Fast Intra- and Intermodal Deformable Registration Based on Local Subvolume Matching - PowerPoint PPT Presentation

Title: Fast Intra- and Intermodal Deformable Registration Based on Local Subvolume Matching


1
Fast Intra- and Intermodal Deformable
Registration Based on Local Subvolume Matching

• Matthias Söhn1, Verena Scheel2, Markus Alber1
• (1) Radiooncological Clinic, Section for
  Biomedical Physics, University of Tübingen,
  Germany
• (2) Laboratory for Preclinical Imaging and
  Imaging Technology, Department of Radiology,
  University of Tübingen, Germany

Forschungszentrum für Hochpräzisionsbetrahlung
2
Deformable Registration for Radiotherapy

• Requirements Challenges
• accuracy
• fast
• no or little user interaction
• versatility

3
Algorithmic Implementation
Cover region of interest in reference image with
regular 3D-grid of featurelets
typical size 1.5x1.5x1.5 cm
4

• Individual rigid registration of each featurelet

reference image (exhale)
target image (inhale)
for each featurelet
maximization of local normalized mutual
information (NMI) allowing 3D-shifts within local
search region
? fast parallelizable!
5

• Individual rigid registration of each featurelet

reference image (exhale)
target image (inhale)
for each featurelet
maximization of local normalized mutual
information (NMI) allowing 3D-shifts within local
search region
? fast parallelizable!
6

• Automatic assessment of local registration quality

7

• Automatic assessment of local registration
  quality -- Result

8

• Relaxation Step Iterative Minimization of
  Deformation Energy for mismatched Featurelets

9

• B-Spline Interpolation of Featurelet shift vectors

target image, final featurelet positions
10
Results

• RCCT Inhale-Exhale deformable registration
  Visual evaluation

before
after registration
11
Results

• CT-ConeBeamCT deformable registration Visual
  evaluation

before
after registration
Elekta XVI ConeBeam-CT data, courtesy D. Yan, Y.
Chi (Beaumont)
12
Results

• CT-MRI deformable registration Visual evaluation

CT
MRI
MRI (backtransformed)
before
after registration
13
Results

• Quantitative evaluation Anatomical landmarks

N55 landmarks altogether
N 3D-residuals mm 3D-residuals mm
N before registration after registration
pat. 1 15 8.24.6 1.30.9
pat. 2 11 4.21.5 1.51.0
pat. 3 14 10.45.5 1.80.7
pat. 4 15 7.85.7 1.81.3
avg. avg. 7.85.1 (max. 21.3) 1.61.0 (max. 4.6)
marked in inhale and exhale CTs of 4 patients
Siemens Somatom Sensation Open RCCT datasets _at_
1x1x3mm voxelsize
14
Results

• Quantitative evaluation Virtual phantom

Virtual thorax phantom known deformation field
used to to deform real lung CT dataset
courtesy D. Yan, Y. Chi (Beaumont Hospital)
15
Results

• Computational performance

test case registered region voxels calculation time (dual-core Xeon PC, 2x2.66GHz)
Thorax (CT-CT) 360x270x120 2min 12sec
HN (CT-CBCT) 225x225x115 49sec
HN (CT-MRI) 378x210x70 1min 23sec

• calculation time mainly depends on
• size of registered region
• size of local search region
• featurelet size

16
Results

• Computational performance

test case registered region voxels calculation time (dual-quadcore Xeon PC, 8x2.66GHz)
Thorax (CT-CT) 360x270x120 38sec
HN (CT-CBCT) 225x225x115 14sec
HN (CT-MRI) 378x210x70 19sec
online deformable registration!
17
Conclusions

• Featurelet-based deformable registration

• fast, parallizable
• model-independent, fully automatic
• enables multi-modality registration due to use
  of mutual information
• sub-voxel registration accuracyas shown by
  landmark-based evaluation and virtual thorax
  phantom

online multi-modality deformable registration
within reach!