Automated three-dimensional analysis of histological and autoradiographic rat brain sections

14 Pages·2007·0.62 MB·English

JournalofCerebralBloodFlow&Metabolism(2007)27,1742–1755
&2007ISCBFMAllrightsreserved0271-678X/07$30.00
www.jcbfm.com
Automated three-dimensional analysis of
histological and autoradiographic rat brain
sections: application to an activation study
Albertine Dubois1, Julien Dauguet1,2, Anne-Sophie Herard3,4, Laurent Besret3,
Edouard Duchesnay1, Vincent Frouin1, Philippe Hantraye3,4, Gilles Bonvento3,4
and Thierry Delzescaux1,4
1UIIBP, Service Hospitalier Frederic Joliot, CEA, Orsay, France; 2Computational Radiology Laboratory,
Harvard Medical School, Boston, Massachusetts, USA; 3Service Hospitalier Frederic Joliot, CEA CNRS URA
2210, Service Hospitalier Frederic Joliot, Orsay, France; 4MIRCen Program, Fontenay-Aux-Roses, France
Besides the newly developed positron emission tomography scanners (microPET) dedicated to the
in vivo functional study of small animals, autoradiography remains the reference technique widely
used for functional brain imaging and the gold standard for the validation of in vivo results. The
analysisofautoradiographicdataisclassicallyachievedintwodimensions(2D)usingasection-by-
sectionapproach,isoftenlimitedtofewsectionsandthedelineationoftheregionsofinteresttobe
analysedisdirectlyperformedonautoradiographicsections.Inaddition,suchapproachofanalysis
does not accommodate the possible anatomical shifts linked to dissymmetry associated with the
sectioningprocess.Thisclassicanalysisistime-consuming,operator-dependentandcantherefore
leadtonon-objective andnon-reproducibleresults.Inthispaper,wehavedevelopedan automated
and generic toolbox for processing of autoradiographic and corresponding histological rat brain
sections based on a three-step approach, which involves: (1) an optimized digitization dealing with
hundredsofautoradiographicandhistologicalsections;(2)arobustreconstructionofthevolumes
basedonareliableregistrationmethod;and(3)anoriginal3D-geometry-basedapproachtoanalysis
of anatomical and functional post-mortem data. The integration of the toolbox under a unified
environment (in-house software BrainVISA, http://brainvisa.info) with a graphic interface enabled a
robust and operator-independent exploitation of the overall anatomical and functional information.
We illustrated the substantial qualitative and quantitative benefits obtained by applying our
methodology to an activation study (rats, n=5, under unilateral visual stimulation).
Journal of Cerebral Blood Flow & Metabolism (2007) 27, 1742–1755; doi:10.1038/sj.jcbfm.9600470; published online
21March2007
Keywords: activation; autoradiography; functional data analysis; rodents; 3D reconstruction
Introduction investigation of the neurotransmission processes
(Araujo etal, 2000; Aznavour et al,2006).However,
The recent development of dedicated small animal these systems still suffer technical limitations
positron emission tomography scanners (microPET) includingalimitedsensitivityandareducedspatial
has opened up the possibility of performing resolution (B2–3mm) compared with autoradio-
repeated functional in vivo studies in the same graphy (B100–200mm). Therefore, the resolution of
animal: longitudinal follow-up of cerebral glucose PET scanning relatively to the size of rodent brain
metabolism and cerebral blood flow; studies on structuresisnotsufficienttoavoidincludingtissues
protein synthesis under different conditions; and with different rates of blood flow and metabolism
within a single voxel or region of interest. Conven-
tional autoradiography images are therefore gener-
Correspondence: Dr T Delzescaux, Service Hospitalier Fre´de´ric ally required to compare and validate in vivo
Joliot,4,placeduGe´ne´ralLeclerc,91401OrsayCedex,France. functional results obtained with small-animal PET
E-mail:[email protected] imaging and microPET technology (Thanos et al,
This project was supported by the Commissariat a` l’Energie 2002;Toyamaetal,2004;SchmidtandSmith,2005).
