Mahua Mam Iiitm Gwalior.1

download Mahua Mam Iiitm Gwalior.1

of 37

Transcript of Mahua Mam Iiitm Gwalior.1

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    1/37

    Discrimination forMalignant and Benign

    Masses in Breast UsingMammogram:

    A Study on

    Adaptive Neuro-FuzzyApproaches  Mahua Bhattacharya 

    AB !ndian !nstitute of !nformation

    "echnology # Management$

    %&alior

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    2/37

    2

    'rigin of "he (or) 

    X-Ray mammography is the most important modality

    which is used in early detection of breast cancer. Women

    have better chance to survive if breast cancer is detected

    early. The failure to detect any abnormal lesion at an early

    stage may lead to disastrous consequences, so the

    improvement of mammographic image quality is essential

    for breast cancer screening. Therefore a need exists to

    automate the process of analyzing a large number of

    mammograms and to discriminate the benign lesion frommalignant one for proper therapy planning.

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    3/37

    3

    'rigin of "he(or)*cont++,

    The present work is continuation of earlier work based on  theory of shape related to gradation of benignancy of tumor in tissue region. Theconcept of symmetry analysis of shape computing a distancefunction D  between the contour of the tumor model and thepattern tumor lesion was utilized to classify the tumors mainly intwo broad categories benign and malignant transformations.

    References

    M. Bhattacharya, D. Dutta Majumder, “Knowledge Based Approach to Medical Image Processing” inPattern Directed Information Analysis(Algorithms, Architecture & Applications, publisher : NewAge Wiely, 2007. (in press)

    M.Bhattacharya and D.Dutta Majumder, “Breast Cancer Screening UsingMammographic Image Analysis' Sixteen International CODATA (France) Conference (8-12 Nov ,Delhi (1998).

    M.Bhattacharya, “Development of Mathematical Model for Radiographic Image Analysis (published,!!"# International $ournal of %omputational and &umerical Analysis and Applications, Academic'ublication ( !!"#.

    D.Dutta Maumder ) Mahua Bhattacharya, “Cybernetic Approach To Medical Technology: Application To Cancer Screening And Other Diagnostics”, Millennium *olume of +ybernetes,International $ournal of ystems ) %ybernetes, M%B publications -+, *ol. , number "/0 , pp10"2304, !!! (5ran6 7eorge Research a8ard 9inning 'aper #.

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    4/37

    4

    (hat is Mass 

    Masses are three-dimensional lesions which may

    represent a localizing sign of Breast Cancer .

    They are described by their Size, Shape, Margincharacteristics, X-ray attenuation, Effects on

    surrounding tissues etc.

    Depending on the Morphologic  criteria of themass, the lielihood of Benignancy/ Malignancy 

    can be established.

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    5/37

    !

    Masses *cont++,

     " mass shape may ha#e one of fi#e characteristics$ Round,

    Oval, o!ulated, "odular   and Stellate. The descriptions

    are fairly self-e%planatory, and a schematic picture of each

    shape is shown below.

     "long with &hape, Margin and &ize of the masses are also

    important indicators to determine Benignancy/ Malignancy# 

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    6/37

    '

    './ective of "he (or) 

    (b)ecti#e of present wor is de#elop a soft-co$puting !ased "euro %uzzy &y!rid Model which will be robust enough to tae decision

    whether the lesion is belonging to either benigngroup or in malignant group.

    *urther ob)ecti#e is not only to predict the+enignancy Malignancy of lesion but to predict

    tendency of growth of the lesion either towardsbenign or malignant i.e. to de#elop a GradationTechnique for prognosis of disease for moreaccurate therapeutic planning .

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    7/37

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    8/37

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    9/37

    /

    Preprocessing of Acquired Images

    0oises appear in the images as abrupt change of grayle#els, thus a#erage of the pi%els contained in theneighborhood of the filter mas is the easiest method fornoise remo#al. These filters are sometimes called a#eragingfilters.

    1n this study 33 smoothing filters used to remo#e irrele#antdetails. The structure of this filter shown below.

    +y replacing the #alue of

    e#ry pi%els of image by thea#erage of the mas reduces

    sharp transitions in gray

    le#el.

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    10/37

    Segmentation of Masses

    The presence of high-contrast fibroglandular tissues in themammograms may confuse the radiologists as the actualcalcified masses.

