Computer Aided Diagnosis: Methodologies: Aurélio Campilho

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 36

INEB - Instituto de Engenharia Computer Aided Diagnosis: Methodologies

Biomédica

Aurélio Campilho
(campilho@fe.up.pt, http://www.fe.up.pt/~campilho)

FEUP – Faculdade de Engenharia da Universidade do Porto


INEB – Instituto de Engenharia Biomédica

Docentes:
Aurélio Campilho
Ana Maria Mendonça
Computer Aided Diagnosis
Master Degree on Bioegineering
Part 2: Computer Aided Diagnosis: methodologies
PART 1: Computer-aided Diagnosis: Overview
1. Introduction
2. CAD System typical organization
3. Developments in CAD systems
PART 2: Methodologies
1. Image Enhancement in CAD Systems
2. Image Segmentation in CAD Systems
 Introduction
 Thresholding and clustering
 Texture analysis
 Organ segmentation: snakes
 Detection of suspicious lesions: 4 case studies
3. Feature Extraction and Classification
4. Evaluation and interpretation
PART 3: Applications
1. Breast Cancer
2. Lung Cancer
3. Ophtalmic Diseases
4. Colon Cancer
N
5. Other applications
P
End Part 2: Contents 2
Computer Aided Diagnosis
Master Degree on Bioegineering
2.2 Image Segmentation in CAD Systems
Categorization of segmentation methods

N
P Slide credit: Fernando Monteiro. PhD Thesis

End Introduction 3
Computer Aided Diagnosis
Master Degree on Bioegineering
2.2 Image Segmentation in CAD Systems
CAD segmentation objectives

In a Computer-aided Diagnosis environment, the segmentation objectives are


1. To detect the region-of-interest, by locating the anatomical region where to focus further
analysis. Examples are anatomical ROIs are:
 Lungs in chest images
 Airway trees in lungs
 Arterial tree in lungs
 Ventricules in the heart
 Vascular trees, macular region, optic disk in retinal images
 Gray and white matter in brain images
 ... and many others

2. To characterize and classify the regions into several categories, depending on the application
problem:
 Microcalcifications in mammograms
 Tumor and masses in mammograms
 Lung nodules
 Microaneurisms, exsudates, drusen in retinal images
 Melanomas in skin images
N  Polyps in colon images
 ... and many other
P
End Introduction 4
Computer Aided Diagnosis
Master Degree on Bioegineering
2.2 Image Segmentation in CAD Systems
Segmentation of anatomical regions

Retinal image and vascular tree (slide credit: Ana Maria Mendonça [ ])
Segmenting carotid US images
slide credit: Rui Rocha. PhD Thesis

Segmenting chest images in CT (Fernando Monteiro. PhD Thesis) Segmenting CT chest images
N Slide credit: Carlos Vinhais. PhD Thesis

P
End Introduction 5
Computer Aided Diagnosis
Master Degree on Bioegineering
2.2 Image Segmentation in CAD Systems
Lesion detection

Lung nodule Malignant masses

N
Plaque in carotid artery
P Polyps in colon rectal images
End Introduction 6
Computer Aided Diagnosis
Master Degree on Bioegineering
Segmentation in the Feature Domain
Threshoding

fT(i, j) = 1 if T1 <f(i, j) < T2


0 if elsewhere

CT Image

CT Image Histogram
10000

9000

8000

7000

6000

5000

4000

3000

2000

1000

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Parenchima Soft tissue Bone

N
P Original Image
End Thresholding and clustering 7
Computer Aided Diagnosis
Master Degree on Bioegineering
2. Feature Domain
Threshoding CT Image Blue=th>0.8 (Bone); Green=0.6<th<0.8 (Soft Tissue); Red=0.18Th<0.4 (Parenchima + ...)

