bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Using a structural root system model
for an in-depth assessment of
root image analysis pipeline
Guillaume Lobet 1,2,*, Iko Koevoets 3, Pierre Tocquin1,
Loïc Pagès4 and Claire Périlleux1
1
InBioS-PhytoSYSTEMS,
University of Liège, 4000 Liège, Belgium
2
Institut fur Bio-und Geowissenschaften: Agrosphare,
Forschungszentrum Jülich, D52425 Jülich, Germany
3
Plant Cell Biology, Swammerdam Institute for Life Sciences,
University of Amsterdam, 1098 XH Amsterdam, The Netherlands
4
INRA, Centre d’ Avignon, UR 1115 PSH,
Site Agroparc, 84914 Avignon cedex 9, France
* Corresponding author: [email protected]
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
1
Abstract
2
Rootsystemanalysisisacomplextask,oftenperformedusingfullyautomatedimageanalysis
3
pipelines.However,thesepipelinesareusuallyevaluatedwithalimitednumberofground-truthed
4
rootimages,mostlikelyoflimitedsizeandcomplexity.
5
6
Wehaveusedarootmodel,ArchiSimpletocreatealargeanddiverselibraryofground-truthed
7
rootsystemimages(10.000).Thislibrarywasusedtoevaluatetheaccuracyandusefulnessof
8
severalimagedescriptorsclassicalyusedinrootimageanalysispipelines.
9
10
Ouranalysishighlightedthattheaccuracyofthedifferentmetricsisstronglylinkedtothetypeof
11
rootsystemanalysed(e.g.dicotormonocot)aswellastheirsizeandcomplexity.Metricsthathave
12
beenshowntobeaccurateforsmalldicotrootsystemsmightfailforlargedicotsrootsystemsor
13
smallmonocotrootsystems.Ourstudyalsodemonstratedthattheusefulnessofthedifferent
14
metricswhentryingtodiscriminategenotypesorexperimentalconditionsmayvary.
15
16
Overall,ouranalysisisacalltocautionwhenautomaticallyanalysingrootimages.Ifathorough
17
calibrationisnotperformedonthedatasetofinterest,unexpectederrorsmightarise,especiallyfor
18
largeandcomplexrootimages.Tofacilitatesuchcalibration,boththeimagelibraryandthe
19
differentcodesusedinthestudyhavebeenmadeavailabletothecommunity.
20
2
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
21
Introduction
22
Rootsareofoutmostimportanceinthelifeofplantsandhenceselectiononrootsystems
23
representsgreatpromiseforimprovingcroptolerance(asreviewedin(Koevoetsetal.,2016)).As
24
such,theirquantificationisachallengeinamultitudeofresearchprojects.Thisquantificationis
25
usuallytwofold.Thefirststepconsistsinacquiringanimageoftherootsystem,eitherusingclassic
26
imagetechniques(CCDcameras)ormorespecializedones(microCT,X-Ray,fluorescence,...).The
27
nextstepistoanalysethepictureinordertoextractmeaningfuldescriptorsoftherootsystem.
28
29
Toparaphrasethefamousbelgiansurrealistpainter,RenéMagritte,figure1Aisnotarootsystem.
30
Figure1Aisanimageofarootsystemandthatdistinctionisimportant.Suchanimageisindeeda
31
twodimensionalrepresentationofarootsystem,whichisusuallyathreedimensionalobject.Until
32
now,measurementsaregenerallynotperformedontherootsystemsthemselves,butontheimages
33
andthisraisessomeissues.
34
35
Imageanalysisis,bydefinition,theobtentionofmetrics(ordescriptors)describingtheobjects
36
containedinaparticularimage.Inaperfectsituation,thesedescriptorswouldaccuratelyrepresent
37
thebiologicalobjectoftheimagewithnegligibledeviationfromthebiologicaltruth(ordata).
38
However,inmanycases,artefactsmightbepresentintheimagessothattherepresentationofthe
39
biologicalobjectisnotaccurateanymore.Theseartefactsmightbeduetotheconditionsinwhich
40
theimagesweretakenortotheobjectitself.Maturerootsystems,forinstance,arecomplex
41
branchedstructure,composedofthousandsofoverlapping(fig.1B)andcrossinglinearsegments
42
(fig.1C).Thesefeaturesarelikelytoimpedeimageanalysisandcreateagapbetweenthe
43
descriptorsandthedata.