Atomique (CEA), France and the ‘Programme Interdisciplinaire
Additionally,thecostofaPETsystempermitsonlya
ImagerieduPetitAnimal’.
few laboratories to be equipped with them. Hence,
Received 21 September 2006; revised and accepted 12 January
2007;publishedonline21March2007 autoradiography remains the reference and widely
Automated3D-geometry-basedanalysisofbrainsections
ADuboisetal
1743
usedtechniqueforfunctionalbrainimaginginsmall 1998; NIH-Image, National Institutes of Health,
animal research. USA; MacPhase, Otter Solution, Whitesboro, NY,
Themajordisadvantageofautoradiographyisthat USA; VoxelView, Vital Images, Fairfield, IA, USA;
an animal can only be studied once. Longitudinal 3D-BrainStation, Loats Associates, Westminster,
studies require the use of multiple animals, adding MD, USA; and SURFdriver, Kailua, HI, USA),
inter-animal variability to other sources of variabi- genericandreliablealgorithmsarestillneededboth
lity. Another significant disadvantage is that for digitization of large numbers of sections and
autoradiography requires brain tissue sectioning, also for automated analysis taking advantage of 3D
entailing the production of up to several hundreds anatomo-functional reconstruction and allowing for
of serial sections and the inherent loss of the three- dissymmetry correction in the sections.
dimensional (3D) spatial consistency. Autoradio- In this paper, we have developed an automated
graphic data are traditionally analyzed in two and generic toolbox for processing of auto-
dimensions (2D) using a limited number of sections radiographic and histological rat brain sections.
and a part of the functional information is therefore This toolbox is based on a three-step approach
not exploited. In addition, depending on the whose strengths are: (1) an optimized data acquisi-
orientation of the cutting plane relative to the tion from large numbers of serial histological and
anteroposteriorandmediolateralaxes,thesymmetry autoradiographic sections (several thousands of
in the brain could have been lost during sectioning, sections obtained from several brains); (2) a reliable
as could section-by-section anatomical correspon- 3D reconstruction of the volumes using an adapted
dence between the right and left hemispheres. This registration method. This method is based on an
is of great importance when the analysis involves a original strategy involving accurate reconstruction
comparison betweenboth hemispheres,because the of both the anatomical volume and the functional
2D section-by-section approach can result in bias if volume by co-registration of each autoradiographic
the user does not take the possible dissymmetry in section to the corresponding registered histological
the sections into account, especially with small section; and (3) a novel approach for the analysis of
regions of interest. Finally, although the users have functional post-mortem data exploiting the overall
the corresponding post-mortem histological stained restored 3D geometry. We specifically applied the
sections available to consider anatomical informa- overall methodology for the characterization of
tion, the delineation of the regions of interest to be the metabolic changes throughout the visual system
analyzed is usually directly performed on the (visual cortex (VC), superior colliculus (SC),
autoradiographicsections,whichisoperator-depen- and lateral geniculate nucleus (LGN)) in lightly
dent and may not be accurate. restrained awake rats during unilateral stimula-
To extract maximum functional information from tion. Since the rat’s chiasm is approximately 90%
the overall autoradiographic brain sections in their crossed (Jeffery, 1984), this allowed comparison of
3D geometrically consistent alignment, a reliable stimulatedversusunstimulatedvisualsysteminthe
3D reconstruction of the data is essential. Many same animal.
methods have been proposed to align 2D histologi-
cal or autoradiographic sections into a 3D volume.
Materials and methods
They include fiducialmarker or artificial landmark-
based methods (Toga and Arnicar, 1985; Goldszal
Visual Stimulation and Measurement of Local
et al, 1995; Hess et al, 1998); principal axes
Cerebral Metabolic Rate of Glucose
alignment (Hibbard and Hawkins, 1988; Hess et al,
1998); feature-based methods, using contours, crest Autoradiographic and histological data sets used in this
lines, or characteristic points extracted from the work were obtained during a previously described
images (Hibbard and Hawkins, 1988; Zhao et al, activationstudy(Herardetal,2005).Thispreviousstudy
1995; Rangarajan et al, 1997); and gray level-based aimedatmeasuringthecerebralmetabolicrateofglucose
registration techniques using the intensities of the (CMRGlu) using the [14C]-2-deoxyglucose autoradio-
whole image, through similarity or correlation graphic method (Sokoloff, 1977) in the SC of adult rats
functions (Andreasen et al, 1992; Zhao et al, 1995; under a complex visual stimulation (n=5) in which the
Kim et al, 1997; Hess et al, 1998; Ourselin et al, lefteyewasleftopened(stimulated)andtherighteyewas
2001).Toberelevant,thefunctionalinformationhas closed with an opaque adhesive tape (unstimulated). In
to be compared and correlated to the corresponding the present work, we have gone further into the analysis
anatomical information, as in human brain studies. oftheSCdataandthevisualsystembyincludinganalysis
However, few of these works have specifically of the metabolic responses within the VC and the LGN.
