From: AAAI Technical Report SS-94-05. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved.
A KNOWLEDGE-GUIDEDALGORITHMFOR BRAIN LESION DETECTION
Yi Lu
Departmentof Electrical and ComputerEngineering
The University of Michigan . Dearborn
Dearborn, MI 48128
Phone: (313) 593-5420
FAX: (313) 593-9967
E.mail: lu@umdsun2.umd.umich.edu
1. Introduction
Thedelineationof brain tumorin CTor
MRIsequencesis important in many
medicalresearch environments
and clinical
applications.It is extremely
usefulin
treatmentplanningandevaluation.
Cunentlyin manymedicalcenters, 3-D
minorvolumesare obtainedby stacking the
boundariesof tumorstraced manuallyor
semi-automatically
on 2-Dimageslices.
Thehumanvisual systemis an ideal
mechanism
for quicklyextracting a general
description of an image.However,in the
detailedanalysisof subtle features, manual
methodsare not only tedious but also
subjective,leadingto substantialinter and
intra-observervariability. Furthermore,
it is
difficult to obtainaccurateevaluationof
therapeuticefficacy basedon manual
delineation, since measurements
of tumor
volumesbased on humanobservation are
not reproducible.
imageseries, determinethe apt size of
morphological
operators, link features
extractedat different scales andeliminate
irrelevant features. Thealgorithmis both
data- and goal-driven. Theexperimentsof
the algorithmis described.
2. Multiscale Image Analysis using
Morphological Operations
In the field of computervision, multiscale
analysis is an importanttechniquefor
extractinginteresting features fromimages.
In low level vision processes, a smoothing
processis often appliedto imagesbefore
the extractionof desiredfeatures. In
general, wecan get fine features from
operatorsof small scales andcoarse
informationfromoperatorsof large scales.
In manyapplications, no single scale of
smoothing
is sufficient. Themultiscale
analysis techniquecan be usednot only to
eliminatefree-scale noisebut also to
separateeventsof different scales arising
fromdistinct physicalprocesses.
This paperaddressesthe problemof the
automaticdetectionof brainlesions in x-ray
computedtomography(cr) imagery.
Automatic
detectionof lesions is a nontrivial problem.Typicallythe boundariesof
lesions in CTimagesare of single-pixel
width,andthe gradientat the lesion
boundaryvaries considerably. As many
studies show,these characteristics of
lesions within CTimages,in conjunction
with the generallylowsignal-to-ratio of CT
images,render simpleboundarydetection
techniquesineffective.
In this paper, weapplymultiscale
morphologicaloperations to brain tumor
imageanalysis. Themathematical
morphology
approachis powerfulin
studying manyvision problems[Lull92].
In general, morphologicalopeningand
closin.g can be viewedas smoothing
operations. Anopeningfilter can remove
objectswithsize smallerthan the structure
element,break narrowpans of a region into
subregions and smooththe rough edges of
object contours,therefore the regions
extractedfroma filtered imagemaynot be
identical to the corresponding
regionsin the
original image.Aclosing operatorwith a
disk structuring elementsmoothesobject
contours, fuses narrowbreaks and long
thin areasof a regionandfills in small
Wepropose a knowledge-guidedtumor
detectionalgorithm.Thealgorithmapplies
multiscale morphological
operationsto CT
nuageswith the guidanceof both anatomic
and physical knowledge.Theknowledgeis
usedto select the initial slice withina CT
88
holes andgapson the object contour.
Morphological
filtering of an imageby an
openingor closing operationcorresponds
to the ideal ban@ass
filters of conventional
linear filtering. If weconsiderthe size of B
r as the scale parameter,as r changesfrom
1 to infinity, the filtered imagesandr form
a scale space.
