Transcript Spatio-Temporal Free-Form Registration of Cardiac MR Image

Spatio-Temporal Free-Form
Registration of Cardiac MR Image
Sequences
Antonios Perperidis
s0094336
10/02/2006
MR-Imaging
 Magnetic Resonance (MR) imaging allows
the acquisition of:
 3D images which describe the cardiac anatomy.
 4D cardiac image sequences which describe the
cardiac anatomy and function.
 Advances in cardiac MR imaging are making
it an important clinical tool:
 The improvement of the spatial and temporal
resolution of the image sequences enabling the
imaging of small cardiac structures.
 The development of tagged MR imaging which
allow the study of cardiac motion.
Cardiac Image Registration



Currently increased need for cardiac registration methods.
Cardiac image registration is a very complex problem due
to 4D nature of the cardiac data:


Complicated non-rigid motion of the heart and the thorax.
Low resolution with which cardiac images are usually acquired.

The construction of anatomical and functional atlases of the
heart.
The analysis of the myocardial motion.
The segmentation of cardiac images.
The fusion of information from different modalities such as CT,
MR, PET, and SPECT.
The comparison of images of the same subject.
The comparisons between the cardiac anatomy and function of
different subjects.
Recently, cardiac image registration has emerged as an
important tool for a large number of applications such as:





Background
 There are currently many registration techniques
for cardiac imaging.
 Most techniques focus on 3D images ignoring any
temporal misalignment between two image
sequences.
 One exception: an approach for the spatial and
temporal registration of cardiac SPECT and MR
images.
 Uses linear interpolation for the temporal mapping
between the end-systolic and end-diastolic frames.
 The heart has a complicated motion pattern.
 This method ignores all the spatial information
contained in the images between the end-systolic
and end-diastolic frames.
Spatio-Temporal Registration
Introduction
 The heart is undergoing a spatially and
temporally varying degree motion during
the cardiac cycle.
 Spatial alignment of corresponding frames
of the image sequences not enough since
frames may not correspond to the same
position in the cardiac cycle of the hearts.
 This is due to differences in:
 The acquisition parameters.
 The length of cardiac cycles.
 The dynamic properties of the hearts .
Spatio-Temporal Registration
Introduction
 Spatio-temporal alignment enables to find the temporal
relationship between the 2 image sequences.
 We present 2 spatio-temporal alignment methods using
image information only.
 The 4-D mapping can be described by the
transformation:
 Mapping can be resolved into decoupled spatial and
temporal components:


Spatial:
Temporal:
Spatial Alignment
 Aim: to relate each spatial point of an
image to a point of the reference image.
 Tspatial can be written:
 Tspatial/global: 3D affine transformation:
 Coefficients  parameterize the 12 degrees
of freedom.
 Tspatial/local: Free-Form deformation (FFD)
model based on B-Splines.
Temporal Alignment
 Ttemporal can be written:
 Ttemporal/global: affine transformation:
 a accounts for scaling differences.
 b accounts for translation differences.
 Ttemporal/local: Free-Form deformation
(FFD) using 1-D B-Splines.
Combined Optimization of the
Spatial and Temporal Components

2 registration algorithms for finding the optimal
transformation T:
1.
2.



Optimizes the spatial and temporal transformation
components simultaneously using image information only
Optimizes the temporal transformation component before
optimizing the spatial component.
In the 1st algorithm: Optimal transformation T is found
by maximizing a voxel based similarity measure.
Normalized Mutual Information (NMI): a measure of
spatio-temporal alignment.
NMI is optimized as a function of Tspatial/global and
Ttemporal/global using:


an iterative downhill descent algorithm.
a simple iterative gradient descent method.
Separate Optimization of the Spatial and
Temporal Components 1





Computational complexity of the
previous method is very high.
Reduced by optimizing each
transformation component
separately.
Ttemporal/global: aligns the temporal
ends of the image sequences
Ttemporal/local: aligns a limited
temporal positions of the cardiac
cycles.
Temporal positions detected by
calculating the normalized
cross-correlation coefficient
between each frame of the
sequence with the first frame.
Separate Optimization of the Spatial and
Temporal Components 2
 The idea behind this
approach:
 during the contraction
phase of the cardiac
cycle each
consecutive image
will be less similar to
the first image.
 during the relaxation
phase of the cardiac
cycle each
consecutive image
will be more similar
to the first image
Results
 Algorithm evaluation: 15 cardiac MR images from
healthy volunteers.
 Reference subject: 32 different time frames
acquired.
 14 4D cardiac MR images were registered to the
reference subject.
 Cardiac Cycle length: 300 to 800msec.
 Qualitative evaluation through visual inspection.
 Quality of registration in spatial domain measured
by Calculating volume overlap for:
 The left and right ventricles.
 The myocardium.
Results – Separate optimization of
the transformation components
 Maximum contraction & and-diastole positions
determined manually.
 Positions compared with corresponding positions
identified by the algorithm:


Mean error in the detection of maximum contraction: 1.2
frames.
Mean error of the end diastole detection: 0.93 frames.
Combined optimization of the transformation
component - Qualitative Evaluation
 Deformable temporal &
spatial registration
improves alignment of
the image sequences in
spatial and temporal
domain
Combined optimization of the transformation
component - Quantitative Evaluation

Optimising
transformation
components
simultaneously
provides
better overlap
measures the
separate
optimisation
method.
Results – Using cross correlation based
method to calculate temporal alignment.


Approach
achieves very
good spatio
temporal
registration of
the images.
Computational
complexity is
reduced by 25%
(approximately)
Applications of the Spatio-Temporal
Registration Method.
 Large number of applications for this SpatioTemporal registration method:
 Comparison between image sequences from the same
subject.
 Comparison between image sequences from different
subjects,
 Building probabilistic
and statistical atlases
of the cardiac
anatomy and
function.
Questions?