    1n cases where the breast may be dense, the margins

    may appear obscured because both benign and malignantlesions de#elop in lobules which are hea#ily surroundedby normal e%tralobular connecti#e tissue. Thus properseg$entation  of masses in the dense breastmammograms is a difficult, important and challenging

    tas.

    The final classification on degree of malignancy of themasses also depends on the superiority of thesegmentation techni5ue.

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    11/37

    Segmentation of Masses*cont++,

    This mass is diagnosed as being Malignant, if not properly

    segmented. Thus segmentation of mass boundary plays an

    important role in discrimination.

    IICAI-07 

      Due to the superimposition of three-dimensional  tissue structures, the margins of a mass may be

      obscured and looed ill-defined. " mass

      with almost o#al shaped circumscribed margin

      obscured by the surrounding tissues and its  proper segmented contour shown below.

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    12/37

    2

    Segmentation of Masses .yFuzzy c-Means 1lustering

    "echni2ue *uzzy 6-means clustering algorithm used forintensity based segmentation of Masses.

    6luster center A  represents the healthy breast tissue. 

    6luster center B  represents false presence of masses.

    6luster center C   represents actual mass region. 

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    13/37

    3

    1nitialize the membership matri% " with random #alues between and as shown in e578.

      78

    6alculate c fuzzy cluster centers using e5728.

      728 

    Fuzzy 1-Mean 1lusteringAlgorithm

    2#(2

    =∑=

     x A   k C 

    ii

    =

    ==n

    m

     N 

    k k 

    m

    i

     x A

     x x A

    k i

    k i

    2

    2

    #:(;

    #:(;

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    14/37

    4

    Fuzzy 1-Mean 1lusteringAlgorithm *1ont++, 6ompute the cost function according to 95 738. &top if its

    impro#ement o#er pre#ious iteration is below a threshold.

     

    738

    6ompute a new membership matri% A using 95 748.

      748

     

    ∑=

    =

      !

    m

    k i

    d d 

     x A

      !k 

    ik 

    2

    #2(

    #(

    2#(

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    15/37

    !

    Superiority of F1M over

    1anny "echni2ue

      (a# (b# (c# (d#

    7a8 (riginal Mammogram, 7b8 6utout from suspicious region,

    7c8 &egmentation using *6M, 7d8 &egmentation using 6anny.

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    16/37

    '

    Boundary Feature3epresentation

    The degree of correct classification strongly depends on

    6lassic *eature &election Method.

    Thus to describe the boundary #ery precisely, we apply

    %ourier 'escriptor  method in our proposed algorithm. 1ncase of abrupt change in boundaries, *ourier descriptors

    contains high fre(uency  components  7a stellate mass

    contains #ery high fre5uency components8. 

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    17/37

    Fourier Descriptors 'f "heBoundary

     6onsider the boundary consists of -points in the %-y

    plane staring at an arbitrary point 7%,y

    8 to the

      coordinate pairs 7%,y

    8, 7%

    2,y

    28,:..7%

    -,y

    -8 .

      9ach co-ordinate pair can be treated as a comple%

      number so that 

    s)*+ )*+ . 0 y)*+

    for ;, , 2, :.., − .

     The greatest ad#antage of this type of representation

    is

      that it reduces the 1-' pro!le$ into 2-' pro!le$.IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    18/37

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    19/37

    /

    "e4tural Feature3epresentation

    Te%ture information plays an important role in imageanalysis and understanding. 1t measures the RelativeS$oothness, 6verage 7nifor$ity, 8ray level 9aria!ility of the images.

    (nly boundary of the masses sometimes representmisleading information 7due to limitations of filmmammography, a circumscribed mass appeared asmalignant one8 thus blending of te%ture features withboundary impro#e the classification results.

    1t is noted that, a benign mass posses S$ooth  and7nifor$ te%ture, it is an indicator to determine the degreeof $alignancy/!enignancy.

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    20/37

    2

    "e4tural Feature3epresentation *cont++,

    Te%tural *eatures are described on the basis of statistical

    properties of the gray le#el image.

    1f z be a random #ariable denoting the gray le#els in the range

    of 7, > − 8.

