andrás horváth

12
András Horváth Segmentation of 3D ultrasound images of the heart Diagnostic ultrasound imaging

Upload: cerise

Post on 23-Jan-2016

49 views

Category:

Documents


0 download

DESCRIPTION

Segmentation of 3D ultrasound images of the heart. Diagnostic ultrasound imaging. András Horváth. Image segmentation. Determine sets of pixels according tocertain visual characteristics Simplify the representation of an image Separate ‘important’ parts. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: András Horváth

András Horváth

Segmentation of 3D ultrasound images of the heart

Diagnostic ultrasound imaging

Page 2: András Horváth

Image segmentation

Determine sets of pixels according tocertain visual characteristics

Simplify the representation of an image

Separate ‘important’ parts

http://www.cs.toronto.edu/~jepson/csc2503/segmentation.pdf

http://www.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/

Page 3: András Horváth

3D cardiac ultrasound

Image analysis in cardiac images:• Segmentation • Motion Analysis.• Strain Analysis.

Extra spatial information vsComputational complexity (curse of dimensionality)

Artifacts:1. Contrast inhomogeneity and data drop-out.2. Spurious signal from side lobes.

‘Extra’ goals:• Machine/operator independence• Handle dropout and inhomogeneity• Handle boundary shape • Non-local methods• Automatic segmentation

www.haskins.yale.edu/conferences/UltrafestV/abstracts/tagare_poster.pdf

Page 4: András Horváth

1980-: Edge Detectors, connected boundaries, local algorithms

1990-: “Snake” + Template matching algorithms Mildly-disconnected boundaries, local algorithms.

1990-00:Probabilistic algorithms (Bayesian) region-based, non-local algorithms.

Discriminative machine-learning algorithms. Spatial inhomogeneity, curse of dimensionality, “drag-and-drop” segmentation.

Some history

G. Stetten, eal-Time 3D Ultrasound Methods for Shape Analysis and �Visualization,�Methods: Special Issue on Real-Time Signal Processing in the Neurosciences (in press, 11/2001).

http://www.escardio.org/communities/EAE/3d-echo-box/3d-echo-atlas

Page 5: András Horváth

Edge-based methods

Step edge vs focal blur, shading artifacts!

Gaussian smoothed edge

Simplest, but difficult method even with high resolution(MR,CT)

2D vs 3D edges

Fisrt order methods:Gradient operatorsHigher order methods:Differential edge detection

Computationaly expensive(Preprocessing, noise filtering)

http://cgv.icu.ac.kr/segmentation

Parameter dependent different intensites

http://www-sop.inria.fr/epidaure/personnel/malandain/segment/edges.html

Page 6: András Horváth

Snakes, template matching, active-contour techniques

G. Hamarneh, J. Hradsky, "DTMRI Segmentation using DT-Snakes and DT-Livewire", The 6th IEEE International Symposium on Signal Processing and Information Technology (ISSPIT), 2006. Keyword(s): Segmentation, Deformable Models, Diffusion Tensor MRI.

Online 3-D Reconstruction of the Right Atrium From Echocardiography Data via a Topographic Cellular Contour Extraction AlgorithmDaniel Hillier, Zsolt Czeilinger, Andras Vobornik, and Csaba Rekeczky IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 57, NO. 2, FEBRUARY 2010

www.haskins.yale.edu/conferences/UltrafestV/abstracts/tagare_poster.pdf

Page 7: András Horváth

Bayesian ApproachAdvantages: more region-based than edge-based The segmentation is optimal (minimum variance) if the generative model is correct • No ad-hoc techniques • Canonical introduction of prior shape information (can be weak)Disadvantages: • Can get trapped in spurious local maxima • Spatially in-homogenous generative models are cumbersome • Many extra parameters • Optimization is difficult

First-order Ultrasound Statistics

Determining region from:MeanStd deviation

www.haskins.yale.edu/conferences/UltrafestV/

abstracts/tagare_poster.pdf

Page 8: András Horváth

Tracking 4D echocardiorgaphy

Pairwise active appearance model and itsapplication to echocardiography trackingS. Kevin Zhou, J. Shao, B. Georgescu, and D. Comaniciu

Page 9: András Horváth

Shape-based Motion Analysis

http://www.ncbi.nlm.nih.gov/pubmed/19565138

http://www.maths.leeds.ac.uk/applied/research.dir/Bio/biological.html

Works well for left ventricle (simple geometry)Works even with simple mathemathical descriptionMore modelbased than observation basedNot applicable for other shapes

Page 10: András Horváth

Cooperative methods

Self organising maps

http://www.viscovery.net/self-organizing-maps

Self organizing maps

Page 11: András Horváth

Segmentation results dependend on:

- Previous expectations (Blood flow, strain, valves)

-Processing time

- Cost (architecture)

Which one is the best solution?

Hybrid methods

Page 12: András Horváth

Biblograpy

Apart from the previously mentioned links

E Angelini, S Homma, "Segmentation of real-time threedimensional ultrasound for quantication of ventricular function: a clinical study on right and left ventricles," Ultrasound in Medicine and Biology, pp. 1143-1158, 2005. 4.1.5

Danping Peng, Barry Merriman and M. Kang, pde based fast local level set methods. Department of Mathematics, University of California at Los Angeles, Los Angeles, California 90095-1555, 1999. 4.1.3

Ch. Brechbuhler and O. Kubler, "Parametrization of closed surfaces for 3-dshape description," Communication Technology Laboratory Image Science SwissFederal Institute of Technology (ETH), p. 1996. 4.1.3

Albert R, "Topology of evolving networks local events and universality,„ Physical Review Letters, p. 5234, 2000. 4.1.1