addressed the co-registration of biologic images A complete and detailed description of the experimental
obtained from different techniques, for example protocolispresentedinHerardetal(2005).Coronalbrain
histology and autoradiography (Humm et al, 2003). sections (approximately 150 per animal, 20-mm thick)
Lastly, despite the fact that automated 3D recon- werecutwithacryostatat(cid:1)201C,mountedonSuperFrost
struction tools based on some of these registration glassslides,rapidlyheatdried,andexposedfor5daysto
methods become more widely available (Diaspro anautoradiographicfilm(KodakBioMaxMR,PerkinElmer,
et al, 1990; Lohmann et al, 1998; Thevenaz et al, Massy, France) along with radioactive [14C] standards
JournalofCerebralBloodFlow&Metabolism(2007)27,1742–1755
Automated3D-geometry-basedanalysisofbrainsections
ADuboisetal
1744
(146C,American RadiochemicalCompany,St Louis,MO, tion resolution needed to be sufficiently high to reveal
USA). The same brain sections were then stained with the main structures of interest in the brain: we chose a
cresyl violet (Nissl stain) to provide complementary 600dpi resolution (pixel size 42(cid:2)42mm2), in view of the
anatomicalinformation. sizeoftheratbrain.Theautoradiographicandhistological
sectionswereacquiredandstoredundertheformofglass
slidecolumnimagescalled‘overallscans’andwhosesize
wasgivenbythescanner’sfieldofview(Figure1A).Thus,
Optimized Data Acquisition: Digitization and
for a data set including approximately 150 sections,
Extraction of Sections from Scans
only five or six columns were needed for each of the
The autoradiographs, the corresponding histological two post-mortem imaging techniques (autoradiography
sectionsetsand the [14C]standards weredigitizedas8-bit and histology). The calibration scale available with the
gray-scale images with a flatbed scanner (ImageScanner, scannergavetherelationbetweenopticaldensityandgray
GE Healthcare Europe, Orsay, France). In-plane digitiza- levelvalues in theautoradiographic images.
Figure1 Proceduresfortheextractionofsectionsfromoverallscans.(A)Overall2Dscansofhistologicalsections(threeglassslides
arranged in a column, 600dpi). (B) Corresponding histogram (three modes: black border, sections, and background). (C) Binary
image,resultofthesectionextractionusingathresholdmethod.(D)Iterativemorphologicerosiononbinaryimage.(E)Separationof
thehorizontallyoverlappingsections.(F)Extractionoftheconnectedcomponents.Theactualsectionnumbercorrespondingtothe
orderinwhichtheyweresectioned,andhencefromwhichtheywereinthebrain,isautomaticallyassignedtoeachcomponentand
depictedby a differentcolor. (G) Theextracted coronalsections tobereconstructed in3D.
JournalofCerebralBloodFlow&Metabolism(2007)27,1742–1755
Automated3D-geometry-basedanalysisofbrainsections
ADuboisetal
1745
Theautomatedprocedureforextractionofsectionsfrom autoradiographic sections with the corresponding anato-
overall scans was based on thresholding and labeling, mical and functional volumes. However, an accurate
using techniques of robust histogram analysis and math- section-to-section registration was necessary to make the
ematical morphology. Histogram analysis was required to volumes spatially consistent in 3D. We used the block-
detect the main modes corresponding to the different matching method (Ourselin et al, 2001) in a propagative
classespresentinthescans(mainlysections,background scheme. This registration technique is especially well
and possible artifacts such as hand-written or manufac- adapted for the 3D reconstruction of biologic volumes
turedinscriptionsontheslides;Figure1B).Thehistogram arising from histological or autoradiographic sections
was iteratively smoothed by a Gaussian filter and the (Malandain et al, 2004; Dauguet et al, 2005a,b). Avector
positionofeachmodewasfollowedalongthescalespace field is computed between the two sections to be
(Mangin et al, 1998). The two modes that remained the registered using the correlation coefficient as similarity
longest were the background and the sections. A region- criterion between blocks and a rigid transformation is
growing method was applied in the histogram from the robustly estimated from this field. A multiresolution
positionscorrespondingtothemaximalvaluesretainedto approach ensures a coarse to fine estimation of the
determine the lower and upper boundaries for the gray optimal transformation. The anatomical volume was
levelsforeachmode,allowingtheautomatedcomputation reconstructed first by registering each anatomical section
of the threshold to be used to derive a binarized image withthefollowingoneinthestack.Then,bycomposition
of the sections (Figure 1C). An a priori knowledge of of the previously assessed transformations, each section
