structure et fonction du cerveau - Université de Sherbrooke [PDF]

U niversité de Sherbrooke

S T R U C T U R E ET F O N C T IO N D U C E R V E A U : L IE N E N T R E LA D E N S IT É V A S C U L A IR E E T L ’A M P L IT U D E D U S IG N A L B O L D

p ar N icolas V igneau-R oy D ép artem en t de m édecine nucléaire et radiobiologie

M émoire présenté à la Faculté de m édecine et des sciences de la san té en vue de l’o b ten tio n du grade de m a ître ès sciences (M .Sc.) en Sciences des rad ia tio n s et im agerie biom édicale

Sherbrooke, Q uébec, C a n a d a 14 sep tem b re 2013

M em bres d u J u ry d ’évalu atio n Kevin W liittin g stall. D ép artem en t de radiologie M axime D escoteaux. D épartem ent d 'inform atique M artin Lepage. D épartem ent de m édecine nucléaire et radiobiologie Pierre-M ichel Bernier, Faculté d 'éd u c a tio n physique et sportive

@ Nicolas Vigneau-Roy, 2013

1+1

Library and Archives Canada

Bibliothèque et Archives Canada

Published Héritage Branch

Direction du Patrimoine de l'édition

395 Wellington Street Ottawa ON K1A0N4 Canada

395, rue Wellington Ottawa ON K1A 0N4 Canada Your file Votre référence ISBN: 978-0-499-00378-2 Our file Notre référence ISBN:

978-0-499-00378-2

NOTICE:

AVIS:

The author has granted a nonexclusive license allowing Library and Archives Canada to reproduce, publish, archive, preserve, conserve, communicate to the public by télécommunication or on the Internet, loan, distrbute and sell theses worldwide, for commercial or noncommercial purposes, in microform, paper, electronic and/or any other formats.

L'auteur a accordé une licence non exclusive permettant à la Bibliothèque et Archives Canada de reproduire, publier, archiver, sauvegarder, conserver, transmettre au public par télécommunication ou par l'Internet, prêter, distribuer et vendre des thèses partout dans le monde, à des fins commerciales ou autres, sur support microforme, papier, électronique et/ou autres formats.

The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission.

L'auteur conserve la propriété du droit d'auteur et des droits moraux qui protégé cette thèse. Ni la thèse ni des extraits substantiels de celle-ci ne doivent être imprimés ou autrement reproduits sans son autorisation.

In compliance with the Canadian Privacy Act some supporting forms may have been removed from this thesis.

Conform ém ent à la loi canadienne sur la protection de la vie privée, quelques formulaires secondaires ont été enlevés de cette thèse.

W hile these forms may be included in the document page count, their removal does not represent any loss of content from the thesis.

Bien que ces formulaires aient inclus dans la pagination, il n'y aura aucun contenu manquant.

Canada

S tr u c tu r e e t fo n c tio n d u c e r v e a u : L ie n e n tr e la d e n s ité v a sc u la ir e r é g io n a le e t l ’a m p litu d e d u s ig n a l B O L D P ar Nicolas V igneau-R ov P rogram m e de Sciences des rad iatio n s et im agerie biom édicale M émoire présenté à la Faculté de m édecine et des sciences de la san té en vue de l’obtention du diplôm e de m aître ès sciences (M .Sc.) en Sciences des rad iatio n s et imagerie biom édicale, Faculté de m édecine et des sciences de la santé, U niversité de Sherbrooke, Sherbrooke, Q uébec, C an ad a, J1H 5N4

L’Im agerie p ar Résonance M agnétique fonctionnelle est devenue u n des outils principaux p our m esurer l’activ ité cérébrale chez l’H om m e de façon non-invasive. Le type de séquence plus fréquem m ent utilisé à ces fins est le signal d ép en d an t d u niveau d ’oxvgène sanguin (BOLD signal) car il reflète les changem ents hém odynam iques associés à l’activité cérébrale et offre l’avantage d ’être relativem ent facile à m esurer. C ependant, cette m esure est d ép en d an te d u niveau de désoxvhém oglobine, pré­ sent dans les veines. L ’architectu re stru ctu relle de ces dernières varie considérablem ent d ’une région du cerveau à l ’a u tre et cela fait en so rte q u ’il est a rd u de savoir si les différences inter-régionales sont dues à des différences au niveau de l’activ ité neuronale ou de la stru c tu re vasculaire. C ’est p o u r rép o n d re à cette question que nous avons observé la relation entre les variations régionales de densité vasculaire (VAD) et l’am plitude du signal BOLD au repos et en présence d ’un stim ulus. Nous avons im pléinenté une m éthode de segm entation des veines p o u r des images en p o n d ératio n de susceptibilité (SW I). Nous avons égalem ent utilisé une m éth o d e de segm entation des tissus (M orphom étrie Basée sur les Voxels (V B M )) p o u r séparer les tissus et cal­ culer une m esure de densité de m atière. Nous avons tro u v é q u ’in d épendam m ent de la taille des voxels, les variations régionales d ’am p litu d e du signal BO LD au repos et du signal évoqué par une tâche sont m ieux corrélées avec la VAD q u ’avec la densité de m atière grise. En utilisant un m odèle linéaire général, nous avons observé q u 'u n e grande p artie des variations régionales au repos pouvait être modélisée p a r la VAD. En somme, nos résultats suggèrent que le signal BO LD au repos est intim em ent relié à la stru c tu re vasculaire. U tiliser la densité vasculaire afin de calibrer les m e­ sures d'am p litu d e du signal BOLD au repos nous p e rm e ttra d 'in te rp ré te r de façon plus précise et d ’obtenir de m eilleures inform ations sur les différences observées entre plusieurs groupes de sujets et de patients. M o ts-clés: IRM ; IRM f ; SW I ; Segm entation ; VBM ; D ensité vasculaire.