44
3
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
45
46
47
Rootimagedescriptorscanbeseparatedintotwomaincategories:morphologicalandgeometrical
48
descriptors.Morphologicaldescriptorsrefertotheshapeofthedifferentrootsegmentsformingthe
49
rootsystem(table1).Theyinclude,amongothers,thelengthanddiameterofthedifferentroots.
50
Forcomplexrootsystemimages,morphologicaldescriptorsaredifficulttoobtainandareproneto
51
errorasmentionedabove.
52
53
Geometricaldescriptorsgivethepositionofthedifferentrootsegmentsinspace.Theysummarize
54
theshapeoftherootsystemasawhole.Thesimplestgeometricaldescriptorsarethewidthand
55
depthoftherootsystem.Sincethesedescriptorsaremostlydefinedbytheoutsideenvelopeofthe
4
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
56
rootsystem,crossingandoverlappingrootshavelittleimpactontheirestimationandtheycanbe
57
consideredasrelativelyerrorless.Geometricaldescriptorsareexpectedtobelooselylinkedtothe
58
actualrootsystemtopology,asidenticalshapescouldbereachedbydifferentrootsystems(the
59
oppositeistrueaswell).Theyareusuallyusedingeneticstudies,toidentifygeneticbasesofroot
60
systemshapeandsoilexploration.
61
62
Severalautomatedanalysistoolsweredesignedinthelastfewyearstoextractbothtypeof
63
descriptorsfromrootimages(Armengaudetal.,2009;Buckschetal.,2014;Galkovskyietal.,2012;
64
Pierretetal.,2013).However,thevalidationofsuchtoolsisoftenincompleteand/orerrorprone.
65
Indeed,fortechnicalreasons,thevalidationisusuallyperformedonasmallnumberofground-
66
truthedimagesofyoungrootsystemsforwhichmostanalysistoolswereactuallydesigned.Inthe
67
fewcaseswherevalidationisperformedonlargeandcomplexrootsystems,itisusuallynoton
68
ground-truthedimages,butincomparisonwithpreviouslypublishedtools(measurementofXwith
69
toolAcomparedwiththesamemeasurementwithtoolB).Thismightseemreasonableapproach
70
regardingthescarcityofground-truthedimagesoflargerootsystems.However,theinherent
71
limitationsofthesetools,suchasscaleorplanttype(monocot,dicot)areoftennotknown.Users
72
mightnotevenbeawarethatsuchlimitationsexistandapplytheprovidedalgorithmwithout
73
furthervalidationontheirownimages.Thiscanleadtounsuspectederrorsinthefinal
74
measurements.
75
76
Onestrategytoaddressthelackofin-depthvalidationofimageanalysispipelinewouldbetouse
77
syntheticimagesgeneratedbystructuralrootmodels(modelsdesignedtorecreatethephysical
78
structureandshapeofrootsystems).Manystructuralrootmodelshavebeendeveloped,eitherto
79
modelspecificplantspecies(Pagèsetal.,1989),ortobegeneric(Pagèsetal.,2004;2013).These
80
modelshavebeenrepeatedlyshowntofaithfullyrepresenttherootsystemstructure(Pagèsand
5
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
81
Pellerin,1996).Inaddition,theycanprovidetheground-truthdataforeachsyntheticrootsystem
82
generated,independentlyofitscomplexity.However,exceptonerecenttooldesignedforyoung
83
seedlingswithnolateralroots(Benoitetal.,2014).theyhavealmostneverbeenusedforvalidation
84
ofimageanalysistools(Rellán-Álvarezetal.,2015).A
85
86
Herewe(i)illustratetheuseofastructuralrootmodel,Archisimple,tosystematicallyanalyseand
87
evaluateanimageanalysispipelineand(ii)evaluatetheusefulnessofdifferentrootmetrics
88
commonlyusedinplantrootresearch.
89
6
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
90
Material and methods
91
Nomenclature used in the paper
92
93
Ground-truthdata:Thereal(geometricandmorphometric)propertiesoftherootsystemasa
94
biologicalobject.Determinedbyeithermanualtracingofrootsorbyusingtheoutputofmodelled
95
rootsystems.
96
(Image)Descriptor:Propertyoftherootimage.Doesnotnecessarilyhaveabiologicalmeaning.