Original Image Thresholded Image


imshow(ImageScaled); title('CT Image')
figure; imhist(ImageScaled);title('CT Image Histogram');
I_RGB(:,:,3)=double(ImageScaled>0.8);
I_RGB(:,:,2)=double(ImageScaled>0.6&ImageScaled<0.8);
I_RGB(:,:,1)=double(ImageScaled>0.18&ImageScaled<0.4);
N
figure;imshow(I_RGB);title('Blue=th>0.8 (Bone); Green=0.6<th<0.8
P (Soft Tissue); Red=0.18Th<0.4 (Parenchima + ...)');
End Thresholding and clustering 8
Computer Aided Diagnosis
Master Degree on Bioegineering
2. Feature Domain
Clustering in the feature domain: K- means
 T 

Φ (clusters,data) = ∑  ∑ ( x j − ci ) ( x j − ci ) 
T
Objective function:
i∈clusters  j∈ith cluster 

Algorithm

1. Choose k data points to act as cluster centers.

2. Until the cluster centers are unchanged

3. Assign each data point to the cluster whose center is closest

4. Replace the cluster centers with the actual mean

5. End

N
P
End Thresholding and clustering 9
Computer Aided Diagnosis
Master Degree on Bioegineering
2. Feature Domain
Clustering in the feature domain: K- means
%% Testing kmeans (1)
%
close all; clear
Image=imread('cameraman.tif');
[r c]=size(Image);
A=Image(:);
[list kcenters]=kmeans(double(A),5);
Image_shape=reshape(list,r,c);imtool(mat2gray(kcenters(I
mage_shape)));
kcenters
figure;imhist(mat2gray(Image))
% hold on;
figure;imhist(kcenters(Image_shape)/255);
% hold off

4000
18000
3500
16000

3000 14000

2500 12000

2000 10000

8000
1500
6000
1000
4000
500
2000
N 0
0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1


P 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

End 6. IMAGE SEGMENTATION 10


Computer Aided Diagnosis
Master Degree on Bioegineering
2. Feature Domain
K = 2 Cluster centers = [0 0.55047]

Clustering in the feature domain: K- means


CT scan K= 2
Cluster centers = [0 0.55]

K = 3 Cluster centers = [0.69031 0.47558 0]

K= 3
Cluster centers = [0,69 0.48 0]

10000

8000
K = 4 Cluster centers = [0.69511 0 0.47243 0.57648]

K= 4
6000
Cluster centers = [0,69 0 0.47 0.58]
4000

2000

N 0

P 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

End 6. IMAGE SEGMENTATION 11


Computer Aided Diagnosis
Master Degree on Bioegineering
Texture analysis
Textures are observed on object surfaces as wood, sand, grass and cloth. The term texture is usually applied
when there is a repetition of basic elements, the texel. A texel is an image patch with various pixels, with a
positioning rule, that can be periodic, random or pseudo-periodic.
Natural textures are often random, while arificial textures are deterministic or periodic.

N
P
End 2.2 Segmentation in CAD systems 12
Computer Aided Diagnosis
Master Degree on Bioegineering
Texture analysis
Texture in medical images

Feature based segmentation methods use pixel and local information in order to provide information for the
segmentation process. Based on that information a classifier or a clustering algorithm can be used to identify
regions, being anatomical regional or lesions. Texture is an important feature often used in medical images, as in
the images below:

Plaque in carotid artery

Malignant masses

Retinal image

N
X-Ray image
P
End 2.2 Segmentation in CAD systems 13
Computer Aided Diagnosis
Master Degree on Bioegineering
Texture analysis
Statistical Analysis of Texture

Basic texture features can be derived from the probability distribution function (PDF) which can be estimated from
the normalized unidimensional gray-level histogram. For characterization intensity variations we may use the kth
central moment of the PDF h(g), defined by:
G −1
mk = ∑ ( g − µ f ) h( g )
k

g =0

where
g = 0,1, ⋅⋅⋅, G − 1,are the level values of image f
µ f is the mean intensity of f
The first moments are:
G −1
1 moment → m1 = ∑ gh( g )
st
(mean of f )
g =0
G −1
2 moment → σ = m2 = ∑ ( g − µ f ) h( g )
nd 2 2
f (variance of f , measures inhomogeneity)
g =0

