Date post: | 20-Jan-2016 |
Category: |
Documents |
Upload: | roland-holt |
View: | 214 times |
Download: | 0 times |
A Novel Image Registration Pipeline for 3-A Novel Image Registration Pipeline for 3-D Reconstruction from Microscopy ImagesD Reconstruction from Microscopy Images
Kun Huang, PhD; Ashish Sharma, PhD; Lee Cooper, MS;Kun Huang, PhD; Ashish Sharma, PhD; Lee Cooper, MS;Tony Pan, MS; Metin Gurcan, PhD; Joel Saltz, MD, PhDTony Pan, MS; Metin Gurcan, PhD; Joel Saltz, MD, PhD
Department of Biomedical InformaticsOhio State University
Creating Geometry from Images
Placenta
H+E Slides Alignment
SegmentationVisualization/Surface
Extraction
AperioScanner
Registration
Registration between different modalities (e.g, MRI and PET)
Mapping of different samples to the same reference (e.g., brain mapping)
3-D reconstruction
An optimization problem Initialization Point feature
matching Automatic vs.
manual
Issues with automatic registration
Initialization Landmark- or image- based? Linear or nonlinear? Error metric / Meaningful morphology /
Domain specific knowledge Computation Structural constraints
Fast initialization using landmarks
S1
Fast initialization
S2
S2’
S1’
S3’Matching pairs:
(S1, S1’) (S1, S3’) (S2, S2’)
S1
S2 S2’
S1’
S3’
d12
d12’d13’
Fast initialization
Maximum Clique Maximum Cyclic Structure
(S1, S1’, S2, S2’, θ, T)
Fast initialization
Difference between two images.
Difference after the automatic initialization using region features.
Difference after the MMI algorithm.
Fast initialization
Registration of Large Images Using Landmarks
Registration of Large Images Using High-Level Features
• No need to globally transform the image• Multi-level registration – rigid to nonrigid• Parallelizable – local operations
Registration of Large Images Using Landmarks
Registration of Large Images Using High-Level Features
• Point feature does NOT contain global information• For global transformation (e.g., rotation and
translation), we need “global” features such as high-level features.
• For nonlinear transformation, which is local, we need “local” features such as point features.
Global first, local second.
Stacks of microscopy images
Principal component analysis (PCA) – based rigid registration
Stacks of microscopy images
Stacks of microscopy images
Stacks of microscopy images
Stacks of microscopy images
Stacks of microscopy images
Stacks of microscopy images
3-D reconstruction vs. registration
• The current metric for registration is between two images and is just for the sake of perfect “registration”.
• We do “registration” for the sake of 3-D reconstruction.
• The structural constraint should be incorporated in the “cost function” instead of just used as a post processing or validation criterion.
New multiple image registration algorithm is needed!
3-D reconstruction via registration
Feature extraction
Feature matching/ tracking
Trajectory generation
Trajectory smoothing and
adjustment
New location for the features in every image
Nonlinear transformation
for every image
Collective adjustment of trajectories
3-D reconstruction via registration
Tracked trajectory
Smoothed trajectory
Registration moves the landmarks to the new locations.
3-D reconstruction via registration
3-D reconstruction via registration
3-D reconstruction via Registration
Summary, future work and discussion
• Technical issues related to automatic registration.• Two step approach to achieve “good” nonlinear
registration.• The paradigm for 3-D reconstruction is different with
pure registration.• New registration pipeline is proposed and implemented.
Summary, future work and discussion
• Parallelization – especially in nonlinear transformation stage.
• Multiresolution / hierarchical approach.
Acknowledgement
• BMI• Imaging group• Collaborators
Thank you !