97
Synthetype:Foreachsimulation,aparametersetisdefinedrandomly.Then,10rootsystemsare
98
created.Sincethemodelhasanintrinsicvariability,eachoftheserootsystemisslightlydifferent
99
fromtheothers,althoughsimilar,formingwhatwecalledasyntheticgenotype,orsynthetype.
100
Rootaxes:firstorderroots,directlyattachedtotheshoot
101
Lateralroot:second(orlower)orderroots,attachedtoanotherroot
102
Creation of a root system library
103
WeusedthemodelArchiSimple,whichwasshowntoallowgeneratingalargediversityofroot
104
systemswithaminimalamountofparameters(Pagèsetal.,2013).Inordertoproducealarge
105
libraryofrootsystems,weranthemodel10.000times,eachtimewitharandomsetofparameters.
106
107
Thesimulationsweredividedintwomaingroups:monocotsanddicots.Forthemonocot
108
simulations,themodelgeneratedarandomnumberoffirst-orderaxesandsecondary(radial)
109
growthwasdisabled.Fordicotsimulations,onlyoneprimaryaxiswasproducedandsecondary
7
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
110
growthwasenabled(theextendofwhichwasdeterminedbyarandomparameter).Forall
111
simulation,onlyfirstorderlateralswerecreated,tolimitcomplexity.
112
113
TherootsystemcreatedfromeachsimulationwasstoredinanRSMLfile.EachRSMLfilewasthen
114
readbytheRSMLReaderpluginfromImageJtoextractmetricsandgenerateground-truthdatafor
115
thelibrary(Lobetetal.,2015).Theseground-truthdataincludedgeometrical,morphologicaland
116
topologicalparameters(table1).ForeachRSMLdatafile,theRSMLReaderpluginalsocreateda
117
PNGimage(ataresolutionof300DPI)oftherootsystem.
118
119
Table1:Rootsystemparametersusedasground-truthdata
Name
Description
Unit
tot_root_length
Thecumulativelengthofallroots
mm
tot_prim_length
Thecumulativelengthofallrootaxes
mm
tot_lat_length
Thecumulativelengthofalllateralroots
mm
mean_prim_length
Themeanfirst-orderrootslength
mm
mean_lat_length
Themeanlateralrootlength
mm
n_primary
Thetotalnumberoffirstorderroots
-
n_laterals
Thetotalnumberoflateralroots
-
Themeanlateralrootdensity:foreachfirst-orderoot,
thenumberoflateralrootsdividedbytheaxislength
120
mean_lat_density
(totallength).
mm-1
mean_prim_diam
Themeandiameterofthefirst-orderroots
mm
mean_lat_diam
Themeandiameterofthelateralroots
mm
mean_lat_angle
Themeaninsertionangleofthelateralroots
°
8
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
121
Root image analysis
122
Eachgeneratedimagewasanalysedusingacustom-madeImageJplugin,RootImageAnalysis-J(or
123
RIA-J).ThesourcecodeofRIA-J,aswellasacompiledversionisavailableattheaddress:
124
https://zenodo.org/record/61509.
125
126
Foreachimage,weextractedasetofclassicalrootimagedescriptors,suchasthetotalrootlength,
127
theprojectedareaorthenumberofvisibleroottips.Inaddition,weincludedshapedescriptors,
128
suchaspseudo-landmarks,ora-dimensionalmetricssuchastheexplorationratio,ofthewidth
129
proportionat50%depth(seeSupplementalfile1fordetailsabouttheshapedescriptors).Thelist
130
ofmetricsandalgorithmsusedbyourpipelineislistedinthetable2.
131
Data analysis
132
DataanalysiswasperformedinR(RCoreTeam).Morphometricanalyseswereperformedusingthe
133
momocs(Bonhommeetal.,2014)andshapes(Dryden,2015)packages.Plotswerecreatedusing
134
ggplot2(Wickham,2009)andlattice(Sarkar,2008).
135
TheRelativeRootSquareMeanErrors(RRSME)wereestimatedusingtheequation:
!
𝑅𝑅𝑀𝑆𝐸 =
(𝑦𝚤 − 𝑦𝑖)
𝑦𝚤
𝑛
136
where𝑛isthenumberofobservations,𝑦𝚤isthemeanand 𝑦𝑖 istheestimatedmean.
137
TheLinearDiscriminantAnalysis(LDA)wasperformedusingtheldafunctionfromtheMASS
138
package(MandD,2002).Foreachanalysis,weusedthesynthetypeinformationasgroupingfactor.
