LEUKO: Computer Decision Support System for Leukemia Diagnosis · 2012. 2. 25. · Leuko + SVM • SVM extension to multiclass task, applying tree-based strategy; • SVM and minimum

LEUKO:LEUKO:Computer Decision Computer Decision Support System for Support System for Leukemia DiagnosisLeukemia Diagnosis

Daniela M. Ushizima

2/4130-May-07 Computer Vision in Leukemia Diagnosis

MotivationMotivation• Leukemia = 30% cancer.

• Pattern recognition: low-cost alternative solution;

• Human visual analysis: • 79.3% to 92.3% correct (lymphocytes);

• 150 cells >> 10 minutes for an hematologist;

• Why lymphoproliferative diseases?

• Software for cell analysis and classification: • computer-aided automation;

• educational purposes;


SynopsisSynopsis1. Human blood cells2. Pattern recognition3. Feature selection4. Results5. Latest developments


1.1. 1.1. WhereWhere leukocytesleukocytes come come fromfrom??


LymphocyteMonocyte

Eosinophil Neutrophil Basophil

1.2. Normal leukocytes1.2. Normal leukocytes


Prolymphocyte

CLLHairy-cell

1.3. Malignant cells1.3. Malignant cells

2.1. Steps in PR design2.1. Steps in PR design

Segmentation

Pre processing

Featureextraction

Classification

Classes

Classes compare

Criteriumfunction

Humanexpertise

Examples

DigitalImage

Feature selection

f={ff={f00,f,f11,…,f,…,fnn--11}}


2.2. Segmentation2.2. Segmentation

segmentsegment

• Regions of interest: nucleus, cytoplasm, red blood cells and background


2.2.1. Methods2.2.1. Methods

• G channel thresholding: where is the nucleus?

• Color segmentation: where are the leukocyte and RBCs?


Green levels

Abs

olut

e fr

eque

ncy

180 200 220 100 98 90 95 95 180 230 221 160 82 95 93 90 230 200 220 100 99 85 82 96

10 200 220 100 50 30 35 25 10 230 221 160 32 35 23 10 20 200 220 100 50 30 35 25

10 200 22 100 50 30 35 25 230 22 160 32 35 23 10 13 100 120 200 15 0 35 25 0

2.2.2. 2.2.2. ThresholdingThresholding


Katz (2000)


Sabino et al (2003)


2.2.3. Color segmentation2.2.3. Color segmentation

trainr

g

b

modelµ,σ

r

g

b


• Image properties;• Nucleus and cytoplasm attributes;• Shape;• Size;• Color;• Texture.

2.3. Feature extraction2.3. Feature extraction

Traditional measures

Subjective descriptions

Shape and sizeShape and size

TextureTexture

ColorColor


2.3.1. Shape and size2.3.1. Shape and size

• Nucleus-cytoplasm ratio (NC):• NC=Anuc/(Anuc+Areacit)

• Factor form=compactness measure (C):• C=P2/A

• Curvature (k):• Concave and convexity points;• Bending Energy (B): 2

1

1 ( )P

tB k t

P == ∑


2.3.2. Fractal dimension2.3.2. Fractal dimension•• CostaCosta: multiscale fractal dimension (2D) curve and

its peak (FDmax) for complexity description of neurons;


2.3.2. 2.3.2. FDFDmaxmax -- the methodthe method

Radius

Cum

mV

f1

ln(Radius)

l n( C

umm

V)

f2=ln(f1)

ln(Radius)

Fra

cD

i m

Peak of max fractality Scale of max fractality Width of max fractality

Multiscale fractal dimension

ln(Radius)

Fra

cD

i m

ln(Radius)

Fra

cD

i m

ln(Radius)

Fra

cD

i m

f3=3-d(f2)dr


2.3.2. Gray Level 2.3.2. Gray Level CoocurrenceCoocurrence MatrixMatrix

• Joint probability distribution of 2 pixels given a displacement (direction+distance) and a window;

• Spatial distribution and spatial dependence.

0 0 1 1 10 0 1 1 10 2 2 2 22 2 3 3 3

d = (1,0) & w = m x n2 0 0 02 4 0 01 0 4 0 0 0 1 2

I(m,n) = gray levels GLCM(i,j) = gray level transitions ratenumber of transitions

ji

nm


2.3.2. GLCM2.3.2. GLCM

2

1 1( , )( )

g gN N

i jInertia p i j i j

= =

= −∑∑

1 1

( , ) log( ( , ))g gN N

i j

Entropy p i j p i j= =

= −∑∑

21 1

1 ( , )1 ( )

g gN N

i j

Homogeneity p i ji j= =

=+ −∑∑

Haralick (1976)

2

1 1

( , )g gN N

i j

Energy p i j= =

=∑∑


2.3.3. Color2.3.3. Color

Pratt (1991), Song et al (1997)

b

N(b)


2.4. 2.4. LeukoLeuko classificationclassification

• Learning: supervised;• Decision rule: maximum likelihood;• Pdf estimation: parametric;• Pdf: Gaussian;• Error estimation methods regarded:

• Resubstituion, holdout, cross-validation,leave-one-out.


3. Feature selection3. Feature selection

• Problems of dealing with many features:• Overfitting (decision surface = many details)• Overtraining (decision surface = few details)• Curse of dimensionality (lacking samples);• Peaking phenomena.

• ∴ Define feature subsets using strategies:• Exaustive;• Heuristics;• Random;


3. Feature selection3. Feature selection


4. Results4. Results• Differentiation among 6 types

of leukocytes, including the CLL (the most common leukemia in adults from Western);

• Project and development of pattern recognition and feature selection tools.


5.1. 5.1. LeukoLeuko + SVM+ SVM

•• SVM extension to SVM extension to multiclassmulticlasstask, applying treetask, applying tree--based based strategy;strategy;

• SVM and minimum spanning tree:• Adapted version of Kruskal

algorithm• Generates a tree in a bottom-up

iterative way• Polynomial complexity

• Including weight assignment• Allows the automatic determination

of the multiclass tree.

1

2 5

3 4

1

0

23

45

6

7

8

9

Graph G

1

2

3 4

1

0

4 5

MST T

2


5.1. 5.1. LeukoLeuko accuracyaccuracyCNPq


• γ(t) = boundary of the nucleus in Ф(x,t=0)• Фt + F| Ф| = 0

F 1 1 G I

5.2. Level set5.2. Level set


• Given C, the set of points of the nucleus contour and p∈ C: DC(p) = min(d(p,q))

• If 0


UshizimaUshizima et al., Leukocyte detection using nucleus contour et al., Leukocyte detection using nucleus contour propagation, LNCS 4091:389propagation, LNCS 4091:389--396, 2006.396, 2006.

43/41

HematlasHematlas

44/41Leukocyte detection using nucleus contour propagation

Acknowledgments:Acknowledgments:• Unisantos: Profa. Marta Rosatelli, Daniella, Rodrigo, Aline, Rafael• NIH: Dr. Edgar Rizzatti e Dr. Rodrigo Calado• Instituto Fleury: Dr. Sérgio Martins• Universidade de Sao Paulo: Prof. André Carvalho e Ana• Universidade Federal de Santa Maria: Prof Marcos e Luciana• Universidade Federal do Ceará: Profa Fátima Medeiros

Contact:Contact:

Documents

LEUKO: Computer Decision Support System for Leukemia Diagnosis · 2012. 2. 25. · Leuko + SVM • SVM extension to multiclass task, applying tree-based strategy; • SVM and minimum