section number and surface was used to perform was aligned to a reference section (chosen because it
iterative erosions, and thereby identify and extract the carriesfewartifactssuchasfoldsortearsandislocatedin
main connected components (Figure 1D). In addition, an themiddleofthevolumetolimiterrorpropagation)soas
automated analysis of width and height parameters for to obtain a consistent 3D anatomical volume (Figures 2B
eachconnectedcomponentextractedaccordingtomedian and 2F). In a second step, this anatomical volume was
valuesallowedustodetectandtocorrectforverticaland usedasareferenceforthereconstructionofthefunctional
horizontaloverlapsbetweentwosections(Figure1E).This data. Each 2D autoradiographic section was directly co-
issueneededtobeautomaticallysolvedsinceevenavery registered with its corresponding registered histological
small overlap, and hence difficult to visually detect, is a section from the anatomical volume using the same
problem for the distinction between two overlapping block-matching method (Figures 2C and 2G). After the
sections, and hence for the individualization of sections 3Dreconstructionofthefunctionalvolume,thegraylevel
as independent connected components. On the basis intensities determined from the autoradiographic images
of X, Y coordinates of the gravity centers of the extracted were calibrated using the coexposed [14C] standard scale
sections, a positioning score was computed and used to and then converted to activity values (nCi/g of tissue)
sortandautomaticallyassigntheactualsectionnumberto usingapolynomialfourthdegreefitmethod,identifiedby
eachconnectedcomponentinthecolumn,corresponding the radioactive [14C] standard curve. The blood samples
to the order in which they were sectioned (vertical top- taken from the animals during the experiment were used
down, horizontal left–right; Figure 1F), and hence from to compute the parameters of the modified operational
which they were in the brain. The computation of a equationofSokoloff(1979)andtoconvertactivityvalues
rectangular bounding box around each connected and to CMRGlu values (mmol/100g/min; Figures 2D and 2H).
labeled component was based on the following: (1) the The3Dreconstructedanatomicalandfunctionalinforma-
horizontal and vertical dimensions of the biggest section tion, respectively, presented both an intra- and inter-
with security margins of 10%; and (2) the gravity center volumeconsistentgeometry(Figures 2I and2J).
information. Finally, the information relative to each
section was extracted from the initial overall scans and
producedindividualizedsectionsarrangedinthesection-
3D-Geometry-Based Analysis
ingprocessorder(Figure1G).Consideringtheoverlapping
sections, the computed bounding boxes included all the The analysis of autoradiographic data was performed the
informationrelativetoonesectionandasmallpartofthe same way through each visual structure. To avoid any
neighboringsection. redundancies, our 3D-geometry-based analysis will be
presentedandillustratedonlyintheSCwherethegreatest
metabolic change occurred during the visual stimulation.
Thein-housesoftwareAnatomist(Rivie`reetal,2003)was
3D Reconstruction of Anatomical and Functional
used for manual segmentation of the right and left SC
Volumes
(RSC and LSC, respectively) on each section of the 3D
Sectiondigitizationandextractionconstitutedaprelimin- rigidlyregisteredanatomicalvolume(Figure3A)yielding
ary step, which was the prerequisite before the 3D two volumes of interest (VOIs) and allowing assessment
reconstruction.Initialvolumeswereobtainedbystacking of their 3D shape (Figures 3B-3D; RSC in red, LSC in
the individualized coronal sections in the Z direction. green). As the anatomical and functional volumes were
The gravity center of each section was aligned with the co-registered, these VOIs were directly mapped on the
centeroftheboundingboxtoperformacoarsealignment functional volume (Figures 3E-G). The segmented VOIs
of the stack (Figures 2A and2E). Thus, the gravity center and the 3D reconstructed functional volume were then
parameters linked the individualized histological and usedtocreateanimageoftheprojectionofmeanCMRGlu
JournalofCerebralBloodFlow&Metabolism(2007)27,1742–1755
Automated3D-geometry-basedanalysisofbrainsections
ADuboisetal
1746
Figure2 3Dreconstructionsofoneratbraindataset(histologicalandcorrespondingautoradiographicsections)inaxialandsagittal
views,respectively.Anatomicalreconstructionbefore(AandE)andafterregistration(BandF).Functionalreconstructionafterco-
registration(CandG)andconversiontoCMRGluvalues(DandH).Surfacerenderingsofthe3Dreconstructedanatomical(I)and
functional(J) volumes.Signal intensitiesare color-coded accordingto the quantitativeCMRGlu scale(bottom).
values in axial incidence. This revealed the existence the subregions of contiguous activated voxels and reject
of a maximally activated CMRGlu subregion in the RSC, the smallest subregions corresponding to noise. The
corresponding to the metabolic response induced by the difference between the CMRGlu values in the activated
visual stimulation (Figure 3H). To compare activated RSC and the non-activated LSC is a measure of the
versusnon-activatedSCinthesameanimal,theactivated metabolicresponse inducedbythevisualstimulation. To
subregion within the RSC (activated SC, corresponding calculate this difference, the 3D subregion of metabolic
to the left-opened eye) and the symmetrized subregion activation, automatically outlined in the RSC had to be
within the LSC (non-activated SC, corresponding to the symmetrized in the LSC. To allow for the possible
closed eye) were automatically delineated. First, mean dissymmetryissueassociatedwiththesectioningprocess,
CMRGlu (m) and standard deviation (s.d.) values we first applied the symmetrization scheme to the entire
werecalculatedfortheLSC(Figure4A).Theyrepresented SC (Figure 4C).A flip over of the RSC around the vertical
the basal reference CMRGlu values. Voxels presenting a central axis of the volume was realized (Figure 4D). From
GMRGlu value more than T=m+2s.d. (significantly thisposition,theflippedRSCwasthenrigidlyregisteredin
higher than the mean CMRGlu value in the LSC) were 3D to the LSC using the block-matching registration
automatically outlined in the RSC, identifying the technique described above (Figure 4E). The activated
metabolic activation induced in the RSC by the visual subregionwasalsoflippedoveraroundtheverticalcentral
stimulation(Figure4B).Median-filtersmoothingandsize axis,andtherigid-bodytransformationestimatedfromthe
thresholding were then applied to respectively regularize entire SC was applied to the flipped activated subregion
JournalofCerebralBloodFlow&Metabolism(2007)27,1742–1755
Automated3D-geometry-basedanalysisofbrainsections
ADuboisetal
1747
Figure3 (A)ManualsegmentationofLSCandRSConeachregisteredsectionoftheanatomicalvolume(whiteareasinA).Three-
dimensionalsurfacerenderingsoftheLSCandRSC:locationinthecorrespondingregisteredanatomicalbrainpart(B)anddepiction
incoronalandaxialviews(CandD).MappingofmanuallysegmentedLSCandRSConeachcorrespondingandco-registeredsection
ofthefunctionalvolume:coronal,axial,andsagittalviews(E–G).AxialprojectionofmeanCMRGluvalueswithinthesegmentedLSC
andRSC(H). Signalintensities arecolor-codedaccording to thequantitativeCMRGlu scale (bottom).
(Figures 4F-4H) to delineate a non-activated subregion within animals. P-values less than 0.05 were considered
(dark green) in the LSC corresponding to the symmetric significant.
form of the activated subregion in the RSC (dark red) All computerized treatments and procedures presented
(Figure 4I). To ensure that only voxels mapping SC tissue in this paper (section extraction, 3D reconstruction of
wereincludedintheanalysis,thevoxelsofeachsubregion anatomical and functional volumes, conversion of func-
lyingoutsideofthecorrespondingSCweremaskedout. tional data values to CMRGlu and 3D-geometry-based
Morphometric parameters relative to the shape and the analysis)werewritteninC++.Theywerealsointegrated
volumeofvariousbrainregionsofinterestwereassessed: within-house software BrainVISA (Cointepas etal, 2001;
(1) the entire RSC (activated) and LSC (non-activated) http://brainvisa.info)andgatheredinplugged-inmodules
obtained after manual segmentation; (2) the activated dedicated to the processing of rat brain histological and
subregionautomaticallyextracted in theRSC;and(3) the autoradiographicsections.Althoughtheyweredeveloped
correspondingflippedandsymmetrizedsubregionsinthe and implemented under BrainVISA environment on a
LSC. Mean CMRGlu values in activated (CMRGlu ) Linuxworkstation,thetreatmentsareabletorunonmost
activated
and non-activated (CMRGlu ) regions were mea- operating systems (Macintosh or Windows) and can be
non-activated
sured and used to compute corresponding relative meta- usedonapersonalcomputer,whichfacilitatestheirdaily
bolic rate changes (MRC, expressed as percentages) using use anddatahandling.
the followingformula:
CMRGlu (cid:1)CMRGlu
relativeMRC ¼100(cid:2) activated non(cid:1)activated
CMRGlu
non(cid:1)activated Results
For2D/3Dcomparativeanalyses,meanCMRGluvaluesin
each segmented section of the RSC and LSC were also To validate the overall methodology, the processing
measured. All results are expressed as means7s.d. stages(fromsectiondigitizationto3D-geometry-based
Student’spairedt-testwasusedtocomparemeanCMRGlu analysis) were applied to five rats with the left eye
JournalofCerebralBloodFlow&Metabolism(2007)27,1742–1755
Automated3D-geometry-basedanalysisofbrainsections
ADuboisetal
1748
Figure 4 Procedures for the 3D-geometry-based analysis of functional information illustrated with rat 2 whose brain was cut in a
slightlydissymmetricalway.Firststep:(A)computationofmean,m,andstandarddeviation,s.d.,CMRGluvaluesintheLSC(non-
activated)and(B)automaticextractionofthemaximallyactivatedCMRGlusubregion(darkred)intheRSC(activated)usingm+
2s.d.threshold.Theresultisrepresentedwith3Dsurfacerenderingsincoronalandaxialviews.Secondstep:(C)three-dimensional
surface renderings of segmented left (light green) and right (light red) SC in axial and coronal views. (D) Flip over around vertical
central axis of the RSC. (E) Rigid registration between LSC and flipped RSC. Third step: (F) Activated subregion in the RSC,
previouslyextractedin(B).(G)Flipoveraroundverticalcentralaxisoftheactivatedsubregion(darkgreen).(H)Applicationofthe
transformation parameters computed with the entire SC to the flipped over activated subregion. Final result: (I) The automated
extractionoftheactivatedsubregionintheRSCisrepresentedindarkredandthesymmetricsubregioninthenon-activatedLSCin
dark green.
open (stimulated) and the right eye closed with sections and autoradiographs, encompassing VC,
opaque adhesive tape (unstimulated). Histological SC,andLGN(approximately300imagesintotalper
andautoradiographicdatasetswereeachcomposed animal),wasacquiredandstoredbyoperatorinless
of approximately 150 sections, divided up as than 10mins. Then, the images were automatically
follows: five columns of five glass slides bearing and successfully extracted from the overall scans in
six sections. Thus, there were approximately 1500 less than 15mins.
sectionsintotal,whichweresuccessfullyprocessed After the 3D reconstruction (one and a half hours
using the above-described methodology. of computing time per data set to be reconstructed
Using our optimized digitization procedure that we generally get working during the night), we
(Figure 1), each series of stained histological obtainedbothconsistentandco-registeredanatomical
JournalofCerebralBloodFlow&Metabolism(2007)27,1742–1755
Automated3D-geometry-basedanalysisofbrainsections
ADuboisetal
1749
and functional volumes. Quality of the registration
and 3D reconstruction process was assessed by
visual inspection of internal structures viewed
in different orthogonal incidences (Figures 2B, 2C,
2F, and 2G) as well as by visual inspection of the
3D surface renderings of the corresponding anato-
mical and functional volumes (Figure 2I and 2J,
respectively).
The 3D functional surface rendering was
sectioned with four different axial cutting planes
moving along the dorso-ventral direction (Figure 5).
Each of them displays the differentially activated
regions obtained in response to the visual stimu-
lation in each visual structure, namely the areas 17
(OC1) and 18a (OC2l) of the VC, the SC, and the
LGN (Figures 5A, 5B, 5C and 5D, respectively). An
increasedCMRGluareaisvisibleineachrightvisual
structurecomparedwiththeleftcorrespondingone.
Using our newly developed procedure for the
3D-geometry-based analysis of functional informa-
tion(Figures3and4),wewereabletoautomatically
delineate the maximally activated subregion(s) and
the symmetrized subregion(s) in each right and left
visual structure and therefore to assess their shape,
their location within the structure and their spatial
extent. In Figure 5E, these subregions are depicted
in their corresponding visual structure and reposi-
tioned within the 3D reconstructed anatomical
volume. Tables 1-3 summarize all the anatomical
and functional information including the relative
MRC obtained for the five animals in each visual
structure (VC, SC, and LGN, respectively). The
volume of each visual structure was very similar
between the left and right hemispheres in the same
animal as well as between animals. The automated
extraction of the activated subregion(s) identified
anatomically restricted volumes that respectively
encompassed 3, 10, and 16% of the volume and 65,
42, and 56% of the sections covering each corre-
spondingrightvisualstructure(rightVC(RVC),RSC,
and right LGN (RLGN), respectively; Figure 5E). The
volumeoftheflippedsubregion(s)ineachleftvisual
structure was lower than the one measured for the
activated subregion(s). They showed an important
variability because only few animals presented a
perfectly symmetric flipped subregion. After the
application of our symmetrization procedure in all
fiverats,thevolumeofthesymmetrizedsubregion(s)
was similar to the one measured for the activated
subregion(s).CMRGluintheentireactivatedRVCand
Figure 5 Four 3D surface renderings of the functional volume
RLGNwassignificantlyincreasedcomparedwiththe indicating the antero-posterior and dorso-ventral location as
non-activated left VC (LVC) and left LGN (LLGN) well as the spatial extent of the metabolic activation in each
(**P<0.01 and *P<0.05, respectively; Figure 6), visual structure (areas 17 and 18a of VC, SC, and LGN; (A),
whereas we did not observe any increase in the (B),(C),(D),respectively)duringvisualstimulationinaratwith
entire activated RSC (P>0.05; Figure 6). For all the one eye closed and one eye open. Signal intensities are color-
three visual structures, CMRGlu in the activated coded according to the quantitative CMRGlu scales (right). (E)
Three-dimensional surface rendering of the activated and
subregions(s) was significantly higher than in the
symmetrizedsubregionsautomaticallydelineatedineachvisual
symmetrized subregions(s) (***P<0.001; Figure 6).
structureofthisanimalandrepositionedwithinthecorrespond-
The relative MRCs determined within these sub-
ing3D reconstructed anatomicalvolume.
regions were +20, +23, and +17% for the VC, SC,
and LGN, respectively.
JournalofCerebralBloodFlow&Metabolism(2007)27,1742–1755
Automated3D-geometry-basedanalysisofbrainsections
ADuboisetal
1750
Table 1 Results of analysis in VC (n=5, left eye open, right eye closed, under visual stimulation): anatomic and functional
informationandrelative MRC computation
Anatomicalinformation:volumes(mm3and%)
Animal Entire Entire Activated %ofRVCvolumetakenup %ofsectionsofRVCtakenup Flipped Symmetrized
RVC LVC subregion bytheactivatedsubregion bytheactivatedsubregion subregion subregion
Rat1 10.30 10.30 0.30 3 64 0.23 0.25
Rat2 11.04 10.86 0.14 1 46 0.00 0.05
Rat3 12.20 12.20 0.27 2 83 0.19 0.25
Rat4 10.30 10.60 0.48 5 71 0.47 0.44
Rat5 12.50 12.00 0.40 3 61 0.35 0.35
Mean 11.27 11.19 0.32 3 65 0.25 0.27
s.d. 1.04 0.86 0.13 1 14 0.18 0.15
Functionalinformation:CMRGluvalues(mmol/100gmin) RelativeMRC(%)
Animal Entire Entire Activated Symmetrizedsubregion Between Betweenright
RVC LVC subregion entireVCs andleft
subregions
Rat1 150.2 141.2 182.0 160.8 6.4 13.2
Rat2 142.7 137.9 171.3 138.0
Rat3 139.0 136.2 175.4 147.1 2.1 19.2
Rat4 132.9 124.4 172.8 145.0 6.8 19.2
Rat5 163.9 152.9 200.3 160.2 7.2 25.0
Mean 145.7** 138.5 180.4*** 150.2 5.6 19.2
s.d. 11.9 10.2 11.9 10.0 2.4 3.4
**P<0.01comparedwiththemeanCMRGluvalueintheentireleftVC(Student’spairedt-test);***P<0.001comparedwiththemeanCMRGluvalueinthe
symmetrizedsubregion(Student’spairedt-test).
Discussion written slide number or slide border and of the
possible section overlaps). In addition, the section
The aim of this paper was to develop a dedicated, extractionprocedureisgeneric:histologicalsections
generic, and automated methodology for 3D-geo- are handled in exactly the same way as autoradio-
metry-based morphometric and functional analysis graphic images.
and to illustrate its substantial qualitative and
quantitative benefits by applying it to an activation
study in the rat.
Three-dimensional Reconstruction Strategy for
Anatomical and Functional Volumes
Sections were first stacked in the Z axis using
Overall Section Digitization and Extraction
gravity center parameters for each section. The
Rather than digitizing sections one by one as it is intermediate volumes thereby already presented
usually the case with a CCD camera and a lighting a good quality of stacking (Figures 2A and 2E).
table,our procedureinvolves amultiple acquisition However, this initial step only provided a coarse
undertheformofglassslidecolumns(overallscans) registration of histological or autoradiographic
using a flatbed scanner. This digitization procedure sections (inner brain structures were not properly
significantly reduces the acquisition time (300 registered).