139
Weusedhalfofthesamples(5)ofeachsynthetypetobuildthemodelandtheotherhalftoassess
140
thediscriminantpoweroftheeachclassofmetrics(morphologyandshape).
9
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
141
Data availability
142
Alldatausedinthispaper(includingtheimageandRSMLlibraries)areavailableattheaddress
143
https://zenodo.org/record/61739
144
Anarchivedversionofthecodesusedinthispaperisavailableattheaddress
145
https://zenodo.org/record/152083
146
10
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
147
Results and discussions
148
Production of a large library of ground-truthed root system images
149
Wecombinedexistingtoolsintoasinglepipelinetoproducealargelibraryofground-truthedroot
150
systemimages.Thepipelinecombinesarootmodel(ArchiSimple(Pagèsetal.,2013)),theRoot
151
SystemMarkupLanguage(RSML)andtheRSMLReaderpluginfromImageJ(Lobetetal.,2015).In
152
short,ArchiSimplewasusedtocreatealargenumberofrootsystems,basedonrandominput
153
parametersets.EachoutputwasstoredasanRSMLfile(fig.2A),whichwasthenusedbytheRSML
154
Readerplugintocreateagraphicalrepresentationoftherootsystem(asa.jpegfile)andaground-
155
truthdataset(fig.2B).DetailsaboutthedifferentstepsarepresentedintheMaterialsandMethods
156
section.
157
158
11
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
159
Weusedthepipelinetocreatealibraryof10,000rootsystemimages,separatedintomonocots
160
(multiplefirstorderrootsandnosecondarygrowth)anddicots(onefirstorderrootandsecondary
161
growth).Foreachinputparameter-setusedforArchiSimple(1.000differentones),10repetitions
162
wereperformedtocreatesyntheticgenotypes,orsynthetypes(fig.2A).Thesynthetyperepetitions
163
weredonesuchasthestructureofthefinaldatasetwouldmimicthestructureofadataset
164
containingphenotypicdataofdifferentgenotypes.Therangesofthedifferentground-truthdataare
165
shownintable2andtheirdistributionisshownintheSupplementalFigure1.Thepipeline
166
producedperfectlythresholdedblackandwhiteimagesandhencethefollowinganalyseswere
167
focusedonthecharacterisationoftherootobjectsthemselves.
168
169
Westartedbyevaluatingwhethermonocotsanddicotsshouldbeseparatedduringtheanalysis.We
170
performedaPrincipalComponentAnalysisontheground-trutheddatasettoassessifthespecies
171
groupinghadaneffectontheoveralldatasetstructure(fig.3A).Monocotsanddicotsformed
172
distinctgroups(MANOVAp-value<0.001),withonlyminimaloverlap.Thefirstprincipal
173
component,thatrepresented33.2%ofthevariationwithinthedataset,wasmostlyinfluencedby
174
thenumberofprimaryaxes.Thesecondprincipalcomponent(19.6%ofthevariation)was
175
influenced,inpart,bytherootdiameters.Thesetwoeffectswereconsistentwiththecleargrouping
176
ofmonocotsanddicots,sincetheyexpressedthemaindifferencebetweenthetwospecies.
177
Therefore,sincethespeciesgroupinghadsuchastrongeffectontheoverallstructure,wedecided
178
toanalysethemseparatelyratherthantogetherforthefollowinganalyses.
179
12
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
180
Table3:Rangesofthedifferentground-truthdatafromtherootsystemsgeneratedusing
181
ArchiSimple
variable
minimumvalue
maximumvalue
unit
MONOCOTS
tot_root_length
8.36
2455.03
cm
width
0.25
33.21
cm
depth
5.49
37.5
cm
n_primary
1
20
-
tot_prim_length
6
327
cm
mean_prim_length
3.22
38
cm
mean_prim_diameter
0.02
0.04
cm
mean_lat_density
0
100.88
cm
n_laterals
0
1378
-
tot_lat_length
0
1630
cm
mean_lat_length
0
4.44
cm
mean_lat_diameter
0
0.03
cm
mean_lat_angle
0
97.74
°
DICOTS
182
tot_root_length
6.91
585.05
cm
width
0.01
15.05
cm
depth
3.89
36.99
cm
n_primary
1
1
-
tot_prim_length
4
37
cm
mean_prim_length
4.4
37.5
cm
mean_prim_diameter
0.02
1.13
cm
mean_lat_density
0
494.54
cm
n_laterals
0
277
-
tot_lat_length
0
437
cm
mean_lat_length
0
5.48
cm
mean_lat_diameter
0
0.23
cm
mean_lat_angle
0
87.63
°
13
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
183
14
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
184
Systematic evaluation of root image descriptors
185
Inordertodemonstratetheutilityofasyntheticlibraryofground-truthedrootsystems,we
186
analysedeveryimageofthelibraryusingacustom-builtrootimageanalysistool,RIA-J.Wedecided
187
todosobecauseourpurposewastotesttheusefulnessofthesyntheticanalysisandnottoassess
188
theaccuracyofexistingtools.Nonetheless,RIA-Jwasdesignedusingknownandpublished
189
algorithms,oftenusedinrootsystemquantification.AdetaileddescriptionofRIA-Jcanbefoundin
190
theMaterialsandMethodssection.
