T17 – Computer-Assisted Diagnostics Can you answer?

T17 – Computer-Assisted Diagnostics
Prof. Dr. Thomas M. Deserno, né Lehmann
Department of Medical Informatics
RWTH Aachen University, Aachen, Germany
Thomas M. Deserno
Computer-Aided Diagnosis
Motivation

Feature extraction techniques



Why do Haralick’s texture features work?

Do I need shape descriptors in image-based
diagnosis?

Take a supervised or unsupervised classifier?

How to generate reliable ground truth?

Can I compare algorithms without ground truth?
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 2

Benign moles

Malignant moles
Texture-based features
Shape-based features
Ground truth
Example: STAPLE
Example

Idefix
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 3
Motivation

How can I chose appropriate texture analysis
approaches?
Evaluation


What is “texture” and can it be measured?

Problem: How to Differ?



T17 – 1
Overview

Can you answer?
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 4
Example Textures
Goal

Gray scale textures

Color textures
Brodatz
image
imaging
image processing
and analysis
VisTex, MIT
diagnosis
patient
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 5
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 6
Medical Textures

Bone structure


Skin melanoma

Photography





Mamma calcification

Cell structures



Mammography
Computer-Aided Diagnosis
T17 – 7






Co-occurrence matrices (CM)
Haralick’s features
Markov Random Fields

Function of displacement vector d
Without spatial correspondence
pixel 1
255
200
Fourier analysis
Gabor analysis
pixel 1
pixel 2
Computer-Aided Diagnosis


50
T17 – 9
Thomas M. Deserno

Computer-Aided Diagnosis
T17 – 10
Idea





Haralick RM, Shanmugam K, Dinstein I
Textural Features for Image Classification
IEEE Trans System Man Cybern 1973; 3(6): 610-21
Further developments


T17 – 11
Capture texture properties from CM
Contrast
Coarseness
Initial publication

Computer-Aided Diagnosis
50 100 150 200 255
Haralick‘s Texture Features
b: [(i,j),(i+1,j)]
c: [(i,j),(i,j+1)]
d: [(i,j),(i+2,j-2)]
e: [(i,j),(i+2,j+1)]
Thomas M. Deserno
pixel 2
0
Displacement vector

100
0
Co-Occurrence Matrix

150
displacement d
Fractal dimension
Box count
Thomas M. Deserno
T17 – 8
Co-occurrence matrix (CM)
Fractal (self similarity)


Computer-Aided Diagnosis
Signal-Theoretical (deterministic signal)


Thomas M. Deserno
representation ambiguous
2nd Order Statistics
Statistical (random process)

unique description
 Mathematical
Texture Analysis

Coarseness
Contrast
Regularity
Direction (rotation (in-)variance)
Resolution (macro- / micro-texture)
 No
Microscopy
Thomas M. Deserno

Visually distinguishable patterns
Radiography
Histology


Texture Properties
De-correlation of measures
Reduction of measure number
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 12
Haralick‘s Texture Features
Markov Random Fields

h1: Angular Second Moment

h7: Sum Variance

h2: Contrast

h8: Sum Entropy

h3: Correlation

h9: Entropy

h4: Sum of Squares (variance)

h10: Difference Variance

h5: Inverse Different Moment

h11: Difference Entropy

h6: Sum Average

h12 - h14: Information
Measures
of Correlation


Statistically predict local gray scales

Left: 1

Right:
P(1) = 58%
P(2) = 42%
Depends on the neighborhood system

Clique
Christine Hivernat & Xavier Descombes
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 13
Thomas M. Deserno
Fourier Analysis



F (u, v)

Spatial domain
Fourier domain
Gabor transform




Windowed Foiurier transform
Gaussian window
Color texture analysis
Drawback

T17 – 14
Gabor Analysis
f (m, n)
Example
Computer-Aided Diagnosis
H
Averaging
the spatial
domain
real
S
imaginary
V
original
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 15
Gabor Color Texture Analysis
H

real

Gabor
T17 – 16


V
No. of self similar objects
to cover the original object
Box count method


spectrum
Self similarity & scale invariance
Infinite detail at every position
Definition
imaginary
complex
representation
spectrum
Computer-Aided Diagnosis
Mandelbrot set

