bandelet transform

MEDICALIMAGERETRIEVALUSINGBANDELET

1SK.SAJIDAPARVEEN,2G.PRATHIBHA,3B.CHANDRAMOHAN

1M.TechStudent,ECEDepartment,ANUcollegeofEngineeringandTechnology,A.P,

India,Email:[email protected]

2AssistantProfessor,ECEDepartment,ANUcollegeofEngineeringandTechnology,A.P,

India,Email:[email protected]

3Professor,ECEDepartment,BapatlaEngineeringCollege,A.P,India,Email:

[email protected]

ABSTRACT

Content based medical image retrieval systems are helpful to the radiologists in

diagnosis of breast cancer. Thispaper presents amethod for retrievingmammogram

images as cancer or normal by using bandelet transform.Bandelet transform is

appropriateforanalysisofedgesandtexturesofimagesbasedonwhichimagefeatures

are extracted. The main contribution in this paper is of three directions. First, Pre‐

processing isvery important tocorrectandadjust themammogramimage for further

studyandprocessing.Second,featureextractionisakeyissueincontentbasedmedical

imageretrieval.Bandeletcoefficientsarecalculatedfromthepre‐processedimagesand

statisticalparametersarecalculatedforthesecoefficientswhichformfeaturevectorfor

the image. Third, based on the similaritymeasurement images are retrieved through

graphical user interface and classified by using KNN. Classification accuracy and

precisionof93.737%and0.945areobtained.

KEYWORDS:contentbasedmedicalimageretrieval,featureextraction,bandelet,KNN,

GUI.

1. INTRODUCTION

In CBMIR system, the term [CBIR]

describes the process of retrieving

desired images from a large collection

onthebasisoffeatures(suchascolour,

texture and shape) that can be

automaticallyextractedfromtheimages

themselves. The features which are

SK. SAJIDAPARVEEN et al. DATE OF PUBLICATION: AUGUST 13, 2014

ISSN: 2348-4098 VOLUME 02 ISSUE 06 JULY 2014

INTERNATIONAL JOURNAL OF SCIENCE, ENGINEERING AND TECHNOLOGY- www.ijset.in 1104

extracted from an image can uniquely

identify an image from others. The

extraction of features from the image

pixels is termed as feature extraction.

Using the extracted features from the

process of feature extraction, similarity

between indexed image and query

imageismeasured.

Breast cancer is the leading type of

cancer inwomen and the secondmost

fatal type of cancer in women .Breast

cancer is amalignant tumor that starts

in the cells of the breast. A malignant

tumorisacollectionofcancercellsthat

canproduceinto(overrun)neighboring

tissuesorspreadtodistantareasofthe

body.Thediseaseoccursaroundwholly

inwomen,but in some casesmenalso.

In themedical imaging context the aim

of CBIR is to provide radiologists with

thediagnosticaidintheformofdisplay

ofrelevantpastcasesalongwithproven

pathology and other suitable

information. Medical image data have

been expanded rapidly in quantity,

content and dimension – due to an

enormous increase in the number of

diverse clinical exams performed in

digital form and to the large range of

image modalities available. It has,

therefore, led to an increased demand

for efficient medical image data

retrieval and management. In current

medical image databases, images are

mainly indexed and retrieved by

alphanumerical keywords, classified by

human experts. However, purely text‐

based retrieval methods are unable to

sufficiently describe the rich visual

propertiesorfeaturesinsidetheimages

content, and therefore pose significant

limitations on medical image data

retrieval. The ability to search by

medical image content is becoming

increasingly important, especially with

the current trend toward evidence‐

based practice of medicine. In the

clinical practice of reading and

interpreting medical images, clinicians

(i.e., radiologists) often refer to and

compare thesimilarcaseswithverified

diagnostic results in their decision

making of detecting and diagnosing

suspiciouslesionsordiseases.However,

searchingforandidentifyingthesimilar

reference cases (or images) from the

large and diverse clinical databases

(either the conventional film/paper

based libraries or advanced digital

image storage systems) is a quite

difficult task. The advance in digital

technologiesforcomputing,networking,

and database storage has enabled the

automated searching for clinically

relevant and visually similar medical




examinations (cases) to the queried

case from the large image databases.