191
192
Table 3: Root image descriptors extracted by RIA-J
Name
Description
Unit Reference
area
Projectedareaoftherootsystem
mm2 (Galkovskyietal.,2012)
length
Lengthoftheskeletonoftherootsystemimage
mm (Galkovskyietal.,2012)
tip_count
Numberofendbranchesintherootsystemskeleton
-
diam_mean
Meandiameteroftherootobjectintheimage
mm width
Themaximalwidthoftherootsystem
mm -
depth
Themaximaldepthoftherootsystem
mm -
width_depth
Ratiobetweenthewidthandthedepthoftherootsystem
-
(Galkovskyietal.,2012)
com_x-com_y Relativecoordinatesofthecenterofmassoftherootsystem
-
(Galkovskyietal.,2012)
convexhull
Areaofthesmallestconvexshapearoundtherootsystem
mm2 (Galkovskyietal.,2012)
exploration
Ratiobetweentheconvexhullareaandtheprojectedarea
-
(Galkovskyietal.,2012)
(ChitwoodandOtoni,
FirstthreePrincipalComponentsofthemorphometricanalysis
PL.PC1-3
usingpseudo-landmarks(seeSupplementalfile1fordetails)
2016;Rellán-Álvarezetal.,
-
2015;Ristovaetal.,2013)
-
(Buckschetal.,2014)
-
(Buckschetal.,2014)
Relativedepthatwhich50%ofthecumulativewidthofthe
width50
rootsystemisreached(seeSupplementalfile1fordetails)
Relativedepthatwhich50%ofthetotalnumberofrootsis
count50
193
reached
15
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
194
Weextracted16descriptorsfromeachrootsystemimage(Table3)andcomparedthemwiththeir
195
ownground-truthdata.Foreachpairofdescriptor/data,weperformedalinearregressionand
196
computeditsr-squaredvalue.Figure4showstheresultsfromthedifferentcombinationsforboth
197
monocotsanddicots.Wecanobservethat,asageneralrule,goodcorrelationswererare,withonly
198
3%ofthecombinationshavinganr-squaredabove0.8.Inaddition,evenagoodcorrelationisnot
199
necessarilydirectlyusefulastherelationshipbetweenthetwovariablesmightnotfollowa1:1rule
200
(fig.4B-C).Insuchcase,anadditionalvalidationmightbeneededtodefinetherelationbetween
201
bothvariables.
202
203
Italsohastobenotedthatthecorrelationsweredifferentbetweenspecies.Asanexample,within
204
thedicotdataset,nogoodcorrelationwasfoundbetweenthetip_countanddiam_meanestimators
205
whilebettercorrelationwasfoundforthemonocots.Asaconsequence,validationofthedifferent
206
imageanalysisalgorithmsshouldbeperformed,atleast,foreachgroupofspecies.Analgorithm
207
givinggoodresultsforamonocotmightfailwhenappliedondicotrootsystemanalysis.
16
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
208
17
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
209
Errors from image descriptors are likely to be non linear
210
Inadditiontobeingrelatedtothespeciesofstudy,estimationerrorsarelikelytoincreasewiththe
211
rootsystemsize.Astherootsystemgrowsanddevelops,thecrossingandoverlappingsegments
212
increase,makingthesubsequentimageanalysispotentiallymoredifficultandpronetoerror.
213
However,asystematicanalysisofsucherrorisseldomperformed.