sections digitized in less than 10mins whereas it A Rigid Pairwise Registration: Three-dimensional
takes 1h with a CCD camera). Unlike other pre- reconstruction generally involves the sequential
viously reported algorithms designed for the same registration of each section to its adjacent section
purpose (Goldszal et al, 1995; Nikou et al, 2003; using linear or non-linear image registration techni-
Nguyen et al, 2004; Lee et al, 2005), we made ques (Hibbard and Hawkins, 1988; Goldszal et al,
extraction of sections from overall scans entirely 1995;Zhao etal,1995). Here, weusedaregistration
automated, reproducible (number assignment and technique based on a rigid body transformation
rectangularboundingboxesaroundeachsectionare betweenadjacentcoronalsections.Thistypeofregi-
automatically computed) and robust (consideration strationisstandard,robust,andwelladaptedforbrain
of the troublemaker modes resulting from various sections obtained with a cryostat because a rigid
artifacts appearing in the scans, such as hand body transformation is sufficient to superimpose
JournalofCerebralBloodFlow&Metabolism(2007)27,1742–1755
Automated3D-geometry-basedanalysisofbrainsections
ADuboisetal
1751
Table 2 Results of analysis in SC (n=5, left eye open, right eye closed, under visual stimulation): anatomical and functional
informationandrelative MRC computation
Anatomicalinformation:volumes(mm3and%)
Animal Entire Entire Activated %ofRSCvolumetakenupby %ofsectionsofRSCtakenup Flipped Symmetrized
RSC LSC subregion theactivatedsubregion bytheactivatedsubregion subregion subregion
Rat1 3.87 3.65 0.21 5 37 0.07 0.19
Rat2 4.15 4.19 0.52 13 43 0.01 0.49
Rat3 3.71 3.39 0.37 10 39 0.31 0.32
Rat4 3.47 3.54 0.55 16 47 0.54 0.54
Rat5 3.30 3.29 0.28 9 44 0.14 0.21
Mean 3.70 3.61 0.39 10 42 0.21 0.35
s.d. 0.33 0.35 0.15 4 4 0.21 0.16
Functionalinformation:CMRGluvalues(mmol/100gmin) RelativeMRC(%)
Animal Entire Entire Activated Symmetrizedsubregion Between Betweenrightand
RSC LSC subregion entireSCs leftsubregions
Rat1 117.6 117.5 162.6 134.9 0.2 20.5
Rat2 120.2 120.9 163.4 125.5 (cid:1)0.6 30.2
Rat3 106.9 114.1 157.5 129.6 (cid:1)6.3 21.5
Rat4 109.7 105.4 151.2 115.5 4.0 31.0
Rat5 134.8 133.7 174.5 153.1 0.8 14.0
Mean 117.9 118.3 161.9*** 131.7 (cid:1)0.4 23.4
s.d. 10.9 10.4 8.6 13.9 3.7 7.1
***P<0.001comparedwiththemeanCMRGluvalueinthesymmetrizedsubregion(Student’spairedt-test).
one section on the next one. Although the data can tion method is very low owing to the thinness
in some cases exhibit deformation artifacts as a and good quality of the post-mortem data sets;
result of sectioning and tissue shrinkage (Kim et al, and (2) the block-matching technique is robust to
1997), non-rigid deformations between the adjacent dissimilarities between sections, missing data, and
sections can distort the brain structures. Therefore, outlying measurements (Ourselin et al, 2001).
it appears better to preserve the shape of each Anatomy as Reference: The block-matching regis-
section without compensating for the deformation tration technique is based on both the section edges
than to take the risk of distorting the overall and and the whole image, so the result of the 3D
regional image information during the registra- reconstruction will depend on the type of data
tion process (Lee et al, 2005). We used the classic (histological or autoradiographic) to be processed,
scheme, consisting in serially propagating the that is to say, on the information available in the
transformations estimated between consecutive sections. Even if we chose the same reference
sections relative to a reference section in the series. section and despite the fact these were the same
This approach has been criticized because it can physical sections, independently registering histo-
leadtodifferenttypesofmisregistrations.According logical and autoradiographic sections would not
to Nikou et al (2003), if an error occurs in the give the same result and would not allow a perfect
registration of a section about the previous section, superposition of each section, which is a prerequi-
this error will be propagated through the entire site for the delineation of ROIs. It is not either
volume. Thus, if the number of sections to be possible to reuse the transformations computed
registered is large, an overall offset of the volume, during histological section registration to recon-
because of error accumulation, is entirely plausible. structthefunctionalvolumeandviceversa.Indeed,
However, these issues are more pronounced when histological and autoradiographic sections were
distant sections are involved in the registration, extracted separately and consequently, they do not
which is not our case (20-mm-thick adjacent serial have the same configuration (dimensions of bound-
sections). Consequently, we believe that our ap- ing boxes, computation of gravity centers). In this
proach of section-to-section registration (Malandain work, one of our objectives was to propose a joint
et al, 2004; Pitiot et al, 2005), in the absence of any 3D-geometry-based anatomo-functional exploitation
3D geometrical reference (such as magnetic reso- of post-mortem data. Thus, the variability between
nance imaging scans or images of the blockface thetypes ofdatapresented problem.Hence,wehad
captured before each section), is the most efficient to develop a 3D reconstruction strategy providing
because: (1) the percentage of errors of this registra- an optimal anatomo-functional section-by-section
JournalofCerebralBloodFlow&Metabolism(2007)27,1742–1755