klasifikasi gambar medis

8/18/2019 Klasifikasi Gambar Medis

1/23

PRESENTED BY : GROUP 3Dina Kusuma Wahyuni/140653!16D"ssy D#i Pu$n%m% /140653666E&&y Su'am(% / 140653!41

)ED*+,- *),GE

+-,SS*.*+,T*ON, S*)U-,T*ON .OR )R* *),GE"$u Sam S"(iai /

1406533-a'smi 2u#i(a /14065311Ris'a Su$yani /


2/23

OUTLINES

Background

Image Clustering

Image Segmentation

Image Thresholding

Image Denoising

Simulation for MRI Image


3/23

BC!"ROUND

In general, digital image processing covers four major sections : Image formation,

visualization, analysis, and management. Image enhancement algorithm can be

use as pre and post processing in all sections.

Modules of Image Processing


4/23

BC!"ROUND

Medical Image : MRI


5/23

IM"E CLUSTERIN"

• Clustering is a method #hih $%us &a(a in(% 7us("$s# $here o%&ects$ithin each clusters ha'e hih &"$"" %8 simi7a$i(y# %ut are dissimilar tothe o%&ects in other clusters(

• Clustering is a method of grou)ing data o%&ects into di*erent grou)s# suchthat similar data o%&ects %elong to the same cluster and dissimilar datao%&ects to di*erent clusters

• Clustering in'ol'es &i9i&in a s"( %8 &a(a %in(s in(% n%n%9"$7ain$%us or clusters of )oints $here )oints in a cluster are +more similar, to

one another than the )oints )resent in other clusters• Clustering of images is done %n (h" asis %8 (h" in($a;7ass simi7a$i(y(

• Target or close images can %e retrie'ed a little faster if it is clustered in a rightmanner

• Data )oints in each cluster are calculated $ith a data )oints in the cluster#similar data )oints are %rought in one cluster( So# each data )oints e-hi%its

same characteristics )resent in one cluster(• good clustering method $ould "


6/23

IM"E CLUSTERIN" / K;)"ans ,7%$i(hm

• !0Means algorithm is the most )o)ular)artitioning %ased clustering techni1ue $hich

generates the nono'erla))ing clusters 234• It is an unsu"$9is"& a7%$i(hm $hich is

used in clustering(

• The use of !means $ould %e to cluster theentities in an image %ased on each )i-el5sfeatures# normall. their %7%$ an& %si(i%n(

• Each cluster has a "n($%i& $hich is used tore)resent the general features of the cluster#it %asicall. chooses an. random centroid andassigns data )oints to the centroid %.com)aring distance of the data )oints $iththe centroid

• The data )oints $hich has 7"as( &is(an"$ith the centroid are made to form onecluster(

• It chooses the centroid smartl. and itcom)ares centroid $ith the data )ointsas"& %n (h"i$ in("nsi(y an&ha$a("$is(is and 6nds the distance

• Ne$ 7k5 centroids are calculated and thus k0clusters are formed % 6ndin out the data


7/23

IM"E CLUSTERIN" / K;)"ans ,7%$i(hm


8/23

IM"E SE"MENTTION

• Segmentation is the )rocess of )artitioning a digital image into multi)le non0

o'erla))ing regions 8sets of )i-els# also kno$n as su)er )i-els9(• The goal of segmentation is to sim)lif. and:or change the re)resentation of an

image into something that is more meaningful and easier to anal.;e(

• Image segmentation is t.)icall. used to locate o%&ects and %oundaries 8lines#cur'es# etc(9 in images( It is a )rocess of assigning a la%el to e'er. )i-el in animage such that )i-els $ith the same la%el share certain 'isual characteristics(

• Com)lete segmentation 0 set of dis&oint regions uni1uel. corres)onding $itho%&ects in the in)ut image( Coo)eration $ith higher )rocessing le'els $hichuse s)eci6c kno$ledge of the )ro%lem domain is necessar.(

•


9/23


10/23

IM"E SE"MENTTION / S%"7 E&" D"("(i%n

• ?hen g- and g. are large 8%ig changes in the 'ertical and hori;ontal

orientation res)ecti'el.9# the gradient magnitude $ill also %e large( It is thegradient magnitudes that are 6nall. )lotted

