method
(using only
emission data)
Uniform photon
attenuation
correction
No photon
attenuation
correction
CT-based PET
reconstruction
ca e
b d f
MR-based PET
reconstruction
Current Work
MRAC CTAC
! We propose reducing the number of free variables in the GPU-based
attenuation estimation algorithm by using conventional MR images to obtain 3D
regions of similar photon attenuation.
! This approach would accelerate the PET-based photon attenuation map
determination by:
1. Reducing memory access during computation
2. Accelerating algorithm convergence
Impact Existing Methods
Using emission data only
T2 STIR
Molecular Imaging
Program at Stanford
School of Medicine
Dept of Radiology
Negin Behzadian1,4, Alexander Mihlin1,2, and Craig S. Levin1,2,3,4
Departments of 1Electrical Engineering, 2Radiology, 3Physics, and 4Bioengineering, Stanford University, California, USA
MR-guided Segmentation-based Acceleration of Emission-based
Attenuation Correction in Whole-body Time-of-Flight PET/MRI
Acknowledgments
I thank the Stanford BioE REU Program, my mentor, Alex Mihlin, and my faculty mentor, Dr. Craig S. Levin for making this project possible.
References
[1] D. E. Opera-Lager et. al., Mol. Imaging Biol. 1536, 1-12 (2015).
[2] A. Mehranian et. al., Journal of Nuclear Medicine. 2159, 1-22 (2015).
[3] P. Khateri et. al., Mol. Imaging Biol. 1536, 1-9 (2015)
Conclusion
Background
! Improve PET imaging accuracy in hybrid Positron Emission Tomography (PET)/Magnetic Resonance
Imaging (MRI) systems
Future Steps
Considerations
! As opposed to the MR-based PET attenuation correction,
incomplete segmentation (left) however, would not cause
an error in attenuation correction since attenuation values
are determined exclusively from PET. Individual
attenuation values would be estimated for un-segmented
regions accordingly.
! Mis-segmentation is problematic when a segmented region
includes voxels with different photon attenuation
characteristics, such as bone and soft tissue in an MR
attenuation correction map (right).
! Current segmentation-based MR attenuation correction methods derive standard 4-class attenuation
maps (lung, air, soft tissue, fat) using a combination of thresholding, region growing, and arithmetic
operations on 4 Dixon MR images. A bone attenuation class is separately generated using either CT
information [2] or advanced, region-specific MR sequences such as STE [3].
! The current project works to devise a general automatic method for organ-specific segmentation of
regions of similar attenuation, exclusively using conventional MR images. Each segmented region is
treated as a single degree of freedom in the PET-based determination of photon attenuation map.
WATER FAT
Co-registered
CT Bone Map
5-class
map
Results
Gradient Anisotropic Diffusion (T2 STIR, Dixon WATER)
Non-linear, edge-enhancing Gaussian smoothing
4-Class Fuzzy c-Means Clustering
- Each voxel may belong to more than one cluster, with
degree of membership determined by Euclidean
distance to cluster centroid
- Few iterations required, but difficult to optimize
parameters
- Fuzzy c-Means and Otsu generate similar segmentation of high-intensity regions, but significant
differences in mid-intensity regions (ie. liver, bone)
- Joint masks of clusters allow selection for region-recombinant (vs. global) histogram shapes compatible
with automatic thresholding segmentation algorithms.
Histogram-based Automatic Thresholding
- Single peak located at an extrema of image range "
Triangle thresholding
Triangle Method
OVERLAID SEGMENTATION RESULTS
Morphological
post-processing
LIVER BONE
1. Obtain segmentations for as many anatomical “attenuation” regions as possible, and accordingly derive
an un-segmented attenuation class for voxel-by-voxel estimation by proposed PET-based algorithm
2. Use segmentation results to constrain the PET-based algorithm, and assess the clinical viability and
accuracy of resulting PET-derived photon attenuation map, using both phantom and clinical data
UsingT2 image ONLY UsingT2 + WATER
! We propose a generalized automatic segmentation method for deriving 3D anatomical regions of similar
attenuation (including bone), using conventional MR images (T2, Dixon sequence) only.
! These segmentations will be used to clinically enable an accurate GPU-based PET-only joint estimation
algorithm for emission activity distribution and photon attenuation map.
[1]
[2]
! In order to improve the accuracy of PET/MRI systems, the Molecular Imaging Instrumentation
Laboratory at Stanford has developed an accurate PET-only (as opposed to MRI-based) joint estimation
method for emission activity distribution and photon attenuation map.
! The current project aims at making this method clinically viable, by accelerating it using MRI information
already available in PET/MRI systems.