The normalized 3rd and 4th moments, skewness and kurtosis, measure simetry and uniformity, respectively:
m
Skewness → s f = 3/3 2
m2
N m4
kurtosis → s f =
P m2 2
End 2.2 Segmentation in CAD systems 14
Computer Aided Diagnosis
Master Degree on Bioegineering
Texture analysis
Texture features
Energy: ∑ Pθ
a ,b
2
,d (a, b)

Entropy: ∑ Pθ
a ,b
,d ( a , b ) log 2 Pθ , d ( a , b )

Maximum
probability: [
max Pθ , d ( a , b ) ]
κ
Contrast: ∑a ,b
a − b Pθλ, d ( a , b )
(tipicamente λ = 1, κ = 2 )

Pθλ, d ( a , b )
Inverse difference
moment:

a ,b ; a ≠ b a−b
κ

( ab ) Pθ , d ( a , b ) − µ x µ y
Correlation: ∑ a ,b σ xσ y
The mean and standard deviation are given by:

µ x = ∑ a ∑ Pθ , d ( a , b ) σ x = ∑ (a − µ x )2 ∑ Pθ , d ( a , b )
a b a b

N µ y = ∑ b ∑ Pθ , d ( a , b ) σ y = ∑ (b − µ y )2 ∑ Pθ , d ( a , b )
b a b a
P
End 2.2 Segmentation in CAD systems 15
Computer Aided Diagnosis
Master Degree on Bioegineering
Texture analysis
Statistical Analysis of Texture

A texture is based on repeated occurrence of some gray-level configuration. This configuration varies rapidly in fine textures, more
slowly in coarse textures. The occurrence of gray-level configuration may be described by 2-dimensional histograms, represented as
matrices of relative frequencies, called grey-level co-occurrence matrices (GLCM), Pθ. d(a, b), defined as

P0º, d (a, b) = {((k , l ), (m, n)) ∈ D : k − m = 0, l − n = d , f (k , l ) = a, f (m, n) = b }


P45º, d (a, b) = {((k , l ), ( m, n)) ∈ D : k − m = d , l − n = −d , f ( k , l ) = a, f ( m, n) = b}
P90º, d (a, b) = {((k , l ), (m, n)) ∈ D : k − m = d , l − n = 0, f (k , l ) = a, f (m, n) = b }
P135º, d (a, b) = {((k , l ), (m, n)) ∈ D : k − m = d , l − n = d , f (k , l ) = a, f ( m, n) = b}

1 1 1 2 2 0
a/b 0 1 2 3 a/b 0 1 2 3
0 0 0 1 1 0
0 10 5 3 0 0 6 4 5 0
0 0 0 1 1 0
f (i, j ) → P0º,1 (a, b) → 1 5 8 2 1 P135º,1 (a, b) → 1 5 6 1 0
0 0 2 2 2 1
2 3 2 10 3 2 5 2 2 6
0 2 2 3 3 2
3 0 1 3 4 3 0 0 6 0
0 1 3 3 2 2

N
P
End 2.2 Segmentation in CAD systems 16
Computer Aided Diagnosis
Master Degree on Bioegineering
Texture analysis
Statistical Analysis of Texture: example Texture Features based on a glcm with ∆x=0; ∆y=3

Fabric.0000 Fabric.0007 Water.0001 Woodgrain-1.2.9 Food.0000

Image asm cont corr idm entrop var prom shade diag
Fabric.0000 5.6346E-004 1.1357E+004 -2.8372E+002 8.8447E-003 3.3217E+000 3.9490E+003 -4.0492E+002 -1.0156E+004 3.9068E+007
Fabric.0007 4.9694E-003 2.7061E+004 -7.3104E+004 7.0106E-005 2.4187E+000 9.1777E+003 3.2177E+003 1.2897E+003 1.8441E+007

Water.0001 3.3531E-001 8.7598E+003 -3.6290E+006 1.1714E-004 5.2304E-001 2.2321E+003 5.6085E+001 9.6205E-002 3.2197E+003

Woodgrain 3.9791E-004 1.4466E+004 -3.7717E+001 7.2021E-003 3.5008E+000 4.1525E+003 2.9711E+003 1.9842E+005 1.6233E+008
Food.0000 2.4603E-004 8.8585E+003 -9.1553E+001 1.2870E-002 3.7252E+000 4.4631E+003 -3.6655E+002 -6.3560E+003 1.3887E+008

