Features-based MRI brain classification with domain knowledge : application to Alzheimer's disease diagnosis
|Michel Couprie||Professor, ESIEE, Université Paris-Est|
|Achille Braquelaire||Professor, LaBRI, Université de Bordeaux|
|Franka Brüchert||Professor, FVA, Baden-Württemberg, Freiburg|
|Fabien Feschet||Professor, ISIT, Université d'Auvergne|
|Fleur Longuetaud||Senior research, INRA, Champenoux|
|Frédéric Mothe||Senior research, INRA, Champenoux|
|Isabelle Debled-Rennesson||Professor, LORIA, Université de Lorraine|
|Bertrand Kerautret||Associate professor, LORIA, Université de Lorraine|
The non-destructive study of wood from X-Ray CT scanners requires to imagine new solutions adapted to analysis of images. Relating both agronomic research and industrial sector of sawmills, segmentation of wood knots is a major challenge in terms of robustness to specificities of each species and to image acquisition conditions. The works carried out in this thesis allow to propose a segmentation process in two phases. It first isolates each knot in a reduced area then it segments the unique knot of each area. Proposed solutions for each phase allow to integrate knowledges about internal organization of trunk and mechanisms inherent to its growth, through classical tools of image analysis and processing. The first phase is essentially based on a movement detection principle borrowed from video analysis and revisited. Two segmentation approaches are then proposed, considering for one the initial CT slices and for the other news slices resampled for each knot orthogonally to its trajectory. The complete process has been implemented in a software dedicated both for experimentation and validation of approach, and to interdisciplinary dialogs.
The applicative support of wood emphasizes the specialization abilities of generic image analysis and processing algorithms, and the relevance to integrate priori knowledges in this perspective.
Keywords: segmentation, CT images, geometric analysis, discrete geometry, wood knots.