    The nth moment of z about the mean is

    where m is the mean #alue of z,

    ∑   −−

    =

    =2

    !#(#(#(

     %

     & n   '  pm '    &  & 

    n '  µ 

    ∑   ×−

    =

    =2

    !#(

     %

     &  ' p 'm  &  & 

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    21/37

    2

    "e4tural Feature3epresentation *cont++,

    1nd moment 7n18, the 9ariance µ1 )z+ is a measure of relati#e

    smoothness of the gray le#el contrast.

    :rd moment 7n:8, measure of S*e;ness  of the histogram, is

    useful for determining the degree of symmetry of histograms.

     "nother useful te%ture measurement is based on histograms?7nifor$ity@778 gi#en by

     "#erage Entropy  is a measure of gray-le#el #ariability. 1t isdefined as

     

    #(2

    !

     '  p   &  %

     & 

    (  ∑−

    =

    =

    #(#( log

    2

    ! '  '    &  & 

     %

     & 

     p pe ∑−

    =

    =

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    22/37

    22

    1lassi5cation of Features 

    Ae ha#e introduced a robust ?6dapti#e "euro %uzzy

    Model for classification of features into benign and

    malignant stages. 

    1t is an 5nnovative Soft Co$puting approaches that

    tacles the uncertainties present in the system. "s a

    result the decision maing by the e%pert system is

    more close to reality.

    &y!rid-earning  rule has been used to train the

    classifier.

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    23/37

    23

    6y.rid 7earning 3ule: 

    Bybrid leaning rule combines Steepest 'ecent 

    method and east-S(uares Esti$ator   for fast

    identification of parameters.

      &ummary of the rule gi#en below  < < %or;ard pass < Bac*;ard pass

    =re$isepara$eters

    %ied 8radient descent

      Conse(uentpara$eters

      east s(uaresesti$ator 

    %ied

    Signals "ode outputs Error signals

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    24/37

    24

    Advantages of 6y.rid7earning

    Hybrid leaning rule combines steepest decent methodand least-squares estimator for fast identification ofparameters in adaptive Neuro-Fuzzy model.

    The hybrid method converges much faster than asingle Neural or Fuzzy approach since it reduces thesearch space dimensions of the original pure backpropagation learning.

    The hybridization of Neuro-fuzzy approaches isrobust, adaptive and handles uncertainties present inthe features much better than a conventional neuralnetwork.

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    25/37

    2!

    Searching of Feature

    su.set using F1M 

    +oundary detection of masses based on %ourier

    'escription  method, implies a large number of

    feature #ectors. 

    Ae introduce %uzzy C-$eans Clustering  techni5ue

    7%CM8 to reduce the shape descriptors into four  

    clusters only. 

    The ob)ecti#e of *6M is that degree of association isstrong for descriptors within the same cluster and

    wea for the descriptors in different clusters.

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    26/37

    2'

    Decision Ma)ing 7ogic

    Ae ha#e defined a distance function 7µ8 such thatµ1 = D1  O1  

    µ2 ; C7'2  O2837'2  O28.!

    ; C7D − (8 2 .! 

    1t determines the de#iation from roundness of themasses 7+enign &tage8. The degree of malignancy ishigher for higher #alue of µ.

    Ae ha#e set the Decision Eule as,

     1f µ2  F; 2, Decision ?Benign>

     1f 2 F; µ2F; 4, Decision ?3endency to;ards Malignancy@

     1f µ2  G 4, Decision ?Malignant@

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    27/37

    2

    Decision Ma)ing 7ogic*cont++,

    Ae ha#e defined another distance function 7µ28 asµ1 ; '1  O1  

    µ1 ; C7'1  O1837'1  O18.!

    ; C7D2 − (28 2 .!

    1t determines the de#iation of smoothness of themasses 7from +enign &tage8. The degree ofmalignancy is higher for higher #alue of µ2.

    Ae ha#e set the Decision Eule as,1fµ1  F; !, Decision ?S$ooth>

    1f !F;µ1F;3, Decision ?3endency to;ards Roughness@

    1fµ1  G 3, Decision ?Coarse@

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    28/37

    2

    Decision Ma)ing 7ogic*1ontd++,

      The *inal Decision on Degree of Malignancy is gi#en below$

    +enign H &mooth

    Benign Stage

    +enign H Eough

    +enign H 6oarse

    Tendency towards Malignant H &mooth

    Tendency towards Malignant H Eough 3endency 3o;ards Malignant Stage 

    Malignant H &mooth

    Tendency towards Malignant H 6oarse

    Malignant H Eough =ossi!ly in Malignant Stage

    Malignant H 6oarse

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    29/37

    2/

    84perimental 3esults

    Ae ha#e applied the proposed algorithm to

    databases consisting of 1?? images.

