Ter Haar Romeny, FEV MIT AI Lab Automatic Polyp Detection.

Post on 31-Mar-2015

219 views 2 download

Tags:

Transcript of Ter Haar Romeny, FEV MIT AI Lab Automatic Polyp Detection.

ter Haar Romeny, FEV

MIT AI Lab

AutomaticPolypDetection

ter Haar Romeny, FEV

Enhancement byGaussian curvature

PMS

CT slice with tagged residual sticking to the wall

Same slice after electronic cleansing

Philips MS

Electronic colon cleansing

ter Haar Romeny, FEV

Current visualization

Normal doseSmooth surface

Low doseBlobs appear

Normal doseRough surface

ter Haar Romeny, FEV

Proposed solutions

Bilateral filtering blobs

Gradient smoothing rough surface

ter Haar Romeny, FEV

Results: normal dose

ter Haar Romeny, FEV

Results: all dose levels

1.6 mAs 6.25 mAs 64 mAs

ter Haar Romeny, FEV

Extract vasculature with ‘vesselness’

From T1w MRI with contrast

Frangi’s vesselness measure [Frangi et al., 1998]

Enhance tubular structures while reducing other morphologies

E. Brunenberg, MSc project

ter Haar Romeny, FEV

Vesselness measure

Based on eigenvalue analysis

of Hessian:

two low eigenvalues

one high eigenvalue

ter Haar Romeny, FEV

Vesselness - 1

Eigenvalue analysis of Hessian:

extract directions of principal curvature

Hessian:

where

and

xx xy xz

yx yy yz

zx zy zz

I I I

H I I I

I I I

G

I2

2 ,I

xx x

G e

2

223

2

1,

2

x

x

ter Haar Romeny, FEV

Vesselness - 2Eigenvalues ordered as |λ1| ≤ |λ2| ≤ |λ3|

Bright vessel region: λ1 small, ideally zero; λ2 and λ3 large

but negative.

Ratio for blobness:

Ratio for plate-like:

Image structure:

BR1

2 3

,1 forblob-like

AR2

3

,0 for line-like

S 2 2 21 2 3 ,1 formuch contrast

ter Haar Romeny, FEV

Vesselness - 3

Total vesselness function:

Parameters:

α = β = 0.5

c = 0.5 * maximum Hessian norm

Multiscale approach:

22 2

22 2

2 3

22 2

0 if 0 or 0

,1 1 otherwise

BARR S

cV

e e e

x

min max

max ,V V

x x

ter Haar Romeny, FEV

Vessel enhancement filtering

Better delineation of small vessels

Preprocessing before MIP

Preprocessing for segmentation procedure

ter Haar Romeny, FEV

Abdominal MRA

Maximum intensity projection

No 3D information

Overlapping organs

ter Haar Romeny, FEV

2D Example: DSA

ter Haar Romeny, FEV

Scale integration

ter Haar Romeny, FEV

Closest Vessel Projection

ter Haar Romeny, FEV

Trabecular Bone

Bone appears in two forms

Cortical Bone

Trabecular Bone

Trabecular Bone

connected network of rods & plates

loading dependent architecture Wiro Niessen, PhD

ter Haar Romeny, FEV

Stress routes

Wolff’s Law

“The internal structure and external shape of a

bone develop in response to the change in function

and forces acting upon it”

Culman Meyer

“Trabecular pattern is oriented with routes of

stress”

ter Haar Romeny, FEV

Clinical Relevance

Trabecular Architecture important parameter in bone strength

(clinically proven)

Applications for in vivo analysis

determine fracture risk

monitoring structure in aging

monitor degree and development of osteoporosis

(treatment available)

monitoring malgrowth near epiphyses

placing implants and evaluating receipt

ter Haar Romeny, FEV

ter Haar Romeny, FEV

Stress Routes in Ankle

ter Haar Romeny, FEV

MR Ankle, FFE, short TE (300m)

ter Haar Romeny, FEV

CT dry femur (250m)

ter Haar Romeny, FEV

Structural Information

2D 3D Orientation PatternHigh High High High High noisy no preferred orientationHigh Low High High Low tubular structure

High Low Low platelike structureLow Low Low Low Low smooth region

ter Haar Romeny, FEV

3D orientaties

ter Haar Romeny, FEV

Dominant orientations

Orientations preferentially along anatomical axis

Histogram of 3D directions: