Title: Integrative System Approaches to Medical Imaging and Image Computing Physiological Modeling In Situ Observation Robust Integration
1Integrative System Approaches to Medical Imaging
and Image ComputingPhysiological Modeling In
Situ ObservationRobust Integration
2Motivations
- Observing in situ living systems across temporal
and spatial scales, analyzing and understanding
the related structural and functional segregation
and integration mechanisms through model-based
strategies and data fusion, recognizing and
classifying pathological extents and degrees - Biomedical imaging
- Biomedical image computing and intervention
- Biological and physiological modeling
3Perspectives
- Recent biological technological breakthroughs,
such as genomics and medical imaging, have made
it possible to make objective and quantitative
observations across temporal and spatial scales
on population and on individuals - At the population level, such rich information
facilitates the development of a hierarchy of
computational models dealing with (normal and
pathological) biophysics at various scales but
all linked so that parameters in one model are
the inputs/outputs of models at a different
spatial or temporal scale - At the individual level, the challenge is to
integrate complementary observation data,
together with the computational modeling tailored
to the anatomy, physiology and genetics of that
individual, for diagnosis or treatment of that
individual
4Perspectives
- In order to quantitatively understand specific
human pathologies in terms of the altered model
structures and/or parameters from normal
physiology, the data-driven information recovery
tasks must be properly addressed within the
content of physiological plausibility and
computational feasibility (for such inverse
problems)
5Philosophy
- Integrative system approaches to biomedical
imaging and image computing - System modeling of the biological/physiological
phenomena and/or imaging processes physical
appropriateness, computational feasibility, and
model uncertainties - Observations on the phenomena imaging and other
medical data, typically corrupted by noises of
various types and levels - Robust integration of the models and
measurements patient-specific model structure
and/or parameter identification, optimal
estimation of measurements - Validation accuracy, robustness, efficiency,
clinical relevance
6Current Research Topics
- Biomedical imaging
- PET activity and parametric reconstruction
- low-count and dynamic PET
- pharmacokinetics
- SPECT activity and attenuation reconstruction
- Medical image computing
- Computational cardiac information recovery
- electrical propagation, electro-mechanical
coupling, material elasticity, kinematics,
geometry - fMRI analysis and applications
- biophysical model based analysis
- Fundamental medical image analysis problems
- Efficient representation and computation platform
- Robust image segmentation
- Level set on point cloud
- Local weakform active contour
- Inverse-consistent image registration
7Tracer Kinetics Guided Dynamic PET Reconstruction
- Shan Tong, Huafeng Liu, Pengcheng Shi
- Department of Electronic and Computer Engineering
- Hong Kong University of Science and Technology
8Outline
- Background and review
- Introduce tracer kinetics into reconstruction, to
incorporate information of physiological
processes - Tracer kinetics modeling and imaging model for
dynamic PET - State-space formulation of dynamic PET
reconstruction problem - Sampled-data H8 filtering for reconstruction
- Experiments
9Background
- Dynamic PET imaging
- Measures the spatiotemporal distribution of
metabolically active compounds in living tissue - A sinogram sequence from contiguous acquisitions
- Two types of reconstruction problems
- Activity reconstruction estimate the spatial
distribution of radioactivity over time - Parametric reconstruction estimate physiological
parameters that indicate functional state of the
imaged tissue
Parametric image of rat brain phantom
Activity image of human brain
10Dynamic PET Reconstruction Review on existing
methods
- Frame-by-frame reconstruction
- Reconstruct a sequence of activity images
independently at each measurement time - Analytical (FBP) and statistical (ML-EM,OSEM)
methods from static reconstruction - Suffer from low SNR (sacrificed for temporal
resolution) and lack of temporal information of
data - Statistical methods assume data distribution that
may not be valid (Poisson or Shifted Poisson) - Prior knowledge to constrain the problem
- Spatial priors smoothness constrain, shape prior
- Temporal priors But information of the
physiological process is not taken into account
11Introduce Tracer Kinetics into Reconstruction
- Motivation
- Incorporate knowledge of physiological modeling
- Go beyond limits imposed by statistical quality
of data - Tracer kinetic modeling
- Kinetics spatial and temporal distributions of a
substance in a biological system - Provide quantitative description of physiological
processes that generate the PET measurements - Used as physiology-based priors
12Tracer Kinetics Guided Dynamic PET Reconstruction
Overview
Represented by PET data
Described by tracer kinetic models
- Tracer kinetics as continuous state equation
- Sinogram sequence in discrete measurement equation