214
215
Figure5showstherelationshipbetweentheground-truthanddescriptorvaluesforthree
216
parameters:thetotalrootlength(fig.5A),thenumberofroots(fig.5B)andtherootsystemdepth
217
(fig.5C).Foreachofthesevariables,wequantifiedtheRelativeRootMeanSquareError(see
218
MaterialsandMethodsfordetails)asafunctionofthetotalrootlength.Wecanobservethatforthe
219
estimationofboththetotalrootlengthandthenumberoflateralroots,theRelativeRootSquare
220
MeanErrorincreasedwiththesizeoftherootsystem(fig.5A-B).Asstatedabove,suchincreaseof
221
theerrorwassomehowexpectedwithincreasingcomplexity.Forothermetrics,suchastheroot
222
systemdepth,noerrorswereexpected(depthissupposedlyanerror-lessvariable)andtheRelative
223
RootMeanSquareErrorwascloseto0whateverthesizeoftherootsystem.
224
225
Suchresultsareacalltocautionwhenanalysingrootimagesasunexpectederrorsindescriptors
226
estimationcanarise.Thisisprobablyevenmoretruewithrealimages,thataresusceptibleto
227
containnon-rootobjects(e.g.dirt)andlowerorderlateralsroots(asstatedabove,simulationsused
228
herewerelimitedtofirstorderlaterals).
18
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
229
230
231
19
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
232
Differentiation power differs between metrics
233
Finally,wewantedtoevaluatewhichmetricswerethemostusefultodiscriminatebetweenroot
234
systemsofdifferentgenotypesorexperimentalseries(controlvstreatment).Asexplainedabove,
235
foreachparametersetusedintheArchiSimplerunforlibraryconstruction,wegenerated10root
236
systems.Giventheintrinsicvariabilityexistinginthemodel,eachofthese10rootsystemswere
237
similaralthoughdifferent,ascouldbeexpectedfromplantsofthesamegenotype.Theseso-called
238
synthetypes,werethenusedtoevaluatehowefficientwerethedifferentmetricstodiscriminate
239
them.
240
241
Toestimatethedifferentiationoftheimagemetrics,weusedaLinearDiscriminantAnalysis(LDA)
242
predictionmodel.Foreachsynthetype,halfoftheplantswereusedtocreatetheLDAmodel.The
243
modelwasthenusedtopredictasynthetypefortheremaininghalfoftheplants.Thisapproach
244
allowedustoevaluatethepredictionaccuracy,ordifferentiationpower,ofthedifferentmetrics.A
245
predictionaccuracyof100%meansthatallplantswerecorrectlyassignedtotheirsynthetype.To
246
evaluatethedifferentiationpowerofsinglemetrics,weusedanapproachinwhicheachmetricwas
247
iterativelyaddedtothemodel,basedonthemodelglobalpredictionpower(seeSupplemental
248
Figure3fordetailsabouttheprocedure).Weperformedtheanalysiseitheronafulldataset(fig.
249
6D-E),oronadatarestrictedtothesmallestplants(fig.6A),inordertotesttheinfluenceofthe
250
underlyingdatastructure.
251
252
Twomainobservationscanbemadeonthefigure6.First,forthreeoutoffourscenarios,only5(or
253
less)descriptorswereneededtoachieveadifferentiationaccuracyof90%.Depth,areaandlength
254
werethemostimportantdescriptorsinalmostallscenarios.Theremainingdescriptorsdidnot
255
increasesignificantlytheaccuracy(somemightevendecreaseit).Thismightbeinterpretedasa
20
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
256
handfulofvariablesweresufficienttodistinguishsynthetypes,andbyextensiongenotypesor
257
treatedplants.However,wecanalsoobservethatthemostimportantparameterschanged
258
dependingontheunderlyingdatastructure(eitherduetospeciesorthesizeofthedataset).This
259
indicatesthatitisdifficulttohaveanapriorievaluationoftheimportantvariables.Keepingas
260
manyvariableapossiblemightalwaysbethemostefficientsolution.
261
21
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
262
22
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
263
Conclusions
264
Theautomatedanalysisofrootsystemimagesisroutinelyperformedinmanyresearchprojects.
265
Hereweusedalibraryof10.000modelledimagestoestimatetheaccuracyandusefulnessof
266
differentimagedescriptorsextractedwithanhome-maderootimageanalysispipeline.Theanalysis
267
highlightedsomeimportantlimitationsduringtheimageanalysisprocess.