There are two types of general

approaches in medical image retrieval

namely, the text (or semantic) based

imageretrieval(TBIR)andthecontent‐

basedimageretrieval(CBIR).Currently,

themostofavailablesearchsystems(or

tools) developed and implemented in

medical informatics and picture

archiving and communication systems

(PACS)useTBIRschemesthatarebased

ontheannotatedtextual informationto

select similar or clinically relevant

references (cases) [1‐4]. This approach

is typically limited to retrieve or select

the same type of medical images (i.e.,

mammograms or CT brain images).

However, the relevant clinical

informationdepictedonmedicalimages

is locally presented (i.e., breastmasses

depicted on mammograms

andemphysemalesionsdepictedonlung

CT images). In the clinical practice of

reading and interpreting medical

images, the nature of the queried

suspicious regions is often un‐

determined. Thus, the CBIR is the only

available and reliable approach to

retrieve the clinically relevant

(reference)casesalongwiththeproven

pathology and other related clinical

information. As a result, developing

CBIR schemes has been attracting

extensive research interest in theareas

ofmedicalinformaticsandPACSforthe

lastdecade[5‐10].Despite the fact that

CBIR approach is still in its early

development stage facing many

technical challenges (i.e., region

segmentation, semantic gap, and

computational efficiency), as the digital

medical images are produced in ever

increasing quantities and used for

diagnosis and therapy, the researchers

believe that the advance of CBIR

development will play more and more

important role in futuremedical image

diagnosis and patient treatment (or

management)[11].

Mammography is one of the effective

toolsinearlydetectionofbreastcancer.

Mammography is a low dose x‐ray

procedure for the visualization of

internal structure of breast.

Mammography has been proven the

most reliable method and it is the key

screeningtoolfortheearlydetectionof

breast cancer. Mammography is highly

accurate, but likemostmedical tests, it

is not perfect. On average,

mammography will detect about 80–

90% of the breast cancers in women

withoutsymptoms.




The common characteristics of the

medical images like as unknown noise,

poor image contrast, in homogeneity,

weak boundaries and unrelated parts

will affect the content of the medical

images. This problem rectified by pre‐

processing techniques. The pre‐

processingarefundamentalstepsinthe

medical image processing to produce

better image quality for segmentation

andfeatureextractions.

2. PROPOSEDMETHOD

THEBANDELETTRANSFORM

Bandelet transform, introduced by Le

PennecandMallatbuilt abaseadapted

to the geometric content of an image.

The bandelet are obtained from a local

deformation of space to align the

direction of regularity with a fixed

direction (horizontal or vertical) and is

reduced to a separable basis. In

bandelettransform,ageometricflowof

vectorsisdefinedtorepresenttheedges

of image. These vectors give the local

directions in which the image has

regular variations. Orthogonal bandelet

bases are constructed by dividing the

image support in regions inside which

the geometric flow is parallel. Image is

portioned into small regions, each

region includes almost one contour. If

theregiondoesnotincludeanycontour,

the imageintensity isuniformandflow

isnotdefined.Inbandelettheseregions

are approximated in separablewavelet

basisof 2L in[12]

Immj

xmj

xmj

xmj

xmj

xmj

xmj

2,

1,

22

,11,

22

,11,

22

,11,

Where I is an index set that depends

uponthegeometryoftheboundaryof ,

andx1,x2denotesthelocationofpixelin

theimage, 2,1, 21xx mjmj 2,1, 21

xx mjmj ,and

2,1, 21xx mjmj arethemodifiedwaveletsat

the boundary. If a geometric flow is

calculated in , this wavelet basis is

replaced by a bandelet orthonormal

basisof 2L in

2,1

1

1

1

,,2

2,11,

22

,11,

22

,11,

mmjljxcx

mjx

mj

xcxmj

xmj

xcxmj

xml

Whichisgotbyintersectingbandeletin

thewarpedwaveletbasisin

WImmjxcx

mjx

mj

xcxmj

xmj

xcxmj

xml

21

1

1

1

,,2

2,11,

22

,11,

22

,11,

Intheaboveexpressions,c(x)denotesa

flow line associated to a fixed

translationparameterx2, 121 , xcxx

is a set of point for x1 varying, and l is




thedirectionofgeometricflowwhichis

more elongated (2l>2j) and c(x) is

definedas

x

x

i duuexcmin

.

