No d'ordre: THESE

Louis Pasteur - Strasbourg

Description

Niveau: Supérieur, Doctorat, Bac+8
No d'ordre: 4616 THESE presentee pour obtenir le grade de Docteur de l'Universite Louis Pasteur - Strasbourg I Ecole doctorale : Sciences pour l'ingenieur Discipline : Electronique, electrotechnique, automatique Specialite : Traitement d'images et vision par ordinateur Modeles statistiques d'apparence non gaussiens. Application a la creation d'un atlas probabiliste de perfusion cerebrale en imagerie medicale English title: “Non-Gaussian Statistical Appearance Models. Application to the Creation of a Probabilistic Atlas of Brain Perfusion in Medical Imaging.” Soutenue publiquement le 21 septembre 2004 par Torbjørn VIK Membres du jury: Mme. Isabelle BLOCH Rapporteur externe M. Jack-Gerard POSTAIRE Rapporteur externe M. Ernest HIRSCH Rapporteur interne M. Philippe RYVLIN Examinateur M. Fabrice HEITZ Directeur de these M. Jean-Paul ARMSPACH Directeur de these

• images coming


• evaluation scheme


• both laboratories


• scientific seminars


• has further been


• normal subjects


• perfusion


• distribution-based intensity


• called appearance-based

Sujets

• Faison
• Kohler
• Hirsch

Informations

• Publié par: profil-zyak-2012
• Publié le: 01 septembre 2004
• Nombre de lectures: 26
• Langue: English
• Poids de l'ouvrage: 3 Mo