• @or e-am)le $e ha'e )i-els arrangement around )i-el 8-#.9 as /

• "radient for So%el o)erator are as follo$ /

8aA ca a9 8aF caG aH9

8aF ca3 aA9 8aH ca a9

• The hori;ontal and 'ertical kernel $ill %e /

•

+ is a constant $ith 'alue A

7


11/23

IM"E SE"MENTTION / S%"7 E&" D"("(i%n

• E-am)le of image after run through So%el edge detection /

Original Image Result


12/23

IM"E TJRESJOLDIN"

• Thresholding is used to "


13/23

IM"E DENOISIN"

• Image denoising is a )rocess for $"m%9in %8 n%is" 8$%m (h"

ima"(• ll denoising algorithm are as"& %n N%is" )%&"7 /

v(i) = u(i) + n(i) ;iϵI

v(i): observed value,u(i): true value,n(i): noise value

• ll the denoising algorithms are achie'ed %. a'eraging( The most common t.)es are/o S)atial domain 6lter

"aussian 6ltering nisotro)ic 6ltering 8@9 Neigh%oring 6ltering

Total Kariation minimi;ationo Non0Local0Means 8NL0means9 algorithm


14/23

IM"E DENOISIN" / O(su )"(h%&

• Based on a 'er. sim)le idea/ @ind the threshold that minimizes the

weighted within-class variance.• This turns out to %e the same as maimizing the between-class

variance.• O)erates directl. on the gra. le'el histogram 2e.g. !"# numbers, $(i)%,

so it&s fast 8once the histogram is com)uted9(• Jistogram 8and the image9 are bimodal.

• No use of s'atial coherence, nor an other notion of o%&ect structure.• ssumes stationar. statistics# %ut can %e modi6ed to %e locall.

ada)ti'e( 8e-ercises9• ssumes uniform illumination 8im)licitl.9# so the %imodal %rightness

%eha'ior arises from o%&ect a))earance di*erences onl.(


15/23


• No$# $e could actuall. sto)here( ll $e need to do is &ustrun through the full range of tvalues *,!"#% and 'ic the'alue that minimi>"s (t)

• But the relationshi) %et$eenthe $ithin0class and %et$een0class 'ariances can %ee-)loited to generate arecursion relation that )ermitsa much faster calculation(

• @or an. gi'en threshold# thetotal 'ariance is the sum of the$ithin0class 'ariances8$eighted9 and the betweenclass variance, $hich is the

sum of $eighted s1uareddistances %et$een the class

TOT,- ?,R*,N+E


16/23


• Since the total is constant and inde)endent of t, the eect o

changing the threshold is merel. to mo'e the contri%utions of thet$o terms %ack and forth(• So# minimizing the within-class variance is the same as maimizing

the between-class variance.• ?e can com)ute the 1uantities in (t) recursivel as we run through

the range o t values.


17/23


• E-am)le of Otsu Method for Image Denoising /


18/23


• Otsus method is aimed in An&in (h" %(ima7 9a7u" 8%$ (h"

7%a7 (h$"sh%7&(• It is %ased on the in("$7ass 9a$ian" maa(i%n( ?ellthresholded classes ha'e $ell discriminated intensit. 'alues(

• M N image histogram/o L intensit. le'el 2F#((# L034o ni )i-el of intensit. i

• Normali;ed histogram/

• Using k# F P k P L Q 3# as threshold# T 7 k/o T$o classes / C3 8)i-el in 2F#k49 and CA 8)i-el in 2k3#L0349o 7


19/23


• m3 mean intensit. of the )i-els in C3 /

• mA# # mean intensit. of the )i-els in CA/

• Mean of glo%al intensit. m" /


20/23


• m# mean intensit. u) to the k le'el /

• Jence /

• The glo%al 'ariance /

• The %et$een0class 'ariance /


21/23


• The goodness of the coice of T 7 k can %e estimated %. the ratio /

• The 1uantities re1uired for the com)utation of # can %e o%tainedfrom the histogram( Jence# for each 'alue of k# 8k9 can %e

com)uted /

$here /

• The o)timal threshold 'alue k satisf. /


22/23

SIMULTION/ )R* *ma" +7assiAa(i%n


23/23

RE@ERENCES

234 Md( !halid Imam Rahmani# Naina

klasifikasi gambar medis

Documents