Themorphologicaloperations can be
extendedfrombinary into gray-scale
imagesby introducingthe conceptof
umbra.The umbrais the volumebelowthe
graylevel surface. In grayscale images,
dilation is accomplished
by taking the
maximum
of a set of sums,anderosionis
accomplishedby taking the minimum
of a
set of sums.Hencegray scale dilation and
erosion have the samecomplexityas
convolution.However,instead of doingthe
summationas in convolution, minimum
or
maximum
is performed.Theset is def’med
bythe structureelementusedin the dilation
or erosion.
Ingray
scale
images,
thedesired
features
canoccur
atanygray
levels.
Indeed,
we
cannotdifferentiate foregroundfrom
backgroundpixels without the knowledge
aboutthe desired features. Basedon the
definition of openixigandclosing
operationsin gray level images,an opening
operation can mean’closing’ to some
regions and’opening’to others. Hencethe
decisionon applyingopeningor closing is
dependenton the gray scale distribution
modelformedby its neighboringregions.
In CTimages,the desired features can
occurat anygray levels. Basedon our
study, wediscoverthat brain lesions have
the followingcharacteristics:
I.brain
lesions
canoccur
indifferent
shape
andsize,
2.thebrightness
ofthelesion
varies
depending
onindividual,
3.theboundaries
areoften
fuzzy
and
sometimesdo not formclosed curves,
4.thebrightness
andthetexture
within
theregion
oflesion
areoften
inconsistent,
and
5.brain
lesion
canhave
thesame
or
opposite
contrast
asthesurrounding
anatomy.
89
Ourstudy on the behaviorof brain tumorin
morphologicalscale space showed[Lull92]
that behaviorof brain tumorcan be
described
byselecting proper
multiscale
morphological
operations
andscale
parameters.If the braintumoris lighter
than its surroundingarea, a sequenceof
openingoperationscan be appliedto the
grayimageto extract the minorboundaries,
otherwisea sequenceof closing operations
can be used. In the multiscaleopening-andclosingfiltered images,the bright areas do
not changemonotonicallywith the
parameter
r., andtheyaxenot sensitiveto
small
holes andcavities. Thedark areas
increase morerapidly than the opening
filtered images.Asthe scale parameter
increases, small dark regions are merged
with neighboringregions instead of being
eliminatedas thosein the openingfiltered
images.In general,the regionsin the
opening-and-closing
filtered imageshave
smootherboundariesthan the filtered
imagesbyopening
orclosing
alone
andcan
also
beused
toextract
lighter
regions.
Basedon the abovestudy, we have
constructedthe followingalgorithmfor
detecting
brain tumors.
3. Detecting Brain TumorsGuided
by Knowledge
Ouralgorithmis bothdata- andgoaldriven. The following knowledgesources
are used to guide various computa6onal
steps in the algorithm:
¯ Normalbrain anatomy.For example,
lateral ventricles,straightsinus,andfalx
cerebri in CTimages,andthird and
fourth ventricle in MRimages.These
objectsare relativelyeasyfeaturesto be
identified andextractedfromthe images
andtherefore
they
canbeused
as
landmarks
toguide
theprocesses
inour
algorithm.
¯ Physicalknowledge.
It includes
scanninganglewithrespect to the
orbitomeatal
plane, initial scanning
location
andthethickness
of theimage
slices. In particular, MR
imagingis
multiparametric
andthe signal contrast
depends on proton density, T1 and T2
relaxation times, and blood flow. This
type of knowledgecan be used to predict
the occurrenceand location of the
anatomic landmarksand to hypothesize
the locations of tumors.
¯ Physicians’
knowledge.
We use
physician’s
knowledge
aboutbrainimage
characteristics to locate brain tumors and
deal with difficult cases in whichthe
tumorboundariesare either not
discernible or ambiguous.
¯ Knowledgeabout object behavior under
multiscale segmentationoperators. Object
can behavedifferently under different 3-D
segmentationoperators and at different
scales [LuJ92]. For any selected 3-D
segmentationoperator, the behavior of
objects in the scale spacewill be an
important knowledgesource for the
reasoning processes in the system.
Thealgorithm
consists
of thefollowing
major
computational
steps:
(1) Selectingan initial slice of CTimages.