S tr u c tu r e a n d fu n c tio n o f t h e b ra in : L in k b e tw e e n r é g io n a l v a scu la r d e n s ity a n d B O L D s ig n a l a m p litu d e By Nicolas V igneau-R oy G rad u ate Program of R ad iatio n Sciences an d Biom édical Im aging M aster's thesis presented to th e Faculty of M edecine an d H ealth Sciences to o b tain the title of m aître ès sciences (M .Sc.) in R ad iatio n Sciences and Biom édical Imaging. Faculty of M edecine and H ealth Sciences, U niversité de Sherbrooke, Sherbrooke, Q uébec, C an ad a, J1H 5N4

Functional M agnetic Résonance Im aging (fM RI) lias becom e one of the p rim arv tools used for non-invasivelv m easuring b rain activ itv in hum ans. For th e m ost p a rt, the blood oxvgen level-dependent (BO LD ) c o n tra st is used, which reflects changes in hem odynam ics associated witli active brain tissue. T he m ain advantage of th e BO LD signal is th a t it is relativelv easy to m easure an d th u s is often used as a proxy for com paring brain function across p o p u latio n groups (i.e. control vs. p atien t). However. it is particularlv w eighted tow ards veins w hose stru c tu ra l arch itectu re is known to varv considerablv across th e brain. T his m akes it difficult to in terp ret w hether différences in BOLD betw een cortical areas reflect tru e différences in neural activitv or vascular structure. We therefore investigated how régional variations of vascular density (VAD) relate to th e am p litu d e of restin g -state an d task-evoked BOLD signais. To address this, we first developed an au to m a ted m eth o d for segm enting veins in images acquired w ith susceptibilitv w eighted im aging (SW I), allowing us to visualize th e venous vascular tree across th e brain. In 19 h ealth v subjects, we then applied Voxel-Based M orphom etry (VBM ) to T l-w eig h ted im ages and com puted régional m easures of grav m a tte r density (G M D ). We found th a t, independent of spatial scale, régional variations in restin g -state an d task-evoked fM RI am p litu d e was b e tte r correlated to VAD com pared to GM D. U sing a général linear m odel (G LM ), it was observed th a t th e bulk of régional variance in re stin g -state activ itv could be modelled by VAD. Taken together, our resuit s suggest th a t re stin g -state BOLD signais are signifîcantlv related to the underlving stru c tu re of th e brain vascular svsteni. C alib ratin g resting BOLD activitv by venous stru c tu re m ay resuit in a m ore a ccu rate in te rp ré ta ­ tion of différences observed betw een cortical areas a n d /o r individuals. K eyw ord s: M RI ; fM RI ; SW I ; S egm entation ; VBM ; V ascular Density. ii

D éd icace & rem erciem en ts Je tiens to u t d ’ab o rd à dédier ce m ém oire à Françoise, m a conjointe, qui m ’appuie depuis toutes ces années dans tous mes projets, dont cette m aîtrise, qui tire m ain ­ ten an t à sa fin. Je veux égalem ent rem ercier G abriel P oulin-L am arre p o u r m ’avoir encouragé quotidiennem ent, Samuel M ercier pour être u n modèle de vie, et G abriel C avanagh pour être un m arin. Je rem ercie plus sérieusem ent ces trois-là p o u r les nom breuses discussions anim ées d u ra n t lesquelles ils m 'o n t fait voir un point de vue extérieur au dom aine grâce auquel j ’ai p u faire fructifier mes recherches. Finalem ent, je rem ercie le professeur K evin W h ittin g sta ll de m ’avoir accueilli d an s son laboratoire et d ’avoir pris le pari de me prendre com m e son prem ier étu d ia n t à la m aîtrise en ta n t que professeur. Je veux rem ercier M ichaël Bernier et le professeur M axime D escoteaux pour leur aide au niveau inform atique, ainsi que to u te l ’équipe responsable de l ’acquisition d ’im age p ar R ésonance M agnétique p o u r l'o b te n tio n des données nécessaires à la réalisation du p ro jet.

E pigraphe “I arn so sm art ! I arri so sinart ! S-M -R -T ... I m ean S-M -A -R-T.”

- H om er Sim pson

Table des m atières D é d ic a c e & r e m e r c ie m e n ts

iii

É p ig r a p h e

iv

A b r é v ia tio n s

xi

In tr o d u c tio n

1

1

4

Le cer v e a u h u m a in 1.1 La stru ctu re du cerveau

......................................................................................

5

Cellule de hase : Le n e u r o n e .................................................................

10

1.1.2 Le réseau vasculaire [ i n ] ........................................................................

11

1.2 A ctivité c é r é b r a le ...................................................................................................

16

1.1.1

1.2.1

2

Notion de hase [ 'i i ] ..................................................................................

16

1.2.2 A ctivation suite* à un stim ulus [ j n ] ......................................................

18

1.2.0 Le cerveau à l’état de r e p o s .................................................................

19

1.4 C o n c lu sio n .................................................................................................................

20

L 'Im a g erie par R é so n a n c e M a g n é tiq u e

21

2.1

Principes de hase de H U M ..................................................................................

22

2.1.1

M agnétism e et p r o t o n s ...........................................................................