268
269
Firstly,generalstructureoftherootsystem(e.gmonocotvsdicots)canhaveastronginfluenceon
270
thedescriptorsaccuracy.Descriptorsthathavebeenshowntobegoodpredictorsforonetypeof
271
rootsystemsmightfailforanothertype.Insomecases,thecalibrationandthecombinationof
272
differentdescriptorsmightimprovetheaccuracyofthepredictions,butthisneedstobeassessed
273
foreachanalysis.
274
275
Asecondfactorinfluencingstronglytheaccuracyoftheanalysisistherootsystemsizeand
276
complexity.Asageneralrule,formorphologicaldescriptors,thelargertherootsystem,thelarger
277
theerroris.Sofar,alargeproportionoftherootresearchhasbeenfocusedonseedlingswithsmall
278
rootsystemsandhavedefactoavoidedsucherrors.However,astheresearchquestionsarelikely
279
tofocusmoreonmaturerootsysteminthefuture,theselimitationswillbecomecritical.
280
281
Finallywehaveshownthatnotallmetricshavethesamebenefitwhencomparinggenotypeor
282
treatments.Again,dependingontherootsystemtypeorsize,differentmetricswillhavedifferent
283
differentiationpowers.
284
285
Itisimportanttohighlightthattheimagesusedinouranalysiswereperfectlythresholded,without
286
anydegradationintheimagequality.Therefore,theerrorscomputedinouranalysisarelikely
23
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
287
under-estimatedcomparedtorealimages(withadditionalbackgroundnoiseandlesserquality).
288
Sincethequalityoftheimagesisdependentontheunderlyingexperimentalsetup,artificialnoise
289
couldbeaddedtothegeneratedimagesinordertomimicanyexperimentallyinducedartifactand
290
toimprovetheanalysispipelineevaluation,asproposedby(Benoitetal.,2014).
291
292
Toconclude,ourstudyisareminderthatthoroughcalibrationsareneededforrootimageanalysis
293
pipelines.Herewehaveusedalargelibraryofsimulatedrootimages,thatwehopewillbehelpful
294
fortherootresearchcommunitytoevaluatecurrentandfutureimageanalysispipelines.
295
296
24
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
297
Conflict of Interest
298
Theauthorsdeclarethattheresearchwasconductedintheabsenceofanycommercialorfinancial
299
relationshipsthatcouldbeconstruedasapotentialconflictofinterest.
300
Author Contributions
301
GL,LP,PTandCPdesignedthestudy.IKdevelopedtheimageanalysispipelineRIA-J.GLgenerated
302
theimagelibrary,didtheimageanalysisanddataanalysis.LPdevelopedtheArchiSimplemodel.All
303
authorshaveparticipatedinthewritingofthemanuscript.
304
Funding
305
ThisresearchwasfundedbytheInteruniversityAttractionPolesProgrammeinitiatedbythe
306
BelgianSciencePolicyOffice,P7/29.GLisgratefultotheF.R.S.-FNRSforapostdoctoralresearch
307
grant(1.B.237.15F).
308
Supplementary Material
309
-Supplementalfigure1:Distributionofthepropertiesofthemodelledrootimages
310
-Supplementalfigure2:Distributionofthedescriptorsofthemodelledrootimages
311
-Supplementalfigure3:Workflowusedfortheaccuracyanalysis
312
-Supplementalfile1:Defintionsoftheshapedescriptors
313
25
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
314
References
315
Armengaud,P.,Zambaux,P.,Hills,A.,Sulpice,R.,Pattison,R.J.,Blatt,M.R.,etal.(2009).EZ-Rhizo:integrated
316
317
softwareforthefastandaccuratemeasurementofrootsystemarchitecture.PlantJ57,945–956.
Benoit,L.,Rousseau,D.,Belin,É.,Demilly,D.,andChapeau-Blondeau,F.(2014).Simulationofimage
318
acquisitioninmachinevisiondedicatedtoseedlingelongationtovalidateimageprocessingroot
319
segmentationalgorithms.ComputersandElectronicsinAgriculture104,84–92.
320
doi:10.1016/j.compag.2014.04.001.
321
Bonhomme,V.,Picq,S.,andGaucherel,C.(2014).Momocs:outlineanalysisusingR.JournalofStatistical….