Theinitial slice is selected basedon the
scanning parametersand the estimation of
the appearanceof the tumorin the scanning
direction. Ideally the initial slice should
contain strong features of brain tumor.
(2) Selecting an appropriate sequence
operators, a priori knowledgeof intensity
contrast of the brain lesion is usedto select
either openingor closing operations. If a
regionof interest is lighter thanthe
surrounding regions, a sequenceof
openingoperations is applied to the gray
scale image, otherwise a sequenceof
closing operations.
(3) Object segmentation. Theobject
segmentationis performedbased on the
histogramsof the filtered imagesat
different scales. As the scale of openingor
closing operator increases, the histograms
of the filtered imagesprovidebetter
informationto separate objects occurringat
different gray levels. Aclustering algorithm
is developedto separate objects with
different gray level distributions. The
90
programoperates on histograms across
multiple scales. It groupsthe gray scales
based on their histogramvalues and their
distance to the next immediategray scale
that has non-zero histogram value. The
stable groupof clusters obtained from the
histogramof the filtered imageat the
minimum
scale is used to extract objects.
Eachcluster correspondsto one class of
objects whichhave the gray levels within
the cluster. Thestability of dusters is
measureuponthe consistency
of dusters
across several different scales. Thus,
objects
arcseparated based
ontheir
gray
level distributions in the image.
(4) Identifying the features of interest.
can uniquelyidentify the desired features
such as tumor based on knowledgesuch as
geometric shape and possible location.
(5)Update
reference area.
The reference
area is a subimageof a slice that must
containthe feature of interest. In the initial
slice, the reference area is the entire image.
Oncewe have detected the feature of
interest in one slice, wecan computean
area within whichthe feature extends to the
two immediateneighboring slices of images
in the sequence. The histogram should
always be computedfrom the reference area
and the reference area should be updatedat
everyslice.
4. Experimental Results
Figure 1 (a) shows a subimageof a
imagewhichcontains a brain lesion with a
necrosis area. Figure 1 (b) showsthe
histogram of the image. The histogram has
one modeand two heaps. The two heaps
represent the brain lesion and the lateral
ventricles. Let’s assumethe desired
features are the lateral ventricles and the
brain lesion. Apparentlythere are no clear
dusters to represent the two heaps. Figure
1. (c) and (d) showthe histogramsof
filtered imageby morphologicalopening
with structure elementsof disks of radius 2
and 3. It is obviousthe histogramsin (c)
and (d) provides moreinformation
separating the two heaps. Ourclustering
program
founda stable
group
of three
meaningful
clusters
fromthehistogram
of
thefiltered
image
bydiskofradius
2.The
groupcontains three clusters, 165 to 187,
and 188 to 208, and 209 to 225. The
segmentedobjects are shownin (e), (0
(g). Thetumorand the lateral ventricles can
be easily identified from(e) and (g)
respectively provided we have geometric
knowledgeabout the desired features.
$. References
[LuH92] Yi Lu and Laurel Harmon,
"Multiscale Analysis of Brain Tumorsin
CTImagery," 21st Applied ImageryPattern
Recognition Workshop, SPIE, Oct. 14-16,
1992.
[LuJ92] Yi Lu and RameshJain,
"Reasoning about Edges in Scale Space,"
IEEETransactions on Pattern Analysis and
MachineIntelligence, Vol. 14, no. 4, April,
1992, PP. 450-468
Figure 1. (d) I2Iistos~ of the image
filtered
byanopeningoperation
witha disk
of size
3.
-- ~..,~.~,-
!"
Fig~e1. (e)Objec==~havegray levels
between 165 and 187.
Figure 1. (a) Input image.
Figure 1. (0 Objects
have gray levels
between 188 and 208.
Figure 1 (b) Histogram
imagein (a)
¯ )
Hgure1. (g) Objects
have gray levels
between 209 and 225.
Figure "1. (c)’Histogramoi the image
filtered by an openingoperation with a disk
of size 2.
91