Detection of Anatomical Landmarks

Detection of Anatomical Landmarks

Bruno Jedynak

Camille Izard

Georgetown University Medical CenterFriday October 6, 2006

Anatomical Landmarks

• Manually defined points in the anatomy ( geometric landmarks)

• !! Landmarker consistency, variability between exerts

• Used as is to analyze shapes• Used as control point for image

segmentation/registration

Landmarking the hippocampus from Brain MRI

Manual landmarking of the Hippocampus

Automatic landmarking

• Given: a set of manually landmarked images

• Goal: build a system that can landmark new images

• The system must adapt to different kind, different number of landmarks

Automatic landmarking Example:

• Given: 38 images expertly landmarked. K landmarks per image

• Goal: landmark new images• Mean error per new image

Or expert evaluation

Stochastic modeling

• Build a likelihood function:

• Learn:

• For each new image, compute:

Landmarks are points

Define

Template matching paradigm

Identify landmarks with a deformation of the 3d space.

Examples of deformations:

Affine

Splines

Diffeomorphisms

Spline model

Define

Identify

Such that

Forward model

Brain MRI gray-values are modeled as a mixture of Gaussians distributions.

There are 6 components in the mixture: CSF,GM, WM, CSF-GM, GM-WM, VeryWhite (Skull, blood vessels, …)

Forward Model

Tissue Probability Map

csfcsf-gm

gmgm-wm

wm outliers

HoH 0 0.04 0.90 0.06 0 0

Estimating the tissue probability map

• Learn the photometry of each image

• Register each image on the template

• Use the E.M. algo. for mixture of Gaussians to estimate

Automatic landmarking of a new image

• Learn the photometry parameters

• Use gradient ascent to maximize

Results

Results

Results

Current work

• Estimating the std. dev. of the Kernels

• Add control points to generate more complex deformations (K=1)

• Test on schizophrenic and other brains