    The classifier was first trained with ob#ious Benign 

    Masses as identified by the e%pert radiologists.

    The non-ob#ious cases ha#e been tested and

    classified during the e%periment.

    *ew of the non-ob#ious case studies and final

    decision on Benignancy/ Malignancy  are gi#en in

    the ne%t slides$

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    30/37

    3

    84perimental 3esults*cont++,

    'ecision$

    Tendency towards

    Malignant &tage

    µ;49.9, µ2;13.48

    'ecision$ Iossibly in

    Malignant &tage

     µ;47.42, µ

    2;29.33

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    31/37

    3

    84perimental 3esults *cont++,

    'ecision$ Iossibly

    in Malignant &tage

    µ;49.58, µ2;28.48 

    'ecision$ Tendency

    towards Malignant

    &tage

    µ;48.72, µ

    2;10.96 

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    32/37

    32

    84perimental 3esults *cont++,

    'ecision$

    Tendency towards

    Malignant &tageµ

    ; 2.59, µ

    2;32.38

    'ecision$ Iossibly in

    +enign &tage

    µ; 18.45, µ

    2;17.86

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    33/37

    33

    1onclusions

    1n proposed methodology, we ha#e attempted tode#elop a techni5ue based on "dapti#e 0euro *uzzymodel by e%tracting the boundary of the lesion orregion of interest E(1 and also te%ture features.

    The proposed classifier has been trained by hybrid-learning rule with ma%imum fifty epochs. The learningprocess is continued till the performance goal

    reaches. The output node #alues indicate the De#iation or

    Distance function of the test masses with respect tothe trained benign masses.

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    34/37

    34

    1onclusions *cont++,

    1t is also noted that there is no sharp boundary

    among the three stages −  Benign, 3endencyto;ards Malignancy and Malignant Stage#

    The partition among the stages is also %uzzy

    =artition#

    Thus when we tae a decision that a mammographic

    masses belongs to the ?Benign stage@, there is a

    little chance that it actually belongs to ?3endency

    to;ards Malignant Stage@ and vice versa.

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    35/37

    3!

    Ac)no&ledgment 

    Ae would lie to than to 'r# S# @# Shar$a 

    of 9=( J-ray and 1maging 1nstitute, =olatafor pro#iding radiological e%pertise.

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    36/37

    3'

    3eferences

    1. L. M. Bruce and R. R. Adhami, “Classifying Mammographic Mass Shapes Using theWavelet Transform Modulus-Maxima Method ”, IEEE Trans. Medical Imaging, vol. 18, no.

    12, pp. 1170-1177, (1999).

    2. J. K. Kim and H. W. Park, “Statistical Textural Features for Detection of

     Microcalcifications in Digitized Mammograms”, IEEE Trans. Medical Imaging, vol. 18,

    no. 3, pp. 231-238, (1999).

    3. M.Bhattacharya and D.Dutta Majumder, “ Breast Cancer Screening Using Mammographic

     Image Analysis”, 16th International CODATA (France) Conference, (8-12) Nov. Delhi

    (1998).

    4. M. Bhattacharya and A. Das, “Fuzzy logic Based Segmentation of Microcalcification in

     Breast Using Digital Mammograms Considering Multiresolution”, Proc. of International

    Machine Vision and Image Processing Conference (IMVIP-07), in NUI Maynooth, Co,

    Kildarem Ireland, UK by IEEE Computer Society, pp. 98-105, 5 -7 Sept, (2007).

    5. Mahua Bhattacharya, “A Computer-Assisted Diagnostic Procedure For Digital

    Mammograms Using Adaptive Neuro Fuzzy Soft Computing” accepted  IEEE Nuclear

    Science Symposium, Medical Imaging Conference, San Diego, California USA , 29 th Oct

    -4 th Nov , (2006).

    IICAI-07 

  • 8/18/2019 Mahua Mam Iiitm Gwalior.1

    37/37

    3

    "han) 9ou