22

2.1.2

D étection de s i g n a l ..................................................................................

23

2.2 Séquences d ’im a g e rie '............................................................................................

25

2.2.1

Séquence en pondérât ion T | .................................................................

26

2.2.2

Séquence en pondération de suscept ibilit é ( S W I ) ..........................

28

2.2.3

Séquence B O L I).........................................................................................

30

v

TABLE DES MATIÈRES 2..'i 3

C o n c lu sio n ...............................................................................................................

32

A r tic le : L es v a ria tio n s r ég io n a le s d e d e n s ité v a sc u la ir e c o rr o ie n t avec les v a r ia tio n s r ég io n a le s d ’a m p litu d e d u sig n a l B O L D

33

3.1

A vant-propos

........................................................................................................

33

3.2

In tr o d u c tio n ...........................................................................................................

38

3.3

M c th o d s ..................................................................................................................

40

3.1

3.5

3.3.1

S u b j e c t s ..............................................................................................

40

3.3.2

Image Acquisit i o n .............................................................................

41

3.4.3

Task-induecd Functional M agnctic R ésonance Imaging ( 1-1MRI)

41

3.3.1

R esting-state Functional M agnctic R ésonance Imaging (RS-fM RI)

42

3.3.5

Stiscept ibility-W eighted Im aging (SW I ) ..................................

42

3.3.0

Image p r o c e s s in g .............................................................................

42

3.3.7

S tructural A n a ly s is ..................................................................................

42

3.3.8

Functional A n a ly s is ...................................................................................

43

3.3.!) S ta n d a rd iz a tio n ................................................................................

44

3.3.10 S tatistieal A n a ly s is .........................................................................

44

Resuit s ......................................................................................................................

46

3.4.1

S tructural R e s u lts ....................................................................................

46

3.4.2

Functional R e s u l t s ...................................................................................

46

D is c u s s io n ...............................................................................................................

48

3.5.1

C orrélation between régional m easures of GMD

W’MD and

BOLD a m p l i t u d e ....................................................................................

48

3.5.2 C orrélation between régional m easures of YAD and BOLD am ­ plitude 3.5.3 Im plication

49 ................................................................................................

51

3.6

C o n c lu sio n ............................................

52

3.7

R é fé re n c é s ...............................................................................................................

53

4

D is c u ss io n

72

5

C o n c lu sio n

75

vi

TABLE DES MATIÈRES A A r tic le : S e g m e n ta tio n d e v e in e s à p a r tir d ’im a g e en p o n d é r a tio n d e s u s c e p tib ilité

93

A .l

Avant-propos

........................................................................................................

93

A .2

In tr o d u c tio n ...........................................................................................................

97

A. 3

Met 1» x l s ..................................................................................................................

98

A .3.1

D ata a c q u isitio n .......................................................................................

98

A.3.2

C om puting the vcssclnoss m easure

................................................

99

R e s u lts ......................................................................................................................

102

A. 4.1

V a lid a tio n .................................................................................................

102

A. 1.2

H mua n vasciilat tire ( l.ô Tesla ) ...........................................................

103

A. 1.3

Rat v a s e u la tu r c .......................................................................................

103

D is c u s s io n ...............................................................................................................

103

A .5.1

T he Snail Veins Segm entation P r o g r a m .........................................

103

A .7.2

Im p lic a tio n s ..............................................................................................

105

A.6

C o n c lu sio n ...............................................................................................................

106

A .7

R e f e r m é e s ...............................................................................................................

107

A .4

A. 5

vi i

L iste des figures 1.1 Division du e-orte’X e-érébral on -1 régions distinctes. Référence : Modifiée1 à p a rtir de [i< >■',]. Pas de droit d 'a u to u r applicable..............................

6

1.2 Division du cerveau on aires de B rodm ann superposé aux divisions structurelles. Référence' : [•!'(]. Pas ele' droit el'auteur applicable1..... 1.3 Divisie»n élu e-orte’x e-érébral eu élu e-e-rve’le’t e'ti aire's feme-tiemnelle’s.

7 ©2]

(M pyright : © 2012 Xue-leus Xleu lie a l Art/Doe-teir S to e -k ................

8

1.4 Coupe- sagittale1élu ceTve'le't. Référe'ne-o : [7 l], Deauaine1 publie-.........

9

1.5

Le's t r o i s p ar ti e>s élu trente- c é r é b r a l : le1 m é s e n e é p h a l e 1 eut ble'U. le1 p o l i t ele1 Vare>le> ( o u ])r ot u b é r a n e e ’ a n n u l a i r e ’) eut r e m g e ’ e>t le> b u l b e ’ rue liielieui

eut jaune. Rélére'iiee' : MeielifieV à p a rtir ele’ © ,]. Deauaine’ public.

...

1.0 Strue-ture1 d'un neurone’. Deauaine' p u b l i e - .............................................

10 11

1.7 Le’s artère's e-érébrale's irrigneait le' e-e>rve’au eai a p p o rta n t élu sang neaif. Référeme-e’ : Encye jopéelie' Larousse1 (www.larousse’.fr). Mie-lied Saeanann. Deanaine- publie-............................................................................................... 1.8 Artère’ e-érébrale’ antériemre'. C opyright : H. Femrnié © 2000 1. 9

Artère1 cérébrale1 moyenne’. Copyright : H. Femrnié © 2000

1.10 Artère* e-érébrale’ peistérieMiiv. CMpyright : H. Femrnié © 2000

...

[72]

[ 72]

12 13

. . . .

14

...