13Tracer Kinetics Guided Dynamic PET Reconstruction
Overview
- Main contributions
- Physiological information included
- Temporal information of data is explored
- No assumptions on system and data statistics,
robust reconstruction - General framework for incorporating prior
knowledge to guide reconstruction
14Two-Tissue Compartment Modeling for PET Tracer
Kinetics
- Compartment a form of tracer that behaves in a
kinetically equivalent manner. Interconnection
fluxes of material and biochemical conversions - arterial concentration of nonmetabolized
tracer in plasma - concentration of nonmetabolized tracer in
tissue - concentration of isotope-labeled metabolic
products in tissue - first-order rate constants
specifying the tracer exchange rates
15Two-Tissue Compartment Modeling for PET Tracer
Kinetics
- Governing kinetic equation for each voxel i
- Compact notation
(1)
(2)
16Two-Tissue Compartment Modeling for PET Tracer
Kinetics
- Total radioactivity concentration in tissue
- Directly generate PET measurements via positron
emission - Neglect contribution of blood to PET activity
(3)
Typical time activity curves
17Imaging Model for Dynamic PET Data
- Measure the accumulation of total concentration
of radioactivity on the scanning time interval - Activity image of kth scan
- AC-corrected measurements
- Imaging matrix D contain probabilities of
detecting an emission from one voxel at a
particular detector pair - Complicated data statistics due to SC events,
scanner sensitivity and dead time, violating
assumptions in statistical reconstruction
(4)
(5)
18State-Space Formulation for Dynamic PET
Reconstruction
- Time integration of Eq.(2)
- where ,
- System kinetic equation for all voxels
- where ,
system noise - A block diagonal with blocks ,
- Activity image expressed as
- Let , construct measurement equation
(6)
(7)
(8)
(9)
19State-Space Formulation for Dynamic PET
Reconstruction
- Standard state-space representation
- Continuous tracer kinetics in Eq.(7)
- Discrete measurements in Eq.(9)
- State estimation problem in a hybrid paradigm
- Estimate given , and obtain activity
reconstruction using Eq.(8)
(7)
(9)
(8)
20Sampled-Data H8 Filtering for Dynamic PET
Reconstruction
- Mini-max H8 criterion
- Requires no prior knowledge of noise statistics
- Suited for the complicated statistics of PET data
- Robust reconstruction
- Sampled-data filtering for the hybrid paradigm of
Eq.(7)(9) - Continuous kinetics, discrete measurements
- Sampled-data filter to solve incompatibility of
system and measurements
21Mini-max H8 Criterion
- Performance measure (relative estimation error)
- , S(t), Q(t), V(t), Po
weightings - Given noise attenuation level , the optimal
estimate should satisfy - Supremum taken over all possible disturbances and
initial states - Minimize the estimation error under the worst
possible disturbances - Guarantee bounded estimation error over all
disturbances of finite energy, regardless of
noise statistics
(10)
22Sampled-Data H8 Filter
- Prediction stage
- Predict state and on time interval
with and as
initial conditions - Eq.(13) is Riccati differential equation
- Update stage
- At , the new measurement is used to update the
estimate with filter gain
(12)
(13)
(14)
(15)
23System Complexity Numerical Issues
- Large degree of freedom
- In PET reconstruction with N voxels (128128)
- Numerical Issues
- Stability issues may arise in the Riccati
differential equation (13), Mobius schemes have
been adopted to pass through the singularities
Number of elements in
J. Schiff and S. Shnider, A natural approach to
the numerical integration of Riccati differential
equations, SIAM Journal on Numerical Analysis,
vol. 36(5), pp. 13921413, 1996.
24Experiments Setup
Time activity curves for different tissue
regions in Zubal phantom
Zubal thorax phantom
Kinetic parameters for different tissue regions
in Zubal thorax phantom
25Experiments Setup
Activity image sequence
Sinogram sequence
- Total scan 60min, 18 frames with 4 0.5min, 4
2min, and 10 5min - Input function
- Project activity images to a sinogram sequence,
simulate AC-corrected data with imaging matrix
modeled by Fesslers toolbox
Prof. Jeff Fessler, University of Michigan
26Experiments Setup
- Different data sets
- Different noise levels 30 and 50 AC events of
the total counts per scan - High and low count cases 107 and 105 counts for
the entire sinogram sequence - Kinetic parameters unknown a priori for a
specific subject, may have model mismatch - Perfect model recovery same parameters in data
generation and recovery - Disturbed model recovery 10 parameter
disturbance added in data generation - H8 filter and ML-EM reconstruction
27Experiments Results
Perfect model recovery under 30 noise
Frame 4
Frame 8
Frame 12
Truth
ML-EM
H8
ML-EM
H8
Low count
High count
28Experiments Results
Disturbed model recovery for low counts data
Frame 4
Frame 8
Frame 12
Truth
ML-EM
H8
ML-EM
H8
30 noise
50 noise
29Experiments Results
- Quantitative results for different data sets
Quantitative analysis of estimated activity
images, with each cell representing the
estimation error in terms of bias variance.
30Future Work
- Current efforts
- Monte Carlo simulations
- Real data experiments
- Planned future work
- Reduce filtering complexity
- Parametric reconstruction using system ID/joint
estimation strategies