In the bandelet representation, the

parameters include the bandelet

coefficientsusedforcomputingandthe

parameters that specify the image

partition and the geometric flow in

square regions, which are subdivided

into four smaller squares,

corresponding to a node having four

childreninthequadtree,asshownfig1.

In order to achieve appropriate image

geometryof image f, thebestgeometry

is employed to an approximation error

Mff .

Figure1:Quadtreeofdyadicsquareimage

segmentation

In each region I of the segmentation,

one must decide if there should be a

geometricflow.Ifthisflowisparallel, c

(t) is calculated as an expansion over

translatedB‐spline functionsdilatedby

a scale factor 2lover a square of

width, the flow at a scale 2l is

characterizedby2k‐lcoefficientsn ,

lk

n

ln ntbtc

2

1

2.

The scale parameter 2l is adjusted

through a global optimization of the

geometry. When the image f has

contoursthatarecurves c whichmeet

atcornersor junctions,andthatf is c

awayfromthesecurves, thisprocedure

leads to a bandelet approximation that

has an optional asymptotic errordecay

rateR.

CMffR M

2 ,

Standard wavelet bases are optimal to

represent functions with point wise

singularities. However they fail to

capture the geometric regularity along

thesingularitiesofthesurfaces,because

of their isotropic support. For C

geometrically regular functions, the

distortion‐rate of a wavelet image

transformcodewithRbitssatisfies

RCRff R log12

The following steps for the proposed

method as shown in fig2. Firstly apply

haar transform to the preprocessed

mammogram image .secondlyto

compute thebestdirection, firstselects

thissetofdirectionsandstoresthemin

Theta,andthencomputestheLagragian

for each direction. It then chooses the




directio

smalles

constru

wavele

segmen

dyadic

Figure

3. ME

3.1IMA

TheDig

(DDSM

availab

data. I

screeni

the tot

the DD

images

normal

thiswo

on which

st Lagra

uctabande

et doma

ntation of

squaresis

e2:Blockdiag

ETHODOL

AGEDATA

gitalforScr

) [13] is

ble databa

It contains

ing mamm

tal number

DSM datab

s consisting

limagesar

ork.

Warped

Best Ge

Bande

Quadtr

gives r

ngian. T

eletbasiso

ain, a

each wave

used.

gramofprop

LOGY

ASET

reeningMa

the larg

se of ma

s approxim

mography

r of images

base, a to

g of both

rechosen,w

d Haar Transfo

ometry Selec

elet Transfor

ree Construct

rise to th

Thirdly T

onthewho

quadtre

elet scale

posedmethod

ammograph

est public

mmograph

mately 262

cases. Fro

s included

otal of 47

cancer an

wereused

form

ction

rm

tion

he

To

ole

ee

in

d

hy

cly

hic

20

om

in

78

nd

in

3

Im

n

o

re

q

im

ap

st

li

w

b

D

im

in

Figure3:Ba

us

.2PREPRO

mage pre‐

ecessary,

rientation

emove the

uality of t

mage‐proce

ppliedonm

teps are v

mit the

without

ackground

Digital ma

mages th

nterpreted,

Query ImaMammog

Feature V

Preproces

Bandeltransfo

asicblockdiag

singproposed

OCESSING

‐processing

in orde

of the

e noise and

the image

essing alg

mammogra

very impor

search fo

undue

dofthemam

ammogram

hat are

,thusapre

age of gram

Vector

ssing

let rm

KNN classif

SimilarMeasure

Retrieved ithrough

gramofcbirs

dmethod

g techniqu

r to fin

mammogr

d to enhan

[14]. Befo

gorithm c

am,prepro

rtant in o

or abnorm

influence

mmogram.

ms are m

difficult

eparationp

Database Iof Mammo

Bandeltransfor

Preproces

fication

Feature V

rity ement

images GUI

system

ues are

nd the

ram, to

nce the

ore any

can be

ocessing

rder to

malities

from

medical

to be

phaseis

Image ogram

let rm

ssing

Vector




needed

quality

results

objectiv

thequa

to furt

reducin

parts

mamm

medica

interpr

essenti

noise a

remove

success

region

Figu

Image

(clarity

Elimina

noise,

enlight

augmen

an im

endoth

low co

the no

d in order

y and ma

more

ve of this

alityofthe

ther proce

ng the un

in the

ogramima

al images

ret. Henc

ial to impr

and high f

ed by fi

sive steps

extraction.