14

[72]

1.11 Le‘S sinus elrainent le* sang usé pour le1 re’îemrneT au co eu r.................

15

1.12 Schéma de la re-acîioti e-ntre la fe-ruiinaiseui synaptiejtte d un axerne- e-î la deaidrite1 élu nemreme' suivant. Le>s ne’uretîransine'tîemrs ine lie i ueuit aux eanattx CCv, m,. i

! ! !, Pl. t I. !_']

R ésu m é : Au cours des dernières années, une m o d alité d 'im ag erie p a r R ésonance M agnétique a été développée afin de séparer les tissu s d u corps ay an t des propriétés m agnétiques différentes. M aintenant disponible dans la m a jo rité des scanner clinique, l'Im agerie en pondération de susceptibilité ( S WI ) utilise la différence de susceptibilité m agnétique des tissus pour extraire le gras, segm enter les tissus riches en fer, séparer la m atière blanche de la m atière grise. Une des habilités les plus im pressionnantes du S W I est d ’être capable de faire ressortir les veines d u cerveau des a u tre s tissu s (m atière grise, m atière blanche, liquide eérébrorachidien). D ans nos trav au x précédents , nous avons revisité un problèm e concernant l’intensité du signal B O LD dans les régions h a u te ­ m ent vascularisées en introduisan t une nouvelle m esure, la densité vasculaire, qui est facilem ent calculable à p a rtir d ’im age S W I. D ans cet article, nous présenterons un toolbox M ATLAB qui im plém ente une m éth o d e de segm entation de veines efficace à p a rtir d ’image en pondération de susceptibilité.

94

A Novel SPM Toolbox for Veins segmentation in Susceptibility Weighted Image N. Vigneau-Roy1, M. Bernier1, K. W hittingstall1,2 and M. Descoteaux3* 1 Department of Diagnostic Radiology, Faculty of Medicine and Health Science, Université de Sherbrooke, 12e Avenue Nord, Sherbrooke, QC, Canada, J1H 5N4 2 Division of Neurosurgery and Neuro-Oncology, Faculty of Medicine and Health Science, Université de Sherbrooke, 12e Avenue Nord, Sherbrooke, QC, Canada, J1H 5N4 3 Computer Science department, Faculty of Science, Université de Sherbrooke, 2500 Boulevard Université, Sherbrooke, QC, Canada, J1K 2R1 *To whom correspondance should be addressed Tel: +1 819-346-1110 poste 14647 E-mail address: [email protected]

Abstract In the last decade, susceptibility-weighted Magnetic Résonance Imaging has been developed and made available for a vast majority o f imaging Systems. Using the différence of phase caused by the différence in magnetic susceptibility of brain tissues, the Susceptibility-Weighted Imaging (SWI) sequence enhances the contrast between tw o tissues with différent

magnetic

susceptibilités. Depending of the chosen sequence, SWI can separate fat from tissue, gray matter from white matter, iron-laden tissues from background ones, amongst other such contrast. One of the most interesting feature of SWI is its capacity to separate veins from other tissues (gray matter, white matter, and even cerebrospinal fluid). In our previous work (VigneauRoy et al. 2013), we reinvestigated an old issue regarding the impact of the vasculature on the amplitude of the BOLD signal obtained via T2*-weighted imaging. We observed that finding the vascular density of the subject to correct the BOLD signal amplitude before interprétation and comparison between subject is of capital importance, both for resting-state fMRI (RS-fMRI) and task-evoked fMRI (T-fMRI), for they are strongly correlated in most région of the brain. The segmentation and quantification of veins is therefore essential to accurately understand the meaning of the observed variations and différences of amplitude o f the fMRI response. In this paper, we présent a MATLAB toolbox, developed as a SPM plugin, implementing the method to segment efficiently brain veins from SWI. The purpose of this toolbox is to make this method available to everyone who wishes to do vascular segmentation.

2

Introduction Magnetic Résonance Imaging (MRI) is a powerful imaging tool to explore the brain in a noninvasive way. Multitudes of imaging sequences have been developed to emphasize various tissue types or to detect brain activity (spin density weighted imaging, Tl-weighted, T2weighted, T2*-weighted, diffusion, functional, etc). Amongst them is a sequence based on susceptibility différences between tissues, whîch can isolate structures of interest, such as ironladen tissues or venous blood vessels: the Susceptibility-Weighted Imaging (SWI) MR protocol (Haacke et al. 2006). In our previous work (Vigneau-Roy et al. 2013), we reinvestigated an old issue regarding the impact of the vasculature on the amplitude of the BOLD signal obtained via T2*-weighted imaging. We observed, as (Menon et al. 1993; Ogawa et al. 1993; Bandettini and Wong 1995, 1997; Davis et al. 1998; Zhao et al. 2006; Jochimsen et al. 2010; Yu et al. 2012; Moon et al. 2013) also did, that in most régions of the human brain, finding the vascular density of the subject to correct the BOLD signal amplitude before interprétation and comparison between subject is of capital importance, both for resting-state fMRI (RS-fMRI) and task-evoked fMRI (T-fMRI), for they are strongly correlated. The segmentation and quantification of veins is therefore essential to accurately understand the meaning of the observed variations and différences of amplitude of the fMRI response. However, even if it is well-known that the veins of the brain are meaningful in the field of fMRI, there are few tools available for veins extraction purposes. One of the first to study the feasibility of segmenting the veins from the gray m atter in 1.5T MR images w ith the purpose to enhance the compréhension of BOLD fMRI were (Singh et al. 1995), when they used voxel intensity and temporal delay to better localize activation sites. In the médical image analysis community, (Frangi et al. 1998) proposed the vesselness measure to identify tubular structures in the brain which became quite popular for the enhancement and segmentation of blood vessel (Krissian 2000; Aylward and Bullitt 2002; Kao et al. 2003; Passât et al. 2006; Descoteaux et al. 2008; Koopmans et al. 2008; Liao et al. 2011; Xu et al. 2012). On the other hand, (Kao et al. 2003) proposed a combination o f différent opérations such as Independent Component Analysis (ICA) and Bayesian estimation to segment perfusion volumes, hence finding veins in the brain.