Haralick Features
 asm - Angular Second Moment of a GLCM.  entrop - Entropy of a GLCM.
 cont - Contrast of a GLCM.  var - Variance of a GLCM.
 corr - Correlation of a GLCM.  prom - Cluster Prominence of a GLCM.
N  idm - Inverse Difference Moment of a GLCM.  shade - Cluster Shade of a GLCM.
P  entrop - Entropy of a GLCM.  diag - Diagonal Moment of a GLCM.

End 2.2 Segmentation in CAD systems 17


Computer Aided Diagnosis
Master Degree on Bioegineering
Texture analysis
Statistical Analysis of Texture: matlab example

I3 = imread('Food.3.png');
I3 = im2double(I3);
glcm3 = graycomatrix_(I3, [0 1], 128, [0 1]);
figure;
subplot(1,2,1); imshow(I3); title('texture 3');
subplot(1,2,2); imshow(mat2gray(glcm3)); title('GLCM:
horizontal offset ');

I4 = imread('Water.4.png');
I4 = im2double(I4);
glcm4 = graycomatrix_(I4, [0 1], 128, [0 1]);
figure;
subplot(1,2,1); imshow(I4); title('texture 4');
subplot(1,2,2); imshow(mat2gray(glcm4)); title('GLCM:
horizontal offset ');

N
P
End 2.2 Segmentation in CAD systems 18
Computer Aided Diagnosis
Master Degree on Bioegineering
Texture analysis
Texture classification

Texture Classification

Develop a texture classifier !

N
P
End 2.2 Segmentation in CAD systems 19
Computer Aided Diagnosis
Master Degree on Bioegineering
Texture analysis
Gabor filters: multichanel filtering approach

This approach to texture analysis is referred to as the multi-channel filtering approach. This
approach is inspired by a multi-channel filtering theory for processing visual information in the early
stages of the human visual system.

The image is decomposed into several filtered images with limited spectral information. The main
issues involved in the multi-channel filtering approach to texture analysis are [1]:
1) functional characterization of the channels and the number of channels,
2) Extraction of appropriate texture features from the filtered images,
3) the relationship between channels (dependent vs. independent), and
4) integration of texture features from different channels to produce a segmentation.

Several approaches uses a bank of Gabor filters. A Gabor filter in the spatial domain is given by

• uo and φ are the frequency and phase of the sinusoidal


 1  x2 y 2  
h( x, y ) = exp  −  2 + 2   cos ( 2π uo + φ )
plane wave along the x -axis (i.e. the 0º orientation).
 2  σ x σ y  • σx and σy are the space constants of the Gaussian
   envelope along the x- and y-axis.

A Gabor filter with arbitrary orientation, θ0, can be obtained via a rigid rotation of the x-y coordinate
N system. These 2D functions have been shown to be good fits to the receptive field profiles of
simple cells in the striate cortex.
P
End 2.2 Segmentation in CAD systems [1] IEEE Jain paper 20
Computer Aided Diagnosis
Master Degree on Bioegineering
Texture analysis
Gabor filters: multichanel filtering approach

 1  x2 y 2   0.01

h( x, y ) = exp  −  2 + 2   cos ( 2π uo + φ )
 2  σ x σ y 
   0.005

-0.005

-0.01
100
50
0 40 50 60 70
0 10 20 30

The frequency- and orientation-selective properties of a Gabor filter are more explicit in its
frequency domain representation. With φ = 0, the Fourier transform of the Gabor function is
 1  ( u − u )2 v 2    1  ( u + u )2 v 2  
H (u , v) = A exp  −  o
+ 2   + A exp  −  o
+ 2 
 2 