ure4:a:origin

image

enhancem

y) of image

ating blur

increasi

tening det

ntation op

mage migh

helial cell a

ontrast and

ise and bl

to improv

ake the s

accurate.

process is

imagetom

ssing by r

nrelated a

backgroun

ages.Mamm

that com

e pre‐pro

rove the q

frequency

ilters. The

take place

.

nal b.pre

i

ent improv

es for hum

rring of i

ing cont

tails are e

erations. F

ht be in

and image

d blurred.

lurring and

ve the imag

egmentatio

The ma

s to improv

makeitread

removing

and surplu

nd of th

mogramsa

mplicated

ocessing

quality. Th

componen

e followin

e for brea

eprocessed

image

ves the cla

man viewin

images an

trast, an

examples

For exampl

use of a

might be

Plummetin

d increasin

ge

on

ain

ve

dy

or

us

he

re

to

is

he

nts

ng

ast

ass

ng.

nd

nd

of

le,

an

of

ng

ng

th

im

th

b

u

st

of

is

ca

p

in

re

if

th

re

of

b

la

im

b

im

h

ar

re

an

sh

3

A

In

of

al

an

he contras

mage. In s

hresholding

inarization

sing Otsu’s

tep permit

fthebreas

s divided

alculate th

art:ifthen

n left par

egionisfro

f the numb

he right p

egion is fro

f the b

ackground

abelling alg

mage into

reastregio

magegener

asthelarg

rea criteri

epresentst

nd to elim

howninfig

.3 k‐

ALGORITHM

npatternr

f the mos

lgorithms

ndregressi

st range co

second ste

g method

n of the

s method i

ts to identi

tregion.F

into two

he number

numberof

rt, the dir

omlefttor

ber of whit

art, the di

om right to

breast re

d the conn

gorithm to

different la

onselection

rated,note

estarea.Fo

ia to selec

thebreast

minate the

g4.

‐NN

M

recognition

t importan

and used

ion.In both

ould enhan

ep, an aut

d to ob

enhanced

is used. Th

ify the orie

Forthat,th

equal par

of pixels

whitepixe

rection of

right(respe

te pixels is

irection of

o left). Sep

gion from

nected com

divide the

abels is us

nbylookin

thatbreast

orthisreas

ct the lab

regionisc

unlikely la

CLASSIFIC

n field,KNN

nt non‐par

d forclassi

h cases, th

nce the

tomated

tain a

image

he third

entation

eimage

rts and

of each

lsisbig

breast

ectively,

s big in

f breast

paration

m the

mponent

ebinary

sed. For

ngatthe

tregion

son,the

bel that

consider

abels as

CATION

Nisone

rameter

ification

he input




consists of thekclosest training

examples in thefeature space. The

output depends on whetherk‐NN is

usedforclassificationorregression:

Ink‐NNclassification, theoutput

is a class membership. An object is

classified by a majority vote of its

neighbors, with the object being

assigned to the class most common

among itsknearest neighbors (kis a

positiveinteger,typicallysmall).Ifk=1,

thentheobjectissimplyassignedtothe

classofthatsinglenearestneighbor.

Ink‐NNregression, theoutput is

the property value for the object. This

value is the average of the values of

itsknearestneighbors.

k‐NN is a type ofinstance‐based

learning, orlazy learning, where the

function is only approximated locally

and all computation is deferred until

classification. Thek‐NN algorithm is

among the simplest of allmachine

learningalgorithms.

Bothforclassificationandregression,it

can be useful to weight the

contributions of the neighbors, so that

thenearerneighborscontributemoreto

theaveragethanthemoredistantones.

For example, a common weighting

schemeconsistsingivingeachneighbor

aweightof1/d,wheredisthedistance

totheneighbor.

The neighbors are taken from a set of

objects for which the class (fork‐NN

classification) or the object property

value (fork‐NN regression) is known.

