François Chung, Ph.D.

Tag: modélisation de l’apparence

CVIU 2013 - Article de revue scientifique

CVIU 2013 – Article de revue scientifique

Publication

François Chung, Hervé Delingette; Regional appearance modeling based on the clustering of intensity profiles; In: Computer Vision and Image Understanding (CVIU), 117 (6), pp. 705-717, 2013.

Abstract

Model-based image segmentation is a popular approach for the segmentation of anatomical structures from medical images because it includes prior knowledge about the shape and appearance of structures of interest. This paper focuses on the formulation of a novel appearance prior that can cope with large variability between subjects, for instance due to the presence of pathologies. Instead of relying on Principal Component Analysis (PCA) such as in Statistical Appearance Models (SAMs), our approach relies on a multimodal intensity profile atlas from which a point may be assigned to several profile modes consisting of a mean profile and its covariance matrix. These profile modes are first estimated without any intra-subject registration through a boosted Expectation-Maximization (EM) classification based on spectral clustering. Then, they are projected on a reference mesh whose role is to store the appearance information in a common geometric representation. We show that this prior leads to better performance than the classical monomodal PCA approach while relying on fewer profile modes.

Mots-clés

  • appearance modeling
  • medical imaging
  • model-based image segmentation
  • unsupervised clustering

Références

Publication

Articles associés

3D Anatomical Human (projet INRIA)
Ph.D. Thesis 2011 (thèse de doctorat)

LAP 2011 - Livre

LAP 2011 – Livre

Publication

François Chung; Regional appearance modeling for model-based image segmentation: Methodological approaches to improve the accuracy of model-based image segmentation; Lambert Academic Publishing (LAP), Saarbrücken, 2011; ISBN: 978-3844322095.

Abstract

This thesis presents a novel appearance prior for model-based image segmentation. This appearance prior, denoted as Multimodal Prior Appearance Model (MPAM), is built upon an Expectation–Maximization (EM) clustering of intensity profiles with model order selection to automatically select the number of profile classes. Unlike classical approaches based on Principal Component Analysis (PCA), the clustering is considered as regional because intensity profiles are classified for each mesh and not for each vertex. Comparative results on liver profiles from Computed Tomography (CT) images show that MPAM outperforms PCA-based appearance models. Finally, methods for the analysis of lower limb structures from Magnetic Resonance (MR) images are presented. A first part deals with the creation of subject-specific models for kinematic simulations of the lower limbs. In a second part, the performance of statistical models is compared in the context of lower limb bone segmentation when only a small number of datasets is available for training.

Références

Publication

Livre (Amazon)
Livre (MoreBooks)
Référence bibliographique (BibTeX)

Articles associés

3D Anatomical Human (projet INRIA)
Ph.D. Thesis 2011 (thèse de doctorat)

En savoir plus

LAP – Lambert Academic Publishing

Ph.D. Thesis 2011 - Thèse de doctorat

Ph.D. Thesis 2011 – Thèse de doctorat

Publication

François Chung; Regional appearance modeling for deformable model-based image segmentation; Thèse de doctorat (Ph.D. Thesis), Mines ParisTech, Centre de Mathématiques Appliquées, 2011.

Abstract

This thesis presents a novel appearance prior for model-based image segmentation. This appearance prior, denoted as Multimodal Prior Appearance Model (MPAM), is built upon an Expectation-Maximization (EM) clustering of intensity profiles with model order selection to automatically select the number of profile classes. Unlike classical approaches based on Principal Component Analysis (PCA), the clustering is considered as regional because intensity profiles are classified for each mesh and not for each vertex.

First, we explain how to build a MPAM from a training set of meshes and images. The clustering of intensity profiles and the determination of the number of appearance regions by a novel model order selection criterion are explained. A spatial regularization approach to spatially smooth the clustering of profiles is presented and the projection of the appearance information from each dataset on a reference mesh is described.

Second, we present a boosted clustering based on spectral clustering, which optimizes the clustering of profiles for segmentation purposes. The representation of the similarity between data points in the spectral space is explained. Comparative results on liver profiles from Computed Tomography (CT) images show that our approach outperforms PCA-based appearance models.

Finally, we present methods for the analysis of lower limb structures from Magnetic Resonance (MR) images. In a first part, our technique to create subject-specific models for kinematic simulations of lower limbs is described. In a second part, the performance of statistical models is compared in the context of lower limb bones segmentation when only a small number of datasets is available for training.

Mots-clés

  • appearance modeling
  • liver
  • lower limbs
  • medical imaging
  • model-based image segmentation
  • unsupervised clustering

Références

Publication

Articles associés

3D Anatomical Human (projet INRIA)
CVIU 2013 (article de revue scientifique)
LAP 2011 (livre)

En savoir plus

MICCAI 2009 - Acte de conférence

MICCAI 2009 – Acte de conférence

Publication

François Chung, Hervé Delingette; Multimodal prior appearance models based on regional clustering of intensity profiles; MICCAI 2009: Medical Image Computing and Computer-Assisted Intervention, London, 2009.

Abstract

Model-based image segmentation requires prior information about the appearance of a structure in the image. Instead of relying on Principal Component Analysis (PCA) such as in Statistical Appearance Models (SAMs), we propose a method based on a regional clustering of intensity profiles that does not rely on an accurate pointwise registration. Our method is built upon the Expectation-Maximization (EM) algorithm with regularized covariance matrices and includes spatial regularization. The number of appearance regions is determined by a novel model order selection criterion. The prior is described on a reference mesh where each vertex has a probability to belong to several intensity profile classes.

Références

Publication

Articles associés

Modèle multimodal (projet INRIA)
3D Anatomical Human (projet INRIA)

En savoir plus

MICCAI – Medical Image Computing and Computer Assisted Intervention

Modèle multimodal

Modèle multimodal

Projet INRIA @Sophia-Antipolis, France (2009). La segmentation d'images basée modèle requiert une information préalable sur l'apparence d'une structure dans l'image. Au lieu de s'appuyer sur l'analyse en composantes principales (ACP), comme c'est le cas avec les modèles d'apparence statistiques (SAM), nous proposons un nouveau modèle d'apparence basé sur un clustering régional de profils d'intensité qui ne repose pas sur un recalage point à point précis.

Ce modèle multimodal d'apparence, dénommé Multimodal Prior Appearance Model (MPAM), repose sur l'algorithme d'espérance-maximisation (EM) combiné avec des matrices de covariance régularisées et inclut une régularisation spatiale. Le nombre de régions d'apparence est déterminé par un nouveau critère de sélection. Le modèle est décrit sur un maillage de référence où chaque sommet a une probabilité d'appartenir à plusieurs classes de profils d'intensité.

Nous avons testé notre méthode avec 7 maillages du foie segmenté à partir d'images tomographiques (CT) et 4 maillages du tibia segmenté à partir d'images par résonance magnétique (IRM). Pour les deux structures, des profils d'intensité composés de 10 échantillons extraits chaque mm ont été générés à partir de maillages composés d'environ 4000 points.

Le principal avantage de notre approche est que les régions d'apparence sont extraites sans nécessiter un recalage point à point précis. Un autre avantage est qu'un modèle peut être construit avec un nombre restreint d'ensembles de données (en fait, un seul ensemble suffit). En outre, le modèle est multimodal, et donc en mesure de faire face à une grande variation au niveau de l'apparence.

Références

Articles associés

MICCAI 2009 (acte de conférence)
Ph.D. Thesis 2011 (thèse de doctorat)