Image segmentation is an important step in the image analysis process. Current image segmentation techniques, however, require that the user tune several parameters in order to obtain maximum segmentation accuracy, a computationally inefficient approach, especially when a large number of images must be processed sequentially in real time. Another major challenge, particularly with medical image analysis, is the discrepancy between objective measures for assessing and guiding the segmentation process, on the one hand, and the subjective perception of the end users (e.g., clinicians), on the other. Hence, the setting and adjustment of parameters for medical image segmentation should be performed in a manner that incorporates user feedback. Despite the substantial number of techniques proposed in recent years, accurate segmentation of digital images remains a challenging task for automated computer algorithms. Approaches based on machine learning hold particular promise in this regard because, in many applications, including medical image analysis, frequent user intervention can be assumed as a means of correcting the results, thereby generating valuable feedback for algorithmic learning. This thesis presents an investigation of the use of evolving fuzzy systems for designing a method that overcomes the problems associated with medical image segmentation. An evolving fuzzy system can be trained using a set of invariant features, along with their optimum parameters, which act as a target for the system. Evolving fuzzy systems are also capable of adjusting parameters based on online updates of their rule base. This thesis proposes three different approaches that employ an evolving fuzzy system for the continual adjustment of the parameters of any medical image segmentation technique. The first proposed approach is based on evolving fuzzy image segmentation (EFIS). EFIS can adjust the parameters of existing segmentation methods and switch between them or fuse their results. The evolving rules have been applied for breast ultrasound images, with EFIS being used to adjust the parameters of three segmentation methods: global thresholding, region growing, and statistical region merging. The results for ten independent experiments for each of the three methods show average increases in accuracy of 5\%, 12\% and 9\% respectively. A comparison of the EFIS results with those obtained using five other thresholding methods revealed improvements. On the other hand, EFIS has some weak points, such as some fixed parameters and an inefficient feature calculation process. The second approach proposed as a means of overcoming the problems with EFIS is a new version of EFIS, called self-configuring EFIS (SC-EFIS). SC-EFIS uses the available data to estimate all of the parameters that are fixed in EFIS and has a feature selection process that selects suitable features based on current data. SC-EFIS was evaluated using the same three methods as for EFIS. The results show that SC-EFIS is competitive with EFIS but provides a higher level of automation. In the third approach, SC-EFIS is used to dynamically adjust more than one parameter, for example, three parameters of the normalized cut (N-cut) segmentation technique. This method, called multi-parametric SC-EFIS (MSC-EFIS), was applied to magnetic resonance images (MRIs) of the bladder and to breast ultrasound images. The results show the ability of MSC-EFIS to adjust multiple parameters. For ten independent experiments for each of the bladder and the breast images, this approach produced average accuracies that are 8\% and 16\% higher respectively, compared with their default values. The experimental results indicate that the proposed algorithms show significant promise in enhancing image segmentation, especially for medical applications.

Ultrasound is one of the primary technologies used for clinical purposes. Ultrasound systems have favorable real-time capabilities, are fast and relatively inexpensive, portable and non-invasive. Recent interest in using ultrasound imaging for tongue motion has various applications in linguistic study, speech therapy as well as in foreign language education, where visual-feedback of tongue motion complements conventional audio feedback. Ultrasound images are known to be difficult to recognize. The anatomical structure in them, the rapidity of tongue movements, also missing segments in some frames and the limited frame rate of ultrasound systems have made automatic tongue contour extraction and tracking very challenging and especially hard for real-time applications. Traditional image processing-based approaches have many practical limitations in terms of automation, speed, and accuracy. Recent progress in deep convolutional neural networks has been successfully exploited in a variety of computer vision problems such as detection, classification, and segmentation. In the past few years, deep belief networks for tongue segmentation and convolutional neural networks for the classification of tongue motion have been proposed. However, none of these claim fully-automatic or real-time performance. U-Net is one of the most popular deep learning algorithms for image segmentation, and it is composed of several convolutions and deconvolution layers. In this thesis, we proposed a fully automatic system to extract tongue dorsum from ultrasound videos in real-time using a simplified version of U-Net, which we call sU-Net. Two databases from different machines were collected, and different training schemes were applied for testing the learning capability of the model. Our experiment on ultrasound video data demonstrates that the proposed method is very competitive compared with other methods in terms of performance and accuracy.