Thiscanbethoughtofasthetrainingset

for the algorithm, though no explicit

training step is required. It is a

supervised learning algorithm. The

classificationrulesaregeneratedbythe

training samples themselves without

any additional data. The KNN

classificationalgorithmpredictsthetest

sample’s category according to the K

training sampleswhich are the nearest

neighbors to the testsample,and judge

ittothatcategorywhichhasthelargest

category probability. The process of

KNNalgorithmtoclassifysampleXis:

Suppose there are j training

categoriesC1,C2,…,Cjandthesumof the

training samples is N after feature

reduction, they become m‐dimension

featurevector.

Make sample X to be the same

featurevectorof theform(X1,X2,…,Xm),

asalltrainingsamples.

Calculate the similarities

between all training samples and X.

Takingtheithsampledi(di1,di2,…,dim)as




an example, the similarity SIM(X, di) is

asfollowing:

2

1

2

1

1jijj dX

di) SIM(X,

m

jij

m

jj

m

dX

Choose k samples which are

larger fromN similarities of SIM(X, di),

(i=1, 2,…, N), and treat them as a KNN

collection of X. Then, calculate the

probabilityofXbelongtoeachcategory

respectivelywiththefollowingformula.

jid

ij CdydXSIMCXP ,,,

Where y(di, Cj) is a category attribute

function,whichsatisfied

y(di,Cj)=

ji

j

Cd

C

,0

d 1, i

Judge sample X to be the

categorywhichhasthelargestP(X,Cj).

3.4SIMILARITYMEASUREMENT

Manhattandistance

Manhattandistanceisgivenby

||1

ii

n

i

yx

The minimum distance value signifies

an exact match with the query.

Manhattan distance is not always the

bestmetric. The fact that the distances

ineachdimensionaremodulatedbefore

summation, places great emphasis on

those features for which the

dissimilarity is large. Hence it is

necessary to normalize the individual

feature components before finding the

distancebetweentwoimages.

3.5GUI‐GraphicalUserInterface

Incomputing, agraphical user

interfaceis a type of interfacethat

allowsuserstointeract with electronic

devicesthrough graphicaliconsand

visual indicators such assecondary

notation, as opposed totext‐based

interfaces,typedcommandlabelsortext

navigation. The GUI is as simple and

intuitiveaspossiblesothatuserdoesn’t

need to spend much time in learning

howtouseit.Itallowsapersontowork

easilywithacomputerbyusingamouse

to point to small pictures and other

elementsonthescreen.Theactionsina

GUI are usually performed

throughdirect manipulationof the

graphical elements. As well as

computers, GUIs can be found inhand‐

held devicessuch asMP3players,

portablemediaplayers,gamingdevices

and smaller household, office and

industryequipment.




Figure5:screenshotsofthemainwindow

Figure5showsscreenshotsofthemain

window.Themainwindowcomposedof

select query image (lower left), no of

returned images (middle), output

(lower right) and retrieved image

(upper).

4. EVALUATION OF THE

RETRIEVALSYSTEM

Performance measures are based on

precisionandrecall.Theyaredefinedas

follows:

Precision=retrieved items all of No.

retrieved itemsrelevant of No.

Recall=retrievedrelevant all of No.

retrieved itemsrelevant of No.

5. RESULTS

The images for the proposed research

were collected from the online digital

base for screening mammography. The

images are preprocessed and

segmented.Featuresareextractedfrom

images using bandelet transform. To

obtain the retrieved images through

GUI,firstlyrunguiapplicationandthen

select the query image from folder.

Secondly select no of returned images

which returns no of retrieved images

and thirdly click output to display the

retrievedimagesasshowninfig6 and

mean while returns classification

accuracy and precision using KNN

classification,alsodisplaythepredictto

which class the image belongs in

commandwindow.

Figure6:RetrievedimagesthroughGUI

6. CONCLUSION

Inthispaper,anewmethodisproposed

for feature extraction which will be

helpful to the radiologist to predict

whether the image is cancer or normal

at early stage. By using this proposed

method the classification accuracy

acquired is 93.709% and precision is

0.945 for ddsm database with KNN

classification algorithm and image




retrieved through graphical user

interface(GUI).

ACKNOWLEDGEMENTS

I take this opportunity to express my

profoundgratitudeanddeepregardsto

G. Prathibha, Assistant professor, ECE

Department, Acharya Nagarjuna

University college of Engineering &

Technology, AP, India for her valuable

guidance and appreciation throughout

thework.