3

(Passât et al. 2006) proposed a hit-or-miss transform in Phase-constrast Magnetic Résonance Angiography to do an automatic détection of the veins. Finally, (Descoteaux et al. 2008) created a segmentation tool to isolate veins in proton density images as well as in angiographie volume. Based on Frangi's (Frangi et al. 1998) vesselness measure, the segmentation is done using a multi scale géométrie flow algorithm. The use of SWI acquisitions to enhance and find veins was discarded until recently. (Koopmans et al. 2008) worked on vessel segmentation using 3T and 7T susceptibility-weighted imaging. They used Frangi's vesselness measure to create their Utrecht vesselness filte r and combined it with vessel enhancing diffusion (VED) to enhance the effect of the vesselness filter in low signalto-noise ratio (SNR) régions thus eliminating artifacts while leaving the veins structure intact. Similarly, (Xu et al. 2012) used another combination of techniques to find veins, adding an additional step during volume acquisition to remove high-frequency background effect before applying the Vessel Enhancing Diffusion (VED) filter using the vesselness measure and voxel intensity to identify the structure of interest. In this paper, we présent a MATLAB SPM toolbox (Mathworks, Natick, USA) to segment veins from

SWI

inspired

by

Frangi's

vesselness

measure.

Using

SPM8-VBM8

toolbox

(http://dbm.neuro.uni-iena.de/vbm/) to display the results, our user-friendly tool gives the community the opportunity to easily segment veins when it is needed in their studies. Results are presented in both human and animal subjects, and compared to the implémentation of (Descoteaux et al. 2008). Finally, we will discuss why segmentation of the veins is im portant and can be applied in both clinical and fundamental research studies (Vigneau-Roy et al. 2013).

Methods 2.1 Data acquisition Nineteen (19) right-handed subjects (4 females, âges 20-33) were recruited fo r a previous study (Vigneau-Roy et al. 2013). Ail subjects were French natives speakers and had no psychiatrie or neurologie symptoms at the time of scanning or in the past. This study was performed according to the guidelines of the Internai Review Board of the Centre Hospitalier Universitaire de Sherbrooke (CHUS). We acquire SWI from each subject fo r veins segmentation purpose. Briefly, this imaging technique benefits from the susceptibility différences between neighboring tissues,

4

which causes différent relaxation times. Depending on the echo tim e chosen, it is possible to visualize large phase différences between tissues. Using those différences to create phase masks, it is possible to enhance certain tissues. It can thus be used to differentiate gray m atter from white matter, iron-laden tissues, venous blood vessels and other tissues which susceptibilités are différent from the background tissue (Haacke et al. 2006).

A 320 x 260

acquisition matrix was used with a TR/TE of 49/40ms, giving a voxel size o f 0.719 x 0.719 x 1.6mm.

2.2 Computing the vesselness measure The purpose of the toolbox is to create a volume representing the venous blood vessel of the brain from SWI. (Frangi et al. 1998) method is used to enhance veins in using the vesselness filter. Briefly, this measure looks for the local tubular shape of the structure computing the Hessian matrix at every voxel of the volume. The measure is done at multiple scales to enhance tubular vessels of différent sizes and can thus differentiate tube-like structures from blob-like and sheet-like shapes, and artifacts. In the spécifie case of SWI, it benefits from the dark contrast of the veins given by the phase enhancement. Note that the technique would still work for bright blood modalities such as angiography (see (Descoteaux et al. 2008)). Statistical Parametric Mapping (SPM8) toolbox (http://www.fil.ion.ucl.ac.uk/spm /software/spm 8/) is used to manage input, output and visualization of NIFTI images. 2.2.1 S e cond o rd e r d e riv a tiv e s The first step of the process is to detect the location of the veins. To do so, we create a derivative of Gaussian filter G(o)ôo . This kind of filter usually enhances the edges in a 2D/3D image. In the SWI volume, it will detect ail the dark voxels of the volume that are contrasting with their brighter neighbors (veins amongst tissue). Once applied in width, height and depth, we obtain a coarse version of the vasculature edges. To detect shape information, the filter is applied a second time, resulting in the second order derivatives (ôx2, ôxy, ôxz, ôyx, Ôy2, ôyz, ôzx, ôzy, ôz2). The we must construct the Hessian Matrix to retrieve the tube-like vascular structure from the second order derivatives. 2.2.2 H essian M atrix tu b u la r s tru c tu re s The Hessian matrix