S
decomp.
complex
representation
Fractal Dimension

original
decomp.
Thomas M. Deserno
Use boxes of side length R  Rn
Sub-divide boxes into (2m)n smaller boxes
Gabor
W. Beyer
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 17
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 18
Example: Osteoporosis Detection
Shape

Fractal dimension  “filling factor”

Bones modeled as cylinders, range between







Hollow  “Osteoporosis”, no trabeculae,  = 1.0
Normal,   1.7 to 1.8
Solid  “Osteopetrosis”,  = 2.0



Algorithm




Segment bone from CT
Select region of interest
Compute 
Edge-based
Perimeter
Chain code
Curvature scale space
Fourier descriptor
Region-based analysis



Bounding box
Best fitting ellipse
Solidity
http://www.rad.washington.edu/exhibits/fractal.html
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 19
Thomas M. Deserno
Perimeter




Direction to next boundary pixel
Neighborhood definition
7
0
1
6
pos
2
5
4
3

Contour represented by curvature

Shape descriptor

Perimeter

T17 – 20
Curvature
Chain code

Computer-Aided Diagnosis
Length of chain code

Location of zero crossings
of curvature relative to starting point
Comparison of shapes
by cyclic shifting of starting points
Example

Chain code <Contour Len=“11“
StartX=“1“ StartY=“1“>
22344657071
<\Contour>

Perimeter = 11
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 21
Curvature Scale Space (CSS)



T17 – 22

Example
Smooth contour
Gaussian kernel
Iteration



Computer-Aided Diagnosis
CSS
Algorithm

Thomas M. Deserno
Repeat until all inflection points vanished
Record positions of inflection point
image
Convergence

Elliptical shape
Thomas M. Deserno
inflections
Computer-Aided Diagnosis
T17 – 23
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 24
Example: Cilia-driven Mixing

Fourier Descriptors
CSS smoothing in each state of progression

Fourier representation of contour



Only main components





Thomas M. Deserno
Computer-Aided Diagnosis



Thomas M. Deserno



Standardized photographs






Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 27


Thomas M. Deserno

Center of gravity
Principal axis
Area



Solidity  1  solid object
Solidity < 1  irregular object or holes contained
Definitions
area
convex area
convexity 
T17 – 29
solidity=1
convex perimeter
perimeter
compactness 
Computer-Aided Diagnosis
T17 – 28
Measure of densitiy
solidity 
Thomas M. Deserno
Computer-Aided Diagnosis
Solidity
Ellipse equals object in

68 female, 86 male skulls
Significant difference (Wilcoxon test, p = 0.0043)
Logistic regression model significant (p = 0.0033)
Model quality poor (r2 = 5.89%)
S alone does not solve the classification problem
Only 63% correct classifications
male
Best Fitting Ellipse

Fourier descriptors
Hierarchical smoothing
(Gaussians)
Convexity index S
Analysis

female
T17 – 26
Processing

Female: smooth
Male: more corner like
Data

Computer-Aided Diagnosis
Example: Fourier Descriptors
Hypothesis: shape of orbit differs in gender

S(0) = ?
S(0), S(1) = ?
S(0), S(1), S(2), = ?
...
Similar to CCS
T17 – 25
Example: Fourier Descriptors
s(t), (x,y)t=1, (x,y)t=2, ...
S(f) = F {s(t)}
Thomas M. Deserno
solidity=0.592
4  area
perimeter 2
Computer-Aided Diagnosis
T17 – 30
Example Compactness

Evaluation of Image Analysis
L-929 fibroblast in contact with ethanol (toxin)

Example: Segmentation

Rank approaches


0%



5%

10%
Thomas M. Deserno

Computer-Aided Diagnosis

Wenzel A, Hintze H: Editorial Review:







