Right Ventricle Segmentation Using Active Shape Model with Inter-Profile Modeling / Mohamed Samir Mohamed Elbaz

Supervisor: Ahmed Fahmy

Thesis (M.A.)—Nile University, Egypt, 2011 .

"Includes bibliographical references"

Contents:
CHAPTER1: INTRODUCTION ..................................................................................... 1
1.1. Heart Anatomy and Physiology 1
1.2. Cardiac Imaging 4
1.2.1. Echocardiography 4
1.2.2. Electron-beam computed tomography (EBCT) 5
1.2.3. Positron emission tomography (PET) 6
1.2.4. Cardiovascular magnetic resonance imaging (CMR) 7
1.2.4.1. Special Cardiac Techniques 8
1.3. Motivation 10
1.4. Thesis Objective 12
CHAPTER 2 : BACKGROUND ................................................................................... 14
2.1. Image Segmentation 14
2.1.1. Definitions 14
2.1.2. Pixel-based Segmentation Algorithms 15
2.1.2.1. Thresholding 15
2.1.2.2. Finite Mixture Models (FMM) 16
2.1.3. 18
2.1.4. Edge-based (boundary -based) Segmentation Algorithms 18
2.1.4.1. Edge-detection and Linking 18
2.1.4.2. Active Contour Model (ACM) 18
2.1.5. In this section the ACM is presented as summarized by T Lee [103]. 18
2.1.6. Segmentation Algorithms Using Prior Information 23
2.1.6.1. Distribution prior-based Segmentation 23
2.1.6.2. Shape prior-based Segmentation 25
2.1.7. Right Ventricle Segmentation 26
2.2. Dimensionality Reduction 28
2.2.1. Principal Component Analysis (PCA) 29
2.2.2. Two Dimensional PCA 31
2.2.2.1. Horizontal 2DPCA 32
2.2.2.2. Vertical 2DPCA 33
2.3. Active Shape Model (ASM) 33
2.3.1. Shapes and Landmarks
2.3.2. Shape Alignment and Mean of the Training Set 35
2.3.3. Shape Modeling 36
2.3.4. Profile Modeling 38
2.3.5. Active Shape Model (ASM) for Segmentation 41
2.3.5.1. Defining Initial Shape 41
2.3.5.2. Searching for the best profile 42
2.3.5.3. Multi-Resolution Approach for Active Shape Models 42
CHAPTER 3: DATASET AND VALIDATION METHOD .......................................... 44
3.1. Dataset 44
3.2. Validation Method 44
CHAPTER 4: RIGHT VENTRICLE SEGMENTATION USING SHAPE-SPECIFIC STATISTICAL MODEL .............................................................................................. 46
4.1. Building RV- Shape Model 46
4.1.1. Defining landmarks 46
4.1.2. Shape Alignment and RV Mean Shape 46
4.1.3. Modeling RV-Shape Variation 47
4.1.4. Building Profile Model 48
4.1.5. Multi-resolution Framework 49
4.2. Results 49
4.3. Discussion 51
4.4. Conclusion 52
CHAPTER 5 : ACTIVE SHAPE MODEL WITH INTER-PROFILE MODELING FOR RIGHT VENTRICLE SEGMENTATION............................................................ 53
5.1. Inter-profile Modeling Paradigm 53
5.2. Capturing the inter-profile variations 55
5.3. Iterative Multi-Stage Bidirectional 2DPCA-Based Best Profile Search Algorithm 58
5.4. Multi-resolution Framework 61
5.5. Methods 62
5.6. Results 63
5.7. Discussion 65
5.7.1. Computational Considerations 68
5.8. Conclusion 69
CHAPTER 6 : CONCLUSION ..................................................................................... 72
REFERENCES .....

Abstract:
Right ventricle (RV) role in cardiac evaluation has been underestimated for long years as opposed to Left Ventricle (LV). Recently, many studies have demonstrated the prognostic value of the RV function in cardiovascular diseases such as heart failure, RV myocardial infarction, congenital heart disease and pulmonary hypertension. Short-axis Magnetic Resonance Imaging (MRI) can provide clear images of the RV that allows quantitative assessment of the RV morphology, volume and function. Nevertheless, this requires manual delineation of the RV borders throughout the cardiac cycle, which is a tedious and time consuming process. Most of the automatic segmentation techniques have been devoted to LV or LV in conjunction with RV, but not RV in particular, resulting in lower performance of RV segmentation results compared to LV. However, the RV border segmentation is a challenging task because of the coarse trabeculations that cross the RV cavity which are difficult to outline because of their complex shape. To handle these RV complex shape challenges, we propose two methods for RV segmentation. The first method is based on capturing the RV-specific shape complexities and variations using the well-known classical Active Shape Model (ASM) method. The contribution here is to allow for segmenting RV depending only on its specific shape rather than combined LV and RV shape properties as currently proposed in literature [55, 60, 61]. In the second method, we propose a novel paradigm for modeling the appearance within the ASM framework. In this paradigm, we model the inter-profile relations between the different landmarks in the shape model rather than modeling each landmark‟s profile independently as commonly done in the classical ASM. This is achieved by constructing a profile matrix for each shape model which contains all the shape profiles. Next, the inter-profile relations are modeled using the Two-Dimensional PCA (2DPCA) in which the two versions of the 2DPCA (horizontal and vertical) are applied on the profiles matrix to capture the inter-profile modes of variations. This enables capturing and maintaining the RV shape complexities in a more natural and effective manner. Then, to benefit from the inter-profile modeling paradigm we propose to use an iterative multi-stage approach to search for the optimal landmark displacement. In the first stage, the best match is found for each profile in the profiles matrix given that all other rows in the matrix are the mean profiles resulted from the 2DPCA modeling step (horizontal or vertical). In the second stage, we find the best profiles matrix that match the model. The second stage is repeated iteratively until convergence is reached. Our experiments show that the proposed methods outperform the classical ASM-based segmentation method and give better results even with bad model initialization.

Text in English, abstracts in English.