No description available.

The goal of fully decoding how the brain works requires a detailed wiring diagram of the brain network that reveals the complete connectivity matrix. Recent advances in high-throughput 3D electron microscopy (EM) image acquisition techniques have made it possible to obtain high-resolution 3D imaging data that allows researchers to follow axons and dendrites and to identify pre-synaptic and post-synaptic sites, enabling the reconstruction of detailed neural circuits of the nervous system at the level of synapses. However, these massive data sets pose unique challenges to structural reconstruction because the inevitable staining noise, incomplete boundaries, and inhomogeneous staining intensities increase difficulty of 3D reconstruction and visualization. In this dissertation, a new set of algorithms are provided for reconstruction of neuronal morphology from stacks of serial EM images. These algorithms include (1) segmentation algorithms for obtaining the full geometry of neural circuits, (2) interactive segmentation tools for manual correction of erroneous segmentations, and (3) a validation method for obtaining a topologically correct segmentation when a set of segmentation alternatives are available. Experimental results obtained by using EM images containing densely packed cells demonstrate that (1) the proposed segmentation methods can successfully reconstruct full anatomical structures from EM images, (2) the editing tools provide a way for the user to easily and quickly refine incorrect segmentations, (3) and the validation method is effective in combining multiple segmentation results. The algorithms presented in this dissertation are expected to contribute to the reconstruction of the connectome and to open new directions in the development of reconstruction methods.

Medical image analysis is a ubiquitous and essential part of modem health care. A crucial first step to this is segmentation, which is often complicated by many factors including subject diversity, pathology, noise corruption, and poor image resolution. Traditionally, manual tracing by experts was done. While considered accurate, this process is time consuming and tedious, especially when performed slice-by-slice on three-dimensional (3D) images over large datasets or on two-dimensional (2D) but topologically complicated images such as a retinography. On the other hand, fully-automated methods are typically faster, but work best with data-dependent, carefully tuned parameters and still require user validation and refinement. This thesis contributes to the field of medical image segmentation by proposing a highly-automated, interactive approach that effectively merges user knowledge and efficient computing. To this end, our work focuses on graph-based methods and offer globally optimal solutions. First, we present a novel method for 3D segmentation based on a 3D Livewire approach. This approach is an extension of the 2D Livewire framework, and this method is capable of handling objects with large protrusions, concavities, branching, and complex arbitrary topologies. Second, we propose a method for efficiently segmenting 2D vascular networks, called ‘Live-Vessel’. Live-Vessel simultaneously extracts vessel centrelines and boundary points, and globally optimizes over both spatial variables and vessel radius. Both of our proposed methods are validated on synthetic data, real medical data, and are shown to be highly reproducible, accurate, and efficient. Also, they were shown to be resilient to high amounts of noise and insensitive to internal parameterization.

Medical image analysis is a ubiquitous and essential part of modem health care. A crucial first step to this is segmentation, which is often complicated by many factors including subject diversity, pathology, noise corruption, and poor image resolution. Traditionally, manual tracing by experts was done. While considered accurate, this process is time consuming and tedious, especially when performed slice-by-slice on three-dimensional (3D) images over large datasets or on two-dimensional (2D) but topologically complicated images such as a retinography. On the other hand, fully-automated methods are typically faster, but work best with data-dependent, carefully tuned parameters and still require user validation and refinement. This thesis contributes to the field of medical image segmentation by proposing a highly-automated, interactive approach that effectively merges user knowledge and efficient computing. To this end, our work focuses on graph-based methods and offer globally optimal solutions. First, we present a novel method for 3D segmentation based on a 3D Livewire approach. This approach is an extension of the 2D Livewire framework, and this method is capable of handling objects with large protrusions, concavities, branching, and complex arbitrary topologies. Second, we propose a method for efficiently segmenting 2D vascular networks, called ‘Live-Vessel’. Live-Vessel simultaneously extracts vessel centrelines and boundary points, and globally optimizes over both spatial variables and vessel radius. Both of our proposed methods are validated on synthetic data, real medical data, and are shown to be highly reproducible, accurate, and efficient. Also, they were shown to be resilient to high amounts of noise and insensitive to internal parameterization.

No description available.

Breast cancer is currently the most common diagnosed cancer among women and a significant cause of death. Breast density is considered a significant risk factor and an important biomarker influencing the later risk of breast cancer. Therefore, ongoing epidemiological studies using MRI are evaluating quantitatively breast density in young women. One of the challenges is segmenting the breast in order to calculate total breast volume and exclude non-breast surrounding tissues. This thesis describes an automatic 3D breast volume segmentation based on 3D local edge detection using phase congruency and Poisson surface reconstruction to extract the total breast volume in 3D. The boundary localization framework is integrated on a subsequent atlas-based segmentation using a Laplacian framework. The 3D segmentation achieves breast-air and breast-chest wall boundary localization errors with a median of 1.36 mm and 2.68 mm respectively when tested on 409 MRI datasets.

Breast cancer is currently the most common diagnosed cancer among women and a significant cause of death. Breast density is considered a significant risk factor and an important biomarker influencing the later risk of breast cancer. Therefore, ongoing epidemiological studies using MRI are evaluating quantitatively breast density in young women. One of the challenges is segmenting the breast in order to calculate total breast volume and exclude non-breast surrounding tissues. This thesis describes an automatic 3D breast volume segmentation based on 3D local edge detection using phase congruency and Poisson surface reconstruction to extract the total breast volume in 3D. The boundary localization framework is integrated on a subsequent atlas-based segmentation using a Laplacian framework. The 3D segmentation achieves breast-air and breast-chest wall boundary localization errors with a median of 1.36 mm and 2.68 mm respectively when tested on 409 MRI datasets.

This thesis develops image segmentation methods for the application of automated cervical cancer screening. The traditional approach to automating this task has been to emulate the human method of screening, where every one of the hundreds of thousands of cells on each slide is analysed for abnormality. However, due to the complexity of cervical smear images and the low error tolerance imposed upon the segmentation stage, only limited success has previously been found. A different approach is to detect malignancy associated changes (MACs) in a relatively small sample of the total population of cells. Under this paradigm, the requirement to segment every cell is loosened, but delineation accuracy and error checking become essential. Following a review of generic and cervical smear segmentation, it is concluded that prior work on the traditional approach to automation is not suitable for a MACs solution. However, the previously proposed framework of a dual-magnification system is found to be relevant and is therefore adopted. Here, scene images are first captured at low resolution in order to rapidly locate the cells on a slide. Cells that are deemed to be suitable for further analysis are then imaged at high resolution for the more accurate segmentation of their nuclei. A water immersion algorithm is developed for low resolution scene segmentation. This method achieves a rapid and robust initial segmentation of the scene without the requirement of incorporating extensive a priori knowledge of the image objects. A global minimum searching contour is presented as a top-down method for segmenting the high resolution cell nucleus images where the image objects are well characterised by shape and appearance. This latter method is tested upon 20,000 images and found to achieve an accurate segmentation rate of 99.47%. An error checking method, that uses segmentation stability as an indicator of segmentation success, is developed that is capable of detecting 100% of the failures of the nucleus segmenter, at the expense of discarding only 9% of the data. Throughout this work, contemporary issues in the field of generic image segmentation are presented and some of these are addressed for the cervical smear application. Finally, an avenue of future work is proposed which may lead to the much wider proliferation of computer vision solutions to everyday problems.