REFERENCES

[1]. Love,H.J.;Antipow,I.;Hersh,W.;

Smith, C.A.; Mailhot, M. Automated

semanticindexingofimagingreportsto

support retrieval of medical images in

the multimedia electronic medical

record.MethInformMed1999,38,303‐

307.

[2]. El‐Kwae,E.; Xu,H.;Kabuka,M.R.

Content‐based retrieval in picture

archiving andcommunication systems. J

DigitImaging2000,13,70‐81.

[3]. Ogiela, M.R.; Tadeusiewicz, R.

Semantic‐oriented syntactic algorithms

for content recognitionand

understanding of images in medical

database. In Proceedings of the second

InternationalConference on Multimedia

andExposition, IEEEComputerSociety,

Tokyo,Japan,2001,621‐624.

[4]. Hersh, W.; Mailhot, M.; Arnott‐

Smith, C.; Lowe,H. Selective automated

indexing of findings and diagnoses in

radiologyreports. JBiomed Informatics

2001,34,262‐273.

[5]. Tagare, H.D.; Jaffe, C.; Duncan, J.

Medical image databases: a content‐

based retrieval approach. J. Am. Med.

InformaticsAssoc.1997,4,184‐198.

[6]. Lehmann, T.M.; Guld, M.O.;

Deselaers, T.; Keysers, D.; Schubert, H.;

Spitzer,K.;Ney,H.;Wein,B.B.Automatic

categorization of medical images for

content‐basedretrievalanddatamining

Comput Med Imaging Graph 2005, 29,

143‐155.

[7]. Long, L.R.; Antani, S.K.; Thoma,

G.R. Image informatics at a national

research center. Comput Med Imaging

Graph2005,29,171‐193.

[8]. Muller, H.; Rosset, A.; Garcia, A.;

Vallie, J.; Geissbuhler, A. Benefits of

content‐based visual data access in

radiology.RadioGraphics2005, 25, 849‐

858.

[9]. Lam,M.O.;Disney,T.;Raicu,D.S.;

Furst, J.;Channin,D.S.BRISC–Anopen

source pulmonary nodule image




retrieval framework. J Digit Imaging

2007,20,63‐71.

[10]. Pourghassem,H.;Ghassemian,H.

Content‐based medical image

classification using a new hierarchical

merging scheme. Comput Med Imaging

Graph2008,32,651‐661.

[11]. Muller, H.; Michoux, N.; Bandon,

D.;Geissbuhler,A.A reviewof content‐

based image retrieval systems in

medical applications – clinical benefits

and future directions.Int J Med Inform

2004,73,1‐23.

[12]. A novel image fusion algorithm

based on bandelets transform by

XiaoboQu,JingwenYan,Guofuxie,Ziqian

Zhu, and Bengang Chen in october

10,2007/vol. 5,no 10/chinese optics

letters.

[13].http://marathon.csee.usf.edu/Mam

mography/Database.html.

[14]. Samir Kumar Bandyopadhyay,

pre‐processing of MammogramImages,

International Journal of Engineering

Science and Technology, Vol. 2(11),

2010,6753‐6758.

BIOGRAPHIES

Sk. SajidaParveen received B. Tech

degree in electronics and

communication engineering from

vignan’slara institute of technology and

science,Gunturin2012.Sheiscurrently

pursuingM.Techincommunicationand

signalprocessinginuniversitycollegeof

engineering and technology, Acharya

Nagarjunauniversity.

Er. G. Prathibha is working as an

Assistantprofessor inuniversitycollege

of engineering and technology, Acharya

Nagarjuna university. She received her

B.Tech (ECE) degree from R.V.R & J.C

college of engineering, Guntur, M. Tech

(system and signal processing) from

JNTU, Hyderabad. Her areas of interest

are medical image processing, video

processingandpatternrecognition.

Dr. B. Chandra Mohan is presently

working as Professor in Bapatla

Engineering College, Bapatla. His areas

of interest are communication, signal

processingandimageprocessing.




bandelet transform

Documents

currentmedical image

medical image data havebeen

extracted features

secondmostfatal type

similaritymeasurement

theextraction of features

textbased retrieval

classified byhuman experts