COMBINING MACHINE LEARNING TECHNIQUES WITH STATISTICAL SHAPE MODELS IN MEDICAL IMAGE SEGMENTATION

ABSTRACT

In this thesis, we implemented Point Distribution Model and basic Active Shape Model algorithm and contributed this to the AUST Computer Vision and Machine Learning code library. We applied the Active Shape Model to segmenting lateral ventricles of 2D brain images and used machine learning – specifically K-Nearest Neighbour algorithm- to improve segmentation results. A statistical shape model is created from a training dataset which is used to search for an object of interest in an image. Active shape model has shown over time to be a reliable image segmentation methodology but its segmentation accuracy is hindered especially by poor initialization which can’t be guaranteed to always be perfect. In our methodology, we extract features for each landmark using Haar filters. We train a classifier with these features and use the classifier to classify points around the final points of an Active shape model search. The aim of this approach is to better place points that might have been wrongly placed from the ASM search. We have used the simple, yet effective K-Nearest Neighbour machine learning algorithm, and have demonstrated the ability of this method to improve segmentation accuracy by segmenting lateral ventricles of the brain.

CHAPTER ONE

1.1 INTRODUCTION

In this thesis, our aim is to segment images, specifically, medical images. We aim to implement the popular Active Shape Model algorithm [16] and demonstrate its usefulness in segmenting 2d medical images. Furthermore, we explore improving the results of the Active Shape Model segmentation using machine learning techniques.

Predominantly, the aim of medical image segmentation is to label each pixel in an image to indicate the anatomical structure it belongs to and delineate such structures of interest for the purposes of visualization, diagnosis or medical research. Segmentation is often a crucial first step in patient diagnosis especially when qualitative and quantitative information about appearance, size, or shape of patient anatomy is desired. Results of medical image segmentation are useful for many purposes including image guided surgery, detection of anatomical changes over time, detection of pathological diseases, volumetric measurement, visualization and research. With the increasing importance of the segmentation process in diagnosis, accuracy of the process is important, as this may impact diagnostic accuracy, treatment planning and subsequently treatment.

The segmentation process is unfortunately as difficult as it is important, and the reasons are easy to comprehend. Computers are not half as good as humans when it comes to ill-defined problems such as object recognition, and when these images contain noise, it makes the process even more difficult for a computer. More often than not, images will be noisy, altering the intensity values of some pixels. This could make anatomical structures difficult to separate from their surroundings, and strong edges may not be present around its borders. Sometimes, the intensity level of a single tissue class varies gradually over the image –a phenomenon known as intensity inhomogeneity or non-uniformity – and this doesn’t make the segmentation task any easier. Other times, an individual pixel may contain mixture of tissue classes such that intensity of a pixel in the image may not be consistent with one class. The gray levels of different tissues, if too close would increase the difficulty of the process. These problems and the variability in the tissue distribution among individuals in the human population means that some degree of uncertainty must be attached to all segmentation results. Segmentation can be done manually, semi-automatically or can be a fully automated process. Manual segmentation is a time consuming task and with the volume of medical image data needed to be processed, it is highly unlikely to be the ideal method considering the fact that results of such a process would depend on operator variability and thus would be difficult to reproduce. The level of confidence ascribed to manual processes suffers accordingly. Automatic methods overcome these drawbacks and are preferred especially with the computing resources available today and the amount of data needed to be processed