Clinical Recommendations for Multimodality Cardiovascular ...
Interpolation Artifacts in Multimodality Image...
Transcript of Interpolation Artifacts in Multimodality Image...
Interpolation Artifacts in Multimodality Image Registration
IEEE Transaction on Medical Imaging, 2003
Jeffrey Tsao (editor)
Keumsil Lee Feb. 9, 2004
Introduction
1. Multimodality image registration? - Images are acquired in different poses- Complementary information: anatomy and physiology
2. Wood’s method VS MI- Voxel intensities at corresponding positions conform to a many-to-one
mapping- No intensity mapping relationship between the images
3. Prefers match metric with fewer assumptions4. Four steps of image registration
- Tentative pose relative to one another- Joint histogram of voxel intensities- Determine match metric- Iterative optimizing of match metric
Mutual Information
v applied to measure the statistic dependence between voxels intensities of the images
v assumed to be maximal if the images are geometrically aligned
Goal of Study
1. Patterns of interpolation artifacts2. Strategies of overcome these artifacts.3. Effects of eight interpolation schemes
- Nearest neighbor (NN)- Linear- Cubic Catmull-Rom- Hamming-windowed sinc- Partial volume- NN with jittered sampling (JIT)- NN with histogram blurring (BLUR)- NN with JIT and BLUR
4. Impacts of using different numbers of intensity bins5. Artifact-reducing effects of image rotation and sampling
Methods
1. Human brain images
2. 4 pairs of images
3. Image degradation
128X128 MRI 128X128 SPECT
Ł
Methods - continue
4. Registration curves- Translational misregistration : 10 vixels
- Rotational misregistration: 180
5. Translation & Rotation applied to SPECT image- Less signal energy
- High spatial frequencies
- Less susceptible to degradation in image quality from interpolation
6. MI was the maximum when the images were registered
Evaluation interpolation artifacts: MI
1. Steps - Images were overlaid
- Joint histogram
- Quantized voxel intensity with fixed number of discrete bins
- Different numbers of intensity bins
2. Determination of MI:
Evaluation interpolation artifacts: Smoothness of Registration Curve
1. Steps- Normalize the Y axis of each registration curve
- Subtract the curve from a smoothed version
- The difference of them represents the interpolation artifacts
2. Filters- Median filtering & Convolution with a Hamming filter
- 1.5 voxels in translation & 5 in rotation along the X axis
3. Fewer interpolation artifacts à smoother à higher smoothness value
Translational Misregistration: 10 Voxels
Rotational Misregistration: 180
Smoothness
Ł # of bins
Discussion
v Number of Intensity Bins
v Classification of Interpolator (Intensity interpolator VS non-)
v Intensity interpolator
v Match metric and Generalization
Summary
v Interpolation artifacts severely affect registration
v Artifacts can be reduced at misaligned voxel gridsà NN interpolation & BLUR can be considered to save time
v Well aligned voxel grids need special attention
v Suggesting strategies1) Avoid extremely large or small number of intensity bins2) Resampling: rotated orientation & unequal voxel sizes3) JIT4) BLUR
By initial transformation estimate
v
v
Example
��������������� � ������ � �������������������
� ��� ��� �����
����� ���������� ����
��������������� ��������
������� ��� ���������
� ����� ������������
���� ��� ��������
Overview of SPM Analysis
Methods
1. Human head images- 128 x 128 axial MR (Magnetic Resonance) image
- 128 x 128 SPECT (single photon emission computed tomography) image
2. 4 pairs of images1. aligned : preregistered using Wood’s Method
2. rotated : SPECT image was 30 counter-clockwise rotated
3. 128/129 : SPECT image was rescaled 129x129
4. /3 : SPECT image was rescaled by increasing the voxel size by /3
3. Image degradation