Computer-Aided Diagnosis

Computer-Aided Diagnosis
Robert M. Haralick
The choice of gold
standard
at SPIE MI 2000:
for evaluation tests
for caries diagnosis.
sometimes it‘s not even
[Dentomaxillofaccopper,
Radiolit‘s
1999;
28:132-136]
plastic!
A robust gold standard is a method that



Is itself precise, i.e. reproducible
Reflects the patho-anatomical appearance of the disease
Is established independently
of the diagnostic method under evaluation
Example


T17 – 33
Caries diagnostic based on radiography exams
Confirmed by histological analysis
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 34
Example: Manual References
Manual references
Thomas M. Deserno
T17 – 32
Wenzel A, Hintze H: Editorial Review:


Generation of Ground Truth


Caries diagnostic based on radiography exams
Confirmed by histological analysis
Thomas M. Deserno
Computer-Aided Diagnosis
Definition: Ground Truth / Gold Standard
Is itself precise, i.e. reproducible
Reflects the patho-anatomical appearance of the disease
Is established independently
of the diagnostic method under evaluation
Example

Thomas M. Deserno
The choice of gold standard
for evaluation tests for caries diagnosis.
[Dentomaxillofac Radiol 1999; 28:132-136]
A robust gold standard is a method that
Which one is best?
Which one is correct?
Ground truth needed
T17 – 31
Definition: Ground Truth / Gold Standard
Ultrasound partitioning

T17 – 35
Inter-observer mean
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 36
Evaluation Without Ground Truth

Recap: Medical Image Processing




Evaluation Without Ground Truth

Exactness & robustness required
Variety & variability
Uncertainty


Which curve is better?
Computer-Aided Diagnosis
Usage


T17 – 37
STAPLE

Establish ground truth
Quality metric for comparison
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 38
Example: Identification of Dental Fixtures
Example

Type











http://www.crl.med.harvard.edu
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 39
Image Processing
Scheme
Thomas M. Deserno

digitization
optimization
APA Ceram
Bonefit
Branemark
Frialit
TPS screw
...
Computer-Aided Diagnosis
T17 – 40
IDeFix: Model for Feature Extraction
acquisition
calibration
Transdental
Transossal
Subperiossal
Enossal
…
Device


Compute probabilistic estimate of true segmentation
 Expectation maximization

Healthy bone tissue
Tumorous tissue
Soft tissue
Background
Thomas M. Deserno
Given a set of segmentations
 Manual
 Automatic
Segmenting uncertainty

Simultaneous Truth and Performance Level
Estimation (STAPLE)
registration
transformation
Measures
d1

Length: l

Diameter: d1, d2, d3
Cross section

 Implant: AO
 Tap hole: AB
filtering
AB
AO
l
d2
t
 Rotation lock: AL
surface
. reconstr..
feature extraction
compression
illumination
segmentation
archiving

No. of turns: t
AL
retrieval

Cone impact: c = d1 / d2
Slimness: s = l / (d1+d2+d3)
d3
shading
classification

display
Thomas M. Deserno
interpretation
measurement
Computer-Aided Diagnosis
communication
T17 – 41

Ground truth from vendor specifications
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 42
IDeFix: Image Processing Chain
IDeFix: Evaluation

Collect images

Establish ground truth

Perform experiments


Branemark

How many?
Manual class labels
Count errors
TPS
Frialit
feature space
Thomas M. Deserno
Computer-Aided Diagnosis
Branemark
T17 – 43
Summary

Texture features






Co-occurrence matrices
Markov random fields
Fourier & Gabor analysis
Fractal dimension (box count)
Shape features





Evaluation




Example

Perimeter (chain code)
Curvature scale space
Fourier descriptor
Best fitting ellipse
Solidity





Thomas M. Deserno
Ground truth
Gold standard
STAPLE algorithm
Computer-Aided Diagnosis
Acquisition (x-ray)
Preprocessing (histogram)
Segmentation (thresholding)
Feature extraction
(shape-based)
Classification (kNN)
Evaluation
T17 – 45
Thomas M. Deserno
Computer-Aided Diagnosis
T17 – 44