5

/ a/2

di2

dx2 dl2

dxdy dl2

dxdz

dydx dl2

dy2 dl2

dydz dl2

\dzdx

dzdy

a/2 \ dl2

dz2 /

contains the local information about the shape of the structure at every voxel. From this matrix, it is possible to extract this information by computing its Eigenvalues and by combining them to find segments of différent shapes: blob structures, sheet structures, and tubular structures, as summarized in Table 1. We want a way to identify the latter from the previous ones, obviously because veins are mainly tubes. Conceptually, those structures differ in the way their normal vector varies. Mathematically, those différences can be found by sorting the Eigenvalues for each voxel ( |AX| < \À2\ < |A3| ): for example, if |A3| is high but the two others are near zéro, it means that the changes in intensity between the current voxel and its surrounding are in a single direction and that we face a sheet-like structure. To detect tubular structures, i.e. the venous blood vessels, we need to find voxels where A jw ill be near zéro while \À2\ and |/l3| will be similar and much greater than \Xx\, meaning that the current voxel is greatly différent from its surrounding in 2 out of the 3 available dimensions, hence suggesting the tube form amongst ail others (Table 1). It is noteworthy that in an effort to optimize speed and efficiency, the Eigenvalues are computed in a matrix fashion over the whole volume instead of the analyzing each voxel sequentially. We were able to implement a numeric formula for each Eigenvalue and compute them for every voxel at the same time using the symmetric nature of the Hessian matrix. 2.2.3 V e sse ln e ss m easure To suppress artifacts, blobs and sheet-like structures and to keep only venous blood vessels, (Frangi et al. 1998) proposed to compute 3 factors:

/?_

—

B

VIW sl'

n - IM Ra ~ ia3r S =

^A

2

+

À

2

+

A3,

where RB is only non-zero for undesired structures such as blobs and noisy artifacts. This situation happens when the smallest Eigenvalue A3 becomes too high, signifying that the

6

detected edge is not foliowing a spécifie direction. RA differentiates tube-like from sheet-like structures by ensuring that both À2 and À3 values are close. Finally, the Frobenius norm S suppresses the random noise effects in the computation o f the vesselness value. Finally the three factors are combined to create the multi-scale vesselness response:

The value a and p were found to be the best at 0.5 in (Frangi et al. 1998), while c was set as the half of the maximum Frobenius norm as suggested in (Descoteaux et al. 2008). Those constants were stable across a large number of experiments and should not be changed by the user. The vesselness measure is computed for ail voxels and a map o f the veins of size a of the brain is produced. The vesselness is computed for many scales o, which we now describe. 2.2.4 M u tli-sca le p ro c e s s in g The derivative of Gaussian filter G(o) is used. The parameter o is linked to the size of the veins. Since the sizes of the veins vary throughout in the brain, we need to repeat the computation of the vesselness measure for différent o value. This multi-scale processing enables the détection of a complété vasculature map, keeping only the highest vesselness measure given by every scale in each voxel and therefore the correct vein size. To résumé, there are 3 major steps to the algorithm: 1. Compute the second order derivatives o f every voxel to construct the Hessian Matrix. 2. Compute the Eigenvalues of the Hessian matrix for each voxel. 3. Compute the vesselness measure of every voxel using the Eigenvalues.

Those three steps are thereafter repeated for each a scale and the maximum vesselness response is kept to obtain a multi-scale measure (Lindeberg 1998)

V'final =

(y O ))-

Once the vesselness volume is computed, it is thresholded according to a user-specified value to extract veins.

7

Results The MATLAB toolbox we developed is available at " pages.usherbrooke.ca/vesselsegmentation". The graphie interface (Figure 1) gives an easy way to load SWI, adjust the parameters and obtain the veins segmentation, which can be saved for further processing.

3.1 Validation To quantify the précision of our method, we compared our results with the C++ implémentation of

(Descoteaux

et

al.

2008)

that

is

based

on

the

MINC

tools

(http://www.bic.mni.mcgill.ca/ServicesSoftware/MINC) using two quantitative measures. The first one is the kappa coefficient (Dice 1945), defined by 2a K

2a + b + c

and the second being

ratio =

a a + b

where a is the number of voxel belonging to both our results and the ground truth results, b the number of voxels belonging only to the ground truth volume and c the number of voxels given by our segmentation program alone. The second measure is The first measure tests the degree to which the correspondence between the two results exceeds chance level (a value of 1 is 100% sure that it is not due to chance). The second ratio indicates the degree to which the ground truth data is accounted for by the test data. The same datasets were processed w ith both version of the algorithm and the results for our 19 subjects are given in Table 2. We set the threshold at 0.015 and used an evenly spaced scale (0.5) ranging from 0.5 to 5.0. It is obvious that our implémentation conserves the accuracy of (Descoteaux et al. 2008) method with a ratio average of 0.99 and kappa average of 0.98.

3.2 Human vasculature (1.5 Tesla) Figure 2 shows the results of the toolbox on one brain. Major veins were identified according to an anatomically accurate atlas obtain via http://www.radnet.ucla.edu/sections/DINR/.

8

3.3 Rat vasculature (7 Tesla) Figure 3 shows the resuit given by the program for veins segmentation on a rat brain obtained using proton density (spin density) sequence. The volume dimensions were 512x512x256 and has 0.059mm isometric voxels. We used a threshold value o f 0.015 and a log sigma scale of 10 values between 0.04 and 5.0 voxels.

Discussion 4.1 The SNAIL Veins Segmentation Program The SNAIL Veins Segmentation Toolbox cornes with numerous parameters that can be adjusted by the user to extract the best possible vesselness volume from the initial SWI. This section discusses the effect of those parameters and their pertinence.

