Curvilinear Structure Enhancement and Centreline Extraction in 2D and 3D Images

Abstract

Extracting meaningful information from digital images, and in particular from curvilinear structures, is a critical challenge in image processing and computer vision applications. Commonly, these curvilinear structures appear in several complexes environments, often with noisy and
inhomogeneous backgrounds. Furthermore, curvilinear structures featured in the same image can have different widths and can be spread across different directions. Thus, the automatic
extraction of the curvilinear structures without tunable parameters becomes increasingly difficult.
The extraction of curvilinear structures is a challenging and difficult task and many research studies are carried out in this field to develop methods to extract these structures. These methods can be broadly divided into two categories: (1) tools to improve the quality of appearance of curvilinear structures in an image and (2) tools to extract relevant features from the curvilinear structures, such as ridges or centrelines.
In this thesis, I present significant contributions to the image processing field, in particular, in
the curvilinear structures enhancement and centreline extraction. These contributions can be
grouped into:
(1) A method was developed to enhance the curvilinear structure in 2D and 3D grayscale images. This method combines the power of complex morphological operations and a local tensor representation to improve the quality of the enhancement results.
(2) After a thorough investigation and examination of the power of the tensor in curvilinear
structure enhancement, a new novel approach for curvilinear structure enhancement is introduced in collaboration with colleagues. This method uses spherical diffusion tensors
in 2D and 3D images.
(3) After reviewing the existing methods that proposed in the literature to extract the centreline, it was found that the majority of these methods relies on a large number of requirements(i.e., tuning parameters), which leads to robustness issues. Based on these findings, a novel
sequential graph-based approach is proposed in order to extract the centreline at different scales and orientations in an iterative stochastic process.
(4) In order to understand the relationship between enhancement and ridge detection in the extraction process of curvilinear structures, a set of experiments were proposed and statistical evaluation for each pair-wise combination was provided with the help and participation of colleagues.