Get Complete Project Material File(s) Now! »
Segmentation Process
The model-based graph cuts segmentation process consists of two steps. The first step is to detect the boundary of the endocardium. In this step, quality of the spatial prior does not affect the result too much because of the thick LV wall. In order to cover a majority of the blood pool, a spatial prior constructed from o Cepi (Figure 3.3b) is used in this step instead of o Cendo . The area inside of o Cepi was filled and then smoothed with a 7×7 Gaussian smoothing kernel to form the spatial prior. The size of the kernel is based on the size of o Cepi . The result with and without using the spatial prior are also shown as two yellow curves in Figure 3.3c and Figure 3.3d, while the red dots give the ground truth.
3 Parameter Optimisation
Two weights are required to be optimized: λ which balances the boundary and regional terms and β which is the weight of the spatial probability in the regional term. Nine sets of parameters with different values of λ from 0.5 to 2 and β from 0.3 to 0.7 were selected for parameter optimisation and an average error was obtained for each patient with each parameter set. Box and whisker plots (or called box plots) of the data are shown in Figure 3.6. The best pair was number 3 which is λ = 2 and β = 0.7.
Achievements
The algorithm is a demonstration of the integration of high-level model-based knowledge into the graph cuts algorithm. It inherits the ability of the graph cuts to provide a global optimum for the entire image, while overcoming the limitation of using only low-level data. The spatial prior derived from the model contains not only spatial information but also some anatomical knowledge. This feature is particularly important when the low-level information is unable to determine the correct segmentation: for example in the cut between the RV myocardium and the LV myocardium. A modified Hausdorff distance measure showed good agreement between the model-based graph cuts and the expert observers in a midventricular slice.
Further Work
An iterative process can be performed if more accurate results are required. The results obtained from the graph cuts algorithm can be considered as the initial data for the model fitting. The parameters of the model, such as shapes, sizes, positions and orientations will be updated by tailoring the model to fit the data. The updated model can then provide more accurate initial contours for further image analysis. The analysis could also be considered in the context of propagation from image to image. The shape of the contours does not change too much from one image to the next, either in the spatial domain or in the temporal domain. The contours propagated from neighbouring images may greatly help the process in practice.
Limitations
Further development and improvement is required to solve failures in some cases. It has been found that because only the septum was used to extract the thickness of the myocardium, the algorithm often failed when the septal myocardium was much thinner than the free wall. Another problem comes from the attached papillary muscles, which affects the dilation step to obtain a good spatial prior for the epicardium. Furthermore, the current model only includes the LV so it is unable to segment the RV. Even if LV RV model is used, the method may still have difficulty in the RV segmentation. The thin RV free wall will cause leaking problems if parts of the boundaries are blurred. The spatial prior can provide proper constraints for the algorithm only when correct location information is available, which is difficult in the RV free wall area.
Table of Contents :
- Abstract
- Acknowledgements
- List of Figures
- List of Tables
- Glossary of Abbreviations
- 1 Introduction
- 1.1 Motivation
- 1.2 Objectives
- 1.3 Cardiac Anatomy and Function Parameters
- 1.4 Data Acquisition
- 1.5 Construction of Heart Model
- 1.6 CMR Image Segmentation
- 1.7 Overview of the Thesis
- 1.8 Achievements in the Thesis
- 2 Automated Detection of the Left Ventricle on 4D CMR Images
- 2.1 Introduction
- 2.2 Method
- 2.3 Results
- 2.4 Discussions
- 2.5 Conclusions
- 3 Model-based Graph Cuts Method for Automated Segmentation
- 3.1 Introduction
- 3.2 Method
- 3.3 Implementation
- 3.4 Experiment
- 3.5 Discussions and Conclusions
- 4 Evaluation of Similarity Measures for Atlas-based Rigid-body Registration
- 4.1 Similarity Measures
- 4.2 SMPL Atlas-based Registration Framework
- 4.3 Protocol for Evaluation
- 4.4 Implementation
- 4.5 Results
- 4.6 Discussion
- 4.7 Conclusions
- 5 Atlas-based Segmentation of Cardiac MR Images
- 5.1 Introduction
- 5.2 Atlas Construction
- 5.3 Free-form Deformation for Non-rigid Registration
- 5.4 Energy Function
- 5.5 Optimisation
- 5.6 Capture Range Analysis
- 5.7 Large Dataset Analysis
- 6 Modified Atlas-based Segmentation for Cardiac MR Images
- 6.1 Inclusion of High Level Information
- 6.2 Integration of the Boundary Term
- 6.3 Integration of Area Term
- 6.4 Results
- 6.5 Discussion
- 6.6 Conclusions
- 7 Model-based 3D Segmentation
- 7.1 Introduction
- 7.2 Feature-based 3D Registration Method
- 7.3 Segmentation on Other Short-axis Slices
- 7.4 Segmentation of the 4-chamber Long-axis slice
- 7.5 Results
- 7.6 Discussions
- 7.7 Conclusions
- 8 Conclusions
- 8.1 Thesis Summary
- 8.2 Further Work
- 9 Publications
- 10 References
GET THE COMPLETE PROJECT
Model-based Strategies for Automated Segmentation of Cardiac Magnetic Resonance Images