4.1.1 Modifying the number of scales The algorithm computes the vesselness measure based on the Hessian matrix o f each voxel with différent scale representing the various sizes of the vessels. Using différent scales is therefore primordial to detect ail the possible veins. Figure 4 shows the effect of choosing différent sigma scales. Giving only one sigma finds veins of corresponding size, but anything too small or too big will not be recognized (as seen in the left column). Finding a more accurate vesselness map will require more sigmas, thus covering a wider range of différent vessel sizes. More scales within a definite range will enhance the quality of the détection by retrieving a broader range o f veins of différent sizes. This is seen by the reduced artifacts présent in the 10 sigma image versus the 5 sigma image. The more scales used, the longer the exécution time. For example, a volume of size 320x260x80, K (c ) for one o takes between 15 and 20 seconds to computer on today's computer.

4.1.2 Modifying the maximum veins width The Maximum Veins Width (Figure lb ) represents the size of the biggest vein in the brain to be segmented. This value is considered as the highest sigma value when the sigma scale is automatically computed. Figure 5 présents the effect of varying the maximum vein size value for a volume. If the given width is too small, the algorithm finds a large number of small veins, and does not find larger veins within the volume. On the contrary, using a larger vein width w ithout increasing the

9

number of scales can dilute the sigma scales and reduce the précision of the results. Using the right maximum vein size will greatly improve the resuit of the segmentation (as shown in Figure 5, where the best veins size is 5).

4.1.3 Modifying the threshold value The threshold value can be changed to modify the segmented resuit. Based on its value, every voxel with a vesselness measure lower than the threshold will be put to 0. It can be used to eliminate voxels below a certain level of confidence of the vesselness measure and conserve those considered as real vein voxels. It can thus serve to reduce the number o f artifacts. However, a too high threshold will lead to the disappearance of true veins (Figure 6).

4.1.4 Other abilities The toolbox also cornes with the ability to segment bright blood volumes (angiographic-like volume). Finally, if the user knows the parameters he needs for ail his volumes, he can select the Multiple Volume Button. This feature enables the user to select multiple volumes and segment them one by one with the previously fixed parameters w ithout any other intervention.

4.1.5 Improvements An extra layer of processing could be added on top of the vesselness for auto-segmentation, as attempted in (Descoteaux et al. 2008), but this requires more computation tim e and more user interaction for seeding. At this stage, we prefer a threshold to give an efficient and interactive segmentation.

4.2 Implications The segmentation of veins can lead to the détection of vascular problems by comparing patients SWI segmentation with healthy subjects segmentation. It can also be used, as in (Vigneau-Roy et al. 2013), to compute vascular density. The tool presented here will increase the analysis possibilities for SWI and promote the use of that modality which is available on the vast majority of clinical imager.

Conclusion In this paper, we have introduced a user-friendly MATLAB SPM toolbox to segment brain vasculature. Results were shown from both human and rat using non-invasive SWI and proton

10

density

sequences.

We

give

this

open

source

MATLAB

SPM toolbox,

available

at

" pages, u s he r b r o o k e. c a / v e s s e l s e g m e n t a t i o n " .

Acknowledgements This research was supported by the Canada Research Chair program (CRC), the Natural Sciences and Engineering Research Council of Canada (NSERC) and the Ministère du Développement Économique, de l'Innovation et Exportation (MDEIE).

11

Bibliography Aylward SR, Bullitt E. Initialization, noise, singularities, and scale in height ridge traversai for tubular object centerline extraction. IEEE Transactions on Médical Imaging [Internet]. IEEE; 2002;21(2):61-75. Available from: http://www.ncbi.nlm.nih.gOv/pubmed/11929106 Bandettini PA, Wong EC. Effects of Biophysical and Physiologie Parameters on Brain Activation-lnduced R2* and R2 changes: Simulations Using a Deterministic Diffusion Model. International Journal of Imaging Systems and Technology [Internet], 1995;6(2/3): 133/152. Available from: http://content.ebscohost.com/pdf13_15/pdf/1995/BWI/01 Jun95/14184282. pdf?T= P&P=AN&K=14184282&S=R&D=a9h&EbscoContent=dGJyMMvl7ESeqK84yNfs OLCmr0uep7BSsqi4S7OWxWXS&ContentCustomer=dGJyMPGusUu0qLVKuePf geyx44Dt6fJJ Bandettini PA, Wong EC. A hypercapnia-based normalization method for improved spatial localization of human brain activation with fMRI. NMR in Biomedicine [Internet], John Wiley & Sons, Ltd.; 1997; 10(4-5): 197-203. Available from: http://www.ncbi.nlm.nih.gov/pubmed/9430348 Davis TL, Kwong KK, Weisskoff RM, Rosen BR. Calibrated functional MRI: Mapping the dynamics of oxidative metabolism. Proceedings of the National Academy of Sciences of the United States of America [Internet], The National Academy of Sciences; 1998;95(4): 1834-9. Available from: http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=19199&tool=pmcentrez &rendertype=abstract Descoteaux M, Collins DL, Siddiqi K. A géométrie flow for segmenting vasculature in proton-density weighted MRI. Médical Image Analysis [Internet]. 2008;12(4):497-513. Available from: http://www.ncbi.nlm.nih.gov/pubmed/18375175 Dice LR. Measures of the amount of écologie association between species. [Internet], JSTOR; 1945;26(3):297-302. Available from: http://www.jstor.org/stable/1932409

Ecology

Frangi AF, Niessen WJ, Vincken KL, Viergever MA. Multiscale vessel enhancement filtering 1 Introduction 2 Method. Wells WM, Colchester A, Delp SL, editors. Computer [Internet]. Springer; 1998; 1496(3): 130-7. Available from: http://www.ncbi.nlm.nih.gov/pubmed/15854841 Haacke EM, Xu Y, Cheng Y-CN, Reichenbach JR. Susceptibility weighted imaging (SWI). Zeitschrift fur medizinische Physik [Internet]. Wiley Online Library; 2006;52(3):612-8. Available from: http://www.ncbi.nlm.nih.gov/pubmed/15334582