σu2
σv 
  2 

σu2
σ v  
   

With σu = 1/2πσx, σv = 1/2πσy, and A = 1/2πσxσy


N
The Fourier domain representation specifies the amount by which the filter modifies or
P modulates each frequency component of the input image.
End 2.2 Segmentation in CAD systems [1] IEEE Jain paper 21
Computer Aided Diagnosis
Master Degree on Bioegineering
Texture analysis
Gabor filters: multichanel filtering approach
Extracting Image Structure with Filter Banks

Edge Filters

Gaussian

Bar Filters

LoG

N
Manik Varma*, Andrew Zisserman, Unifying statistical texture classification frameworks, Image
P and Vision Computing 22 (2004) 1175–1183

End 2.2 Segmentation in CAD systems 22


Computer Aided Diagnosis
Master Degree on Bioegineering
Texture analysis
Gabor filters: multichanel filtering approach
Extracting Image Structure with Filter Banks

N
P
End 2.2 Segmentation in CAD systems 23
Computer Aided Diagnosis
Master Degree on Bioegineering
Organ segmentation: snakes
Snakes

A snake is a deformable curve that from an initial position evolves towards the contour of
the object.
Medical Applications [3]

Main references:
[1] M. Kass, A. Witkin, and D. Terzepoulos, “Snakes: Active Contour
Models”, International Journal of Computer Vision, 321-331 (1988).
[2] M. Nixon, A. Aguado, Feature Extraction & Image Processing (Chapter
6), Newnes, 2002.
[3] A. Blake and M. Isard, Active Contours, Springer, 1998
N [4] In Chenyang Xu and Jerry L. Prince, Active Contours, Deformable
Models, and Gradient Vector Flow, in http://iacl.ece.jhu.edu/projects/gvf/
P
End 2.2 Segmentation in CAD systems 24
Computer Aided Diagnosis
Master Degree on Bioegineering
Organ segmentation: snakes
Snakes: Contours

A snake is an elastic curve defined by the parameter curve v( s ) = ( x( s ), y ( s ))

2 5

1
6

v1 = (x1, y1)
x(s) y(s)
35 70
4 6
30 60

125 6
1 50

20
2 40

15
5
5 30
2
10
3 20

5
N 4
10

0
1 3 1
0 50 100 150 200 250 0
P 0 50 100 150 200 250

End 2.2 Segmentation in CAD systems 25


Computer Aided Diagnosis
Master Degree on Bioegineering
Organ segmentation: snakes
Snakes: Contours

The evolution from an initial contour to the contour of the object is expressed by an energy
minimisation process. The energy functional includes contour properties that control the way
the contour can strech and curve and image properties like edge magnitude, characterized by
the contour’s internal energy Eint, and the external image energy Eimage. Thus, the snake energy,
for a normalized length s ∈ [0, 1] around the snake is given by

1
Esnake = ∫E
s =0
int ( v( s )) +Eimage ( v ( s ))ds

The goal is to minimize this energy functional, i.e. the final energy of the contour is minimal.

N
P
End 2.2 Segmentation in CAD systems 26
Computer Aided Diagnosis
Master Degree on Bioegineering
Organ segmentation: snakes
Snakes: Contours
Internal Energy:

The internal energy depends on the intrinsic propoerties of the curve and is a weigthed
summation of elastic energy and bending energy, ie, of 1st and 2nd order derivatives around the
contour

2 2 2
dv ( s ) d v( s)
Eint = α ( s ) + β (s)
ds ds 2

Elastic energy: a high value implies a Curvature energy: measuring the energy
high rate of change of the contour due to bending

The choice of the values of α and β controls the shape the snake aims to attain. Low values of α imply the
points can change in space greatly, whereas higher values imply that the snake aims to attain evenly spaced
contour poins. Low values of β imply that curvature is not minimized and the contour can form corners whereas
high values predispose the snake to smooth contours (note that the bending energy is minimum for a circle) .

N
P
End 2.2 Segmentation in CAD systems 27
Computer Aided Diagnosis
Master Degree on Bioegineering
Organ segmentation: snakes
Snakes: The greedy algorithm for snakes
There are many approaches for minimizing the Energy functional. It is to start with the study of the
Greedy Algorithm, which implements the minimization as purely discrete process, as illustrated in
the block diagram.
Define snake points and
parameters α, β and γ

Start with first snake point

Initialise minimum energy


and co-ordinates

Determine co-ordinates of
neighbourhood point with
lowest energy

Yes
Set new co-ordinate points
to new minimum

More snake
points ?