322
Bucksch,A.,Burridge,J.,York,L.M.,Das,A.,Nord,E.,Weitz,J.S.,etal.(2014).Image-basedhigh-throughput
323
324
fieldphenotypingofcroproots.doi:10.1104/pp.114.243519.
Chitwood,D.H.,andOtoni,W.C.(2016).MorphometricanalysisofPassifloraleavesI:therelationship
325
betweenlandmarksofthevasculatureandellipticalFourierdescriptorsoftheblade.
326
doi:10.1101/067512.
327
Dryden,I.L.(2015).shapespackage.Vienna,Austria.
328
Galkovskyi,T.,Mileyko,Y.,Bucksch,A.,Moore,B.,Symonova,O.,Price,C.A.,etal.(2012).GiARoots:software
329
forthehighthroughputanalysisofplantrootsystemarchitecture.BMCPlantBiol12,116.
330
doi:10.1186/1471-2229-12-116.
331
Koevoets,I.T.,Venema,J.H.,Elzenga,J.T.M.,andTesterink,C.(2016).RootsWithstandingtheir
332
Environment:ExploitingRootSystemArchitectureResponsestoAbioticStresstoImproveCrop
333
Tolerance.FrontPlantSci07,91–19.doi:10.3389/fpls.2016.01335.
334
Lobet,G.,Pound,M.P.,Diener,J.,Pradal,C.,Draye,X.,Godin,C.,etal.(2015).RootSystemMarkupLanguage:
335
TowardaUnifiedRootArchitectureDescriptionLanguage.PlantPhysiol167,617–627.
336
doi:10.1104/pp.114.253625.
337
M,V.W.,andD,R.B.(2002).ModernAppliedStatisticswithS
338
.New-York:Springer.
339
Pagès,L.,andPellerin,S.(1996).Studyofdifferencesbetweenverticalrootmapsobservedinamaizecrop
340
andsimulatedmapsobtainedusingamodelforthethree-dimensionalarchitectureoftherootsystem.
341
PlantandSoil182,329–337.
26
bioRxiv preprint first posted online Sep. 14, 2016; doi: http://dx.doi.org/10.1101/074922. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
Lobetetal.2016-Usingmodelstoevaluaterootimageanalysistools
342
Pagès,L.,Bécel,C.,Boukcim,H.,Moreau,D.,Nguyen,C.,andVoisin,A.-S.(2013).Calibrationandevaluationof
343
ArchiSimple,asimplemodelofrootsystemarchitecture.EcologicalModelling290,76–84.
344
doi:10.1016/j.ecolmodel.2013.11.014.
345
Pagès,L.,Jordan,M.O.,andPicard,D.(1989).Asimulationmodelofthethree-dimensionalarchitectureofthe
346
347
maizerootsystem.PlantandSoil119,147–154.doi:10.1007/BF02370279.
Pagès,L.,Vercambre,G.,Drouet,J.-L.,Lecompte,F.,Collet,C.,andLeBot,J.(2004).RootTyp:agenericmodel
348
349
todepictandanalysetherootsystemarchitecture.PlantandSoil258,103–119.
Pierret,A.,Gonkhamdee,S.,Jourdan,C.,andMaeght,J.-L.(2013).IJ-Rhizo:anopen-sourcesoftwareto
350
measurescannedimagesofrootsamples.PlantandSoil,1–9.doi:10.1007/s11104-013-1795-9.
351
RCoreTeamR:ALanguageandEnvironmentforStatisticalComputing.
352
Rellán-Álvarez,R.,Lobet,G.,Lindner,H.,Pradier,P.-L.,Sebastian,J.,Yee,M.-C.,etal.(2015).GLO-Roots:an
353
imagingplatformenablingmultidimensionalcharacterizationofsoil-grownrootsystems.eLife4,
354
e07597.doi:10.7554/eLife.07597.
355
Ristova,D.,Rosas,U.,Krouk,G.,Ruffel,S.,Birnbaum,K.D.,andCoruzzi,G.M.(2013).RootScape:Alandmark-
356
basedsystemforrapidscreeningofrootarchitectureinArabidopsisthaliana.
357
doi:10.1104/pp.112.210872.
358
Sarkar,D.(2008).Lattice:MultivariateDataVisualizationwithR.NewYork:Springer.
359
Wickham,H.(2009).ggplot2.NewYork,NY:SpringerNewYorkdoi:10.1007/978-0-387-98141-3.
360
27