Extrait

oN d’ordre: 4616
THESE
presentee pour obtenir le grade de
Docteur de l’Universite Louis Pasteur - Strasbourg I
Ecole doctorale : Sciences pour l’ingenieur
Discipline : Electronique, electrotechnique, automatique
Specialite : Traitement d’images et vision par ordinateur
Modeles statistiques d’apparence non gaussiens.
Application a la creation d’un atlas probabiliste de
perfusion cerebrale en imagerie medicale
English title: \Non-Gaussian Statistical Appearance Models.
Application to the Creation of a Probabilistic Atlas of Brain Perfusion in Medical Imaging."
Soutenue publiquement
le 21 septembre 2004
par
Torbj rn VIK
Membres du jury:
Mme. Isabelle BLOCH Rapporteur externe
M. Jack-Gerard POSTAIRE Rapp
M. Ernest HIRSCH Rapporteur interne
M. Philippe RYVLIN Examinateur
M. Fabrice HEITZ Directeur de these
M. Jean-Paul ARMSPACH de theseThanks
Working and writing this thesis during four years has been an incredible experience to me. It
has given me the opportunity to learn a great many things, as well on the professional level as
on the personal level. It has further given me the opportunity to work with many competent
and resourceful persons. As with most research I believe, the achievement of this thesis has
followed a route of ups and downs on which I have not been alone. The interaction with and
the help from my collegues, friends and family has been indispensable and I would hereby like
to express my gratefulness to everybody.
First of all, I would like to thank Isabelle Bloch, Jack-Gerard Postaire and Ernest Hirsch
for their careful evaluation and judicious comments concerning the manuscript of this work.
I would also like to thank Ernest Hirsch for his guidance as a pedagogical mentor and for his
engagement in organizing bilingual scienti c seminars which has permitted me to enlarge my
professional vision. I would further like to thank Philippe Ryvlin for the interest he has shown
in my work, an extremely di cult task for a non-specialist.
I am very grateful for having had the opportunity to work under the guidance of Fabrice
Heitz and Jean-Paul Armspach who, through engagement and dedication, have created a rich
and dynamic working environment for me and others. Furthermore, I would like to thank
both for their support and the many discussions we have had during these years.
I thank Daniel Grucker and Jean-Fran cois Dufourd for having received me at their respec-
tive laboratories. A special thank to the personnel and co-workers at both laboratories.
I would also like to thank the persons I have met during my teaching duties, Sophie Kohler,
Yoshi Takakura, Jean Martz, Laurent Thoraval and others.
I would like to thank my fellow doctoral students, Marcel Bosc, Sylvain Faisan, Vincent
Noblet, Rozenn Dahyot, Farid Flitti, Aicha Far and others. In particular, I would like to thank
Marcel Bosc, who taught me linux/unix and who took me to a new level of C++ programming.
His concern and engagement for out-of-work issues has further been an immense personal
enrichement to me. I would also like to thank Sylvain Faison who I could easily win for
scienti c discussions (mostly futile) through which I again caught pleasure in my work after a
rude rst year. I hope we shall have time again for our scienti c and non-scienti c bike-rides
when stu has calmed down after the thesis. During these years, I have also enjoyed the
companionship of a long list of internships and temporary workers, most lately Thomas Berst,
Nicolas Wiest-Daessle and Samuel Sinapin.
A warm thanks goes to my wife Aude and to my son Emil. You make me a very happy
person. I would also like to thank our families and friends for their support and for many
pleasant occasions during these years and in the years to come.ivAbstract
Single Photon Emission Computer Tomography (SPECT) is a 3D functional imaging tech-
nique that yields information about the blood o w in the brain (also called brain perfusion).
This imaging technique has found application in the diagnostics of head trauma, dementia,
epilepsy and other brain pathologies. To this end, SPECT images are analyzed in order to nd
abnormal blood o w patterns. For localized abnormalities such as stroke, this characterization
remains an accessible task, whereas for di use and variable abnormalities such as beginning
dementia, near-drowning episodes and toxic substance exposure, characterization is di cult.
It is therefore necessary to develop quantitative methods in which computer-aided statistical
analysis can take advantage of information present in a database of normal subjects.
This work deals with the construction and evaluation of a probabilistic atlas of brain
perfusion in normal subjects as observed in SPECT images. The goals of such an atlas are
twofold: (1) to describe perfusion patterns of the population represented by the atlas in a
compact manner, and (2) to identify statistically signi can t di erences between an individual
brain perfusion pattern and the probabilistic atlas. The successful creation of a computerized,
probabilistic atlas may have far-reaching impact on clinical applications where qualitative
(visual) analysis of SPECT images is current practice.
Three issues have been central in this work: the statistical models that actually describe
brain perfusion, the image processing tools used to make brains \comparable" and the experi-
mental evaluation of the atlas. For the rst issue, we have explored so-called appearance-based
approaches. These have been developed in computer vision where they have also been widely
adopted. Recent developments have given these models a proper statistical basis. In this
work, we have introduced an original non-linear model based on principal component analysis
(PCA) and Bayesian estimation theory.
The second issue is related to the spatial normalization of images (i.e. image registration)
and intensity normalization. In order to compare brain images coming from di eren t subjects,
their relative positions must be found. This is done by calculating a non-linear mapping be-
tween corresponding anatomical regions. A registration scheme speci cally adapted to SPECT
images had to be developed for this task. Furthermore, since the gray values in SPECT im-
ages only represent relative measures of blood o w, the observed values must be normalized to
allow for comparison between images. For this we devised an e cien t, joint distribution-based
intensity normalization scheme.
Finally, because of the lack of absolute knowledge about the brain perfusion in a normal
population, an elaborate evaluation scheme had to be developed. The scheme is based on the
detection of simulated abnormalities combined with a leave-one-out strategy. This scheme was
used to evaluate and compare the di eren t models and normalization schemes considered in
this work. For evaluation on a clinical application, the atlas was also applied to characterize
seizure foci in patients with epilepsy.Resume
Cette these a ete redigee en anglais. Voici un resume detaille de la these en fran cais.
La tomoscintigraphie par emission mono-photonique (TEMP) est une methode d’imagerie
fonctionnelle 3D qui apporte des informations sur le debit sanguin cerebral (egalement appele
perfusion cerebrale). Cette methode d’imagerie, par la detection visuelle d’anomalies de per-
fusion caracterisees par des zones hypo- ou hyper-intenses, est utilisee pour le diagnostic chez
des patients atteints d’accidents vasculaires cerebraux, de demence, d’epilepsie ou d’autres pa-
thologies cerebrales. La detection d’anomalies focalisees observees chez les patients ayant une
attaque cerebrale est relativement aisee, alors que les anomalies di uses, observees en debut
de demence, lors d’un accident entra^ nan t une oxygenation insu san te du cerveau ou suite a
une exposition a une substance toxique, sont plus di cilemen t observables. Dans ces cas, une
analyse quantitative des images, utilisant un atlas et des outils statistiques s’appuyant sur une
base d’images de cas normaux, peut apporter une aide precieuse au diagnostic.
Le travail presente dans cette these est centre sur la problematique de la construction et
de l’evaluation d’un atlas probabiliste de perfusion cerebrale a partir des images TEMP de
sujets dits normaux. Les objectifs d’un tel atlas sont doubles : (1) creation d’une cartographie
statistique de la perfusion cerebrale d’une population normale, decrite de maniere compacte, et
(2) identi cation des di erences de perfusion cerebrale qui sont statistiquement signi cativ es
entre une image TEMP d’un individu et l’atlas probabiliste. L’utilisation d’un atlas devrait
avoir un impact important sur les applications cliniques ou l’analyse qualitative d’images
TEMP est pratique courante.
A n d’atteindre ces objectifs, trois points ont ete abordes : le developpement de modeles
statistiques qui decrivent de fa con d ele la perfusion cerebrale, les outils de traitement d’images
utilises pour rendre les cerveaux < comparables >, et en n, l’evaluation experimentale de
l’atlas.
Pour le premier point, nous avons explore les approches dites < par modeles d’apparence >.
Ceux-ci ont ete developpes dans le domaine de la vision par ordinateur ou ils ont ete largement
appliques. Des developpements recents ont rede ni ces modeles dans un cadre statistique.
Dans ce travail, nous avons introduit un modele original non li