Greedy algorithm
N No
Finish iteration
P
End 2.2 Segmentation in CAD systems 28
Computer Aided Diagnosis
Master Degree on Bioegineering
Organ segmentation: snakes
Snakes: The greedy algorithm for snakes

The energy minimised for each snake point is:

Esnake = Eint ( v ( s )) + Eimage ( v ( s ))


or
2 2
dv ( s ) d 2 v( s)
Esnake = α ( s) + β (s) 2
+ γ ( s ) Eimage
ds ds
The contour derivatives can be numerically approximated by

d d2
v( s ) ≈ v s +1 − v s 2
v( s ) ≈ v s +1 − 2 v s + v s −1
ds ds

The image energy functional can be implemented using the magnitude of the edge operator
after smoothing, or
Eimage = −∑ Gx ( xi , yi ) + G y ( xi , yi )
2 2

with
N ∂ ∂
Gx = Gσ * I Gy = Gσ * I
P ∂x ∂y
End 2.2 Segmentation in CAD systems 29
Computer Aided Diagnosis
Master Degree on Bioegineering
Organ segmentation: snakes
Snakes: The greedy algorithm for snakes

Number of iterations (α = 0.05, β = 0, γ = 0,6):


0 2 4 20 40

Heart Image
Number of iterations (α = 0.05, β = 0, γ = 0,6):

N
P 0 10 100
End 2.2 Segmentation in CAD systems 30
Computer Aided Diagnosis
Master Degree on Bioegineering
Organ segmentation: snakes
Snakes: The greedy algorithm for snakes

CT chest image
Number of iterations (α = 0.05, β = 0, γ = 0,6):

0 10 60

N
P
End 2.2 Segmentation in CAD systems 31
Computer Aided Diagnosis
Master Degree on Bioegineering
Detection of suspicious lesions
4 case studies: A CAD system in mammography [2]

[2] Rianne Hupse, Nico Karssemeijer, “Use of Normal Tissue Context in Computer-Aided Detection of Masses in Mammograms”,
N IEEE Trans on Medical Imaging, 12, 2033-2041, 2009.
P
End 2.2 Segmentation in CAD systems 32
Computer Aided Diagnosis
Master Degree on Bioegineering
Detection of suspicious lesions
4 case studies: A CAD system in Lung Cancer [7]

Architecture of multiple
MTANNs (multi-MTANN)

LUNG CANCER continues to rank as the leading cause of cancer deaths among Americans; the number of lung cancer deaths in
each year is greater than the combined number of breast, colon, and prostate cancer deaths. Because CT is more sensitive than
chest radiography in the detection of small nodules and of lung carcinoma at an early stage, lung cancer screening programs are
being investigated in the United States and Japan with low-dose helical computed tomography (LDCT) as the screening modality.
It may be difficult, however, for radiologists to distinguish between benign and malignant nodules on LDCT. Computer-aided
diagnostic (CAD) schemes for distinction between benign and malignant nodules in LDCT would be useful for reducing the
number of “unnecessary” follow-up examinations.

[7] Kenji Suzuki, Feng Li, Shusuke Sone, and Kunio Doi, Computer-Aided Diagnostic Scheme for Distinction Between Benign and
N Malignant Nodules in Thoracic Low-Dose CT by Use of Massive Training Artificial Neural Network IEEE Trans on Medical
P Imaging, VolL. 24, no. 9, September 2005.

End 2.2 Segmentation in CAD systems 33


Computer Aided Diagnosis
Master Degree on Bioegineering
Detection of suspicious lesions

4 case studies: A CAD system in Colorectal cancer [5]

Colorectal cancer is the second leading cause of mortality due to cancer in the western world. Adenomatous colorectal polyps are
considered important precursors to colon cancer. It has been shown that screening for such polyps can significantly reduce the
incidence of colon cancer. Computer-aided detection (CAD) of polyps may enhance the efficiency and also increase the
sensitivity. This is specifically important for large-scale screening. Recent studies show that the sensitivity of CAD systems is
already comparable to the sensitivity of optical colonoscopy and radiologists using CTC.