12

Jochimsen TH, Ivanov D, Ott DVM, Heinke W, Turner R, Môller HE, et al. Whole-brain mapping of venous vessel size in humans using the hypercapnia-induced BOLD effect. Neurolmage [Internet]. Elsevier Inc.; 2010;51(2):765-74. Available from: http://www.ncbi.nlm.nih.gov/pubmed/20188189 Kao Y-H, Guo W-Y, Wu Y-T, Liu K-C, Chai W-Y, Lin C-Y, et al. Hemodynamic segmentation of MR brain perfusion images using independent component analysis, thresholding, and Bayesian estimation. Magnetic Résonance in Medicine [Internet]. 2003;49(5):885-94. Available from: http://www.ncbi.nlm.nih.gov/pubmed/12704771 Koopmans PJ, Manniesing R, Niessen WJ, Viergever MA, Barth M. MR venography of the human brain using susceptibility weighted imaging at very high field strength. Magma New York NY [Internet]. Springer; 2008;21 (1-2): 149-58. Available from: http://www.ncbi.nlm.nih.gov/pubmed/18188626 Krissian K. Model-Based Détection of Tubular Structures in 3D Images. Computer Vision and Image Understanding [Internet], Elsevier; 2000;80(2): 130-71. Available from: http://linkinghub.elsevier.com/retrieve/pii/S107731420090866X Liao W, Rohr K, Kang C-K, Cho Z-H, Worz S. A generative MRF approach for automatic 3D segmentation of cérébral vasculature from 7 Tesla MRA images. 2011 IEEE International Symposium on Biomédical Imaging From Nano to Macro IEEE; 2011 p. 2041-4. Lindeberg T. Edge détection and ridge détection with automatic scale sélection. Proceedings CVPR IEEE Computer Society Conférence on Computer Vision and Pattern Récognition [Internet], IEEE Comput. Soc. Press; 1998;30(2):465-70. Available from: http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=517113 Menon RS, Ogawa S, Tank DW, Ugurbil K. Tesla gradient recalled echo characteristics of photic stimulation-induced signal changes in the human primary visual cortex. Magnetic Résonance in Medicine [Internet]. 1993;30(3).380-6. Available from: http://www.ncbi.nlm.nih.gov/pubmed/8412612 Moon CH, Fukuda M, Kim S-G. Spatiotemporal characteristics and vacular sources of neural-specific and -nonspecific fMRI signais at submillimeter columnar resolution. Neurolmage. 2013;64(1):91-103. Ogawa S, Menon RS, Tank DW, Kim SG, Merkle H, Ellermann JM, et al. Functional brain mapping by blood oxygénation level-dependent contrast magnetic résonance imaging. Biophysical Journal. 1993;64(3):803-12. Passât N, Ronse C, Baruthio J, Armspach J-P, Maillot C. Magnetic résonance angiography: from anatomical knowledge modeling to vessel segmentation. Médical Image Analysis [Internet]. 2006;10(2):259-74. Available from: http://www.ncbi.nlm.nih.gov/pubmed/16386938

13

Singh M, Kim T, Kim H, Khosla D. Séparation of veins from activated brain tissue in functional magnetic résonance images at 1.5 T. IEEE Transactions on Nuclear Science 1995. Vigneau-Roy N, Bernier M, Descoteaux M, Whittingstall K. Régional variations in vascular density correlate with resting-state and task-evoked BOLD signal amplitude. Human Brain Mapping. 2013; Xu X, Dou F, Wang C, Shuo S, Chen J, Xu J, et al. Segmentation of Cérébral Venous Vessel in SWI Based on Multi-Adaptive Threshold with Vessel Enhancement and Background Effects Elimination. International Conférence on Internet Computing for Science and Engineering [Internet]. 2012. p. 107-11. Available from: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6239729 Yu X, Glen D, Wang S, Dodd S, Hirano Y, Saad Z, et al. Direct imaging of macrovascular and microvascular contributions to BOLD fMRI in layers IV-V of the rat whisker-barrel cortex. Neurolmage [Internet]. Elsevier Inc.; 2012;59(2): 1451-60. Available from: http://www.ncbi.nlm.nih.gov/pubmed/21851857 Zhao F, Wang P, Hendrich K, Ugurbil K, Kim S-G. Cortical layer-dependent BOLD and CBV responses measured by spin-echo and gradient-echo fMRI: insights into hemodynamic régulation. Neurolmage [Internet]. 2006;30(4):1149-60. Available from: http://www.ncbi.nlm.nih.gov/pubmed/16414284

14

Help

Load Volume

Segmentation Options C

: Bright Blood

Max V essel Width k (in Voxel) 5

r

Advanced

j Faster Segmentation (5 scale) ] Give your scale

C . 1 :C 2.C ? . C . £ . C . S : C . c : C . 7 : C . 8 : C 9. 1 C

Multiple Volumes

Close

Figure 1: The User Interface o f our toolbox. The red frame is the “Advanced Options” that appears once the ‘Advanced’ button is clicked. The ‘Compute V esselness’ button becomes available once the volume has been loaded, and the ‘Save V esselness’ is enabled once the vesselness volume is computed. One can also choose the ‘Multiple Volumes’ option to segment multiple volumes at the same time.

15

cn

4—*

13 (/) O m c

o