Date on Master's Thesis/Doctoral Dissertation

5-2012

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Electrical and Computer Engineering

Committee Chair

Graham, James H.

Author's Keywords

Object modeling; Object categorization; Object segmentation; Lung nodules; Object detection

Subject

Diagnostic Imaging; Lungs--Imaging

Abstract

Active Shape Models (ASM), Active Appearance Models (AAM) and Active Tensor Models (ATM) are common approaches to model elastic (deformable) objects. These models require an ensemble of shapes and textures, annotated by human experts, in order identify the model order and parameters. A candidate object may be represented by a weighted sum of basis generated by an optimization process. These methods have been very effective for modeling deformable objects in biomedical imaging, biometrics, computer vision and graphics. They have been tried mainly on objects with known features that are amenable to manual (expert) annotation. They have not been examined on objects with severe ambiguities to be uniquely characterized by experts. This dissertation presents a unified approach for modeling, detecting, segmenting and categorizing small objects under uncertainty, with focus on lung nodules that may appear in low dose CT (LDCT) scans of the human chest. The AAM, ASM and the ATM approaches are used for the first time on this application. A new formulation to object detection by template matching, as an energy optimization, is introduced. Nine similarity measures of matching have been quantitatively evaluated for detecting nodules less than 1 em in diameter. Statistical methods that combine intensity, shape and spatial interaction are examined for segmentation of small size objects. Extensions of the intensity model using the linear combination of Gaussians (LCG) approach are introduced, in order to estimate the number of modes in the LCG equation. The classical maximum a posteriori (MAP) segmentation approach has been adapted to handle segmentation of small size lung nodules that are randomly located in the lung tissue. A novel empirical approach has been devised to simultaneously detect and segment the lung nodules in LDCT scans. The level sets methods approach was also applied for lung nodule segmentation. A new formulation for the energy function controlling the level set propagation has been introduced taking into account the specific properties of the nodules. Finally, a novel approach for classification of the segmented nodules into categories has been introduced. Geometric object descriptors such as the SIFT, AS 1FT, SURF and LBP have been used for feature extraction and matching of small size lung nodules; the LBP has been found to be the most robust. Categorization implies classification of detected and segmented objects into classes or types. The object descriptors have been deployed in the detection step for false positive reduction, and in the categorization stage to assign a class and type for the nodules. The AAMI ASMI A TM models have been used for the categorization stage. The front-end processes of lung nodule modeling, detection, segmentation and classification/categorization are model-based and data-driven. This dissertation is the first attempt in the literature at creating an entirely model-based approach for lung nodule analysis.

COinS