The best indicator of the risk that a polyp is malignant or turns malignant over time is size. The consensus is that patients with a
polyp of at least 10 mm must be referred to optical colonoscopy for polypectomy.

[5] van Ravesteijn, V. F.; van Wijk, C.; Vos, F. M.; Truyen, R.; Peters, J. F.; Stoker, J.; van Vliet, L. J.; Computer-Aided Detection
of Polyps in CT Colonography Using Logistic Regression, IEEE Transactions on Medical Imaging, Volume 29, Issue 1, Jan. 2010
N Page(s):120 – 131, 2010.
P
End 2.2 Segmentation in CAD systems 34
Computer Aided Diagnosis
Master Degree on Bioegineering
Detection of suspicious lesions

4 case studies: A CAD system in Diabetic Retinopathy [8]

Exudates identification results

Currently, there is an increasing interest for setting up medical systems that can screen a large number of people for sight
threatening diseases, such as diabetic retinopathy. This paper presents amethod for automated identification of exudate
pathologies in retinopathy images based on computational intelligence techniques. The color retinal images are segmented using
fuzzy c-means clustering following some preprocessing steps, i.e., color normalization and contrast enhancement. The entire
segmented images establish a dataset of regions. To classify these segmented regions into exudates and nonexudates, a set of
initial features such as color, size, edge strength, and texture are extracted.

[8] Alireza Osareh, Bita Shadgar, and Richard Markham, A Computational-Intelligence-Based Approach for Detection of Exudates
N in Diabetic Retinopathy Images, IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 13, NO. 4,
JULY 2009.
P
End 2.2 Segmentation in CAD systems 35
Computer Aided Diagnosis
Master Degree on Bioegineering
Part 2: Computer Aided Diagnosis: methodologies
[1] Kunio Doi, “Computer-aided diagnosis in medical imaging: Historical review, current status and future potential”,
Computerized Medical Imaging and Graphics 31 (2007) 198–211.
[2] Rianne Hupse, Nico Karssemeijer, “Use of Normal Tissue Context in Computer-Aided Detection of Masses in
Mammograms”, IEEE Trans on Medical Imaging, 12, 2033-2041, 2009.
[3] Rui Ribeiro, “Lung Nodule Detection in Chest Radiographs”, MSc Thesis, FEUP, 2008.
[4] M. Giger, “Computer-aided Diagnosis in Diagnostic Mammography and Multimodality Breast Imaging”, RSNA
Categorical Course in Diagnostic Radiology Physics: Advances in Breast Imaging—Physics, Technology, and Clinical
Applications 2004; pp 205–217, 2004.
[5] van Ravesteijn, V. F.; van Wijk, C.; Vos, F. M.; Truyen, R.; Peters, J. F.; Stoker, J.; van Vliet, L. J.; Computer-Aided
Detection of Polyps in CT Colonography Using Logistic Regression, IEEE Transactions on Medical Imaging, Volume
29, Issue 1, Jan. 2010 Page(s):120 – 131, 2010.
[6] Rui Rocha, Aurélio Campilho, Jorge Silva, Elsa Azevedo, Rosa Santos, Segmentation of the carotid intima-media
region in B-mode ultrasound images, Image and Vision Computing, Volume 28, Issue 4, Pages 614-625, 2010.
[7] Kenji Suzuki, Feng Li, Shusuke Sone, and Kunio Doi, Computer-Aided Diagnostic Scheme for Distinction Between
Benign and Malignant Nodules in Thoracic Low-Dose CT by Use of Massive Training Artificial Neural Network IEEE
Trans on Medical Imaging, VolL. 24, no. 9, September 2005
[8] Alireza Osareh, Bita Shadgar, and Richard Markham, A Computational-Intelligence-Based Approach for Detection of Exudates in
Diabetic Retinopathy Images, IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 13, NO. 4,
JULY 2009.

N
P
End References 36

You might also like