Date et lieu
10 novembre 2011 à 14h à Mines ParisTech
Amphithéâtre Georges Charpak
60, boulevard Saint Michel
75272 Paris

Jury
M. Michel JOURLIN, Professeur, Université St Etienne, rapporteur
M. Jean SEQUEIRA, Professeur, Université de la Méditérranée, Marseille, rapporteur
Mme Christine GRAFFIGNE, Professeur, Université Paris Descartes, examinateur
M. Philippe GRENIER, Professeur, UPMC - LIF (UMR_S 678), examinateur
M. Pierre-Yves BRILLET, Professeur, Université Paris-Nord, UPRES EA 2363, examinateur
M. Frédéric BANEGAS, Dr., Directeur Technique Intrasense, examinateur
Mme Françoise PRETEUX, Professeur, Mines ParisTech, Directeur de thèse, examinateur
M. Catalin FETITA, Maître de Conférences, TSP/ARTEMIS UMR 8145 MAP5, encadrant, examinateur

Résumé
The motivation of this work is the development of a computer-aided diagnosis (CAD) system for pulmonary diseases such as asthma and COPD. The analysis targets the airway system, where these diseases induce biological and morphological changes and lead to impaired respiratory function. The clinical interest lies in understanding the mechanisms and relationships between airway structure/physiology and the clinical phenotype and genotype. This knowledge could enable the prediction of disease progression, allow for the determination of the therapeutic response and improve the patients comfort and healing process.

Such a CAD system adopts the concept of "Image as Marker" and exploits routinely available MDCT image data, which enable a volumetric and non-invasive quantitative analysis. Our contribution addresses image-based dedicated investigation approaches, in order to identify relevant pathological markers. The latter ones are found by capturing morphological changes in the airway lumen, the airway wall and bronchial tree subdivision pattern, each raising the challenging problem of segmentation and quantification. Other issues concern the automatic location and severity determination of abnormalities, and a patient-specific modeling of the ventilation. In addition, validation of each proposed solution is necessary to determine its accuracy and robustness.

The key issue of our work is the volumetric segmentation of the airway wall represented by a dual-surface mesh delineation of internal and external airway wall borders. The developed original approach relies on a patient-specific, active surface, mesh model evolving over the image space according to Lagrangian dynamics, under the constraints of an external diffusion vector field. Based on the dual-surface representation of the airway wall border, the morphometric quantification of the airway system can be achieved. First, a local maximal airway lumen caliber is introduced and its computation performed on the internal airway wall border surface. Its visual feedback using color-coded information enables the semi-quantitative assessment of the airway shape variation. This knowledge is further exploited for the automatic detection of shape abnormalities, such as stenosis or bronchiectasis, and their location and severity. Regions of interest are thus identified on the central axis of the airway tree for an in-depth investigation based on a dense sample of cross-section measurements. Secondly, the airway wall remodeling subject to a pathology or an applied therapy, can be captured by analyzing the local wall thickness information available all over the bronchial tree, including subdivisions. Intuitive visualization, navigation and interaction capabilities associated with the extracted quantitative data complete the developed system for airway pathology assessment and follow-up.

All the solutions were quantitatively and qualitatively validated, either on synthetic or real MDCT images. Subsequently, computer phantoms based on patients’ geometries were generated and the CT image acquisition simulated, to obtain a ground truth model where different pathophysiological configurations could be recreated. The qualitative evaluation was performed on a large dataset of routine clinical data with feedback provided by radiologist experts.

The results reported a precision in the range of the CT image resolution and the robustness towards different acquisition protocols, pathologies and inter-patient variability. The successful application of the proposed methods on several clinical studies involving 150 patient cases assures the relevance of the extracted markers and the feasibility of high-throughput analyses in clinical research.