szirmay-kalos, lászló budapest uni of tech

20
GPU-based Image Processing Methods in Higher Dimensions and their Application to Tomography Reconstruction Szirmay-Kalos, László Budapest Uni of Tech Sapporo, 2010

Upload: fawzia

Post on 14-Jan-2016

33 views

Category:

Documents


0 download

DESCRIPTION

GPU-based Image Processing Methods in Higher Dimensions and their Application to Tomography Reconstruction  . Szirmay-Kalos, László Budapest Uni of Tech. Sapporo, 2010. Positron Emission Tomography. Intensity: x. e -. e +. Line Of Response : y. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Szirmay-Kalos, László Budapest Uni of Tech

GPU-based Image Processing Methods in Higher Dimensions and their Application

to Tomography Reconstruction  

Szirmay-Kalos, LászlóBudapest Uni of Tech

Sapporo, 2010

Page 2: Szirmay-Kalos, László Budapest Uni of Tech

Positron Emission Tomography

e-

e+

Line Of Response: y

Intensity: x

Page 3: Szirmay-Kalos, László Budapest Uni of Tech

Iterative Maximum Likelihood Reconstruction

Measureddetectorresponse

Source intensity as a 3D voxel array

Source estimation

Source correction

Compute expecteddetector response

Expecteddetector response

Page 4: Szirmay-Kalos, László Budapest Uni of Tech

Ill-posed reconstructionerror

Iteration number

Maximum likelihood estimate

Page 5: Szirmay-Kalos, László Budapest Uni of Tech

Regularization

• Additional information– Penalty term added to the

likelihood• Prevents overfitting• TV norm (L1 optimization)

– No smoothness condition– Preserves edges

x

dttf )('

V

dvvx )(

Page 6: Szirmay-Kalos, László Budapest Uni of Tech

TV minimalization

• In steepest descent search the derivative of the TV term is needed:– Function |x| cannot be differentiated:

• Add a small term (blurring)• Primal-dual methods

– Only local values are needed: parallelization

V

dvvx )(

xV

Page 7: Szirmay-Kalos, László Budapest Uni of Tech

Detector scattering compensation

Path probability inside the detector can be pre-computed or measured

photon

crystals

intercrystalscattering

absorption

Electronicsnumber of hits

Page 8: Szirmay-Kalos, László Budapest Uni of Tech

Pre-computation

q

Page 9: Szirmay-Kalos, László Budapest Uni of Tech

dxxwxXL )()( 1

0

))(( dttxXL

L

w

L

=

Quasi-Monte Carlo filtering

Page 10: Szirmay-Kalos, László Budapest Uni of Tech

Random sampling

Random sampling

undersampling

oversampling

Page 11: Szirmay-Kalos, László Budapest Uni of Tech

Delta-Sigma modulator

Filter kernel

pixels

Page 12: Szirmay-Kalos, László Budapest Uni of Tech

Filter kernel

Delta-Sigma modulator

Page 13: Szirmay-Kalos, László Budapest Uni of Tech

Delta-Sigma modulator

Filter kernel

Floyd-Steinberg halftoning!

Page 14: Szirmay-Kalos, László Budapest Uni of Tech

Sampling with Sigma-Delta modulation

Page 15: Szirmay-Kalos, László Budapest Uni of Tech

GPU Implementation• Simulation step:

• GPU: Quasi-SIMD massively parallel machine– Gathering = threads to equations (outputs)– “No” conditional statements or variable length loops

• Reconstruction algorithm– Geometric LOR marching: threads to LORs

(adjoint problem)– LOR filtering: threads to output LORs– TV regularization: threads to voxels

xAy high dim. integrals

108 voxels108 LORs

Page 16: Szirmay-Kalos, László Budapest Uni of Tech

TV regularization results

=0.005

=0.05=0.008

No TV

Page 17: Szirmay-Kalos, László Budapest Uni of Tech

TV results

=0.001 =0.0005 =0.0001=0.005

Page 18: Szirmay-Kalos, László Budapest Uni of Tech

Scattering in the detector

3D reconstruction, no detector scattering

compensation

Detector scattering compensation

2D reconstruction:SSRB + OSEM

Page 19: Szirmay-Kalos, László Budapest Uni of Tech

F18 mouse

Page 20: Szirmay-Kalos, László Budapest Uni of Tech

Conclusions

• Image processing algorithms can be and are worth being generalized to higher dimensions, but

• beware the curse of dimensions and use Monte Carlo methods.

• GPUs are good platforms for image processing, but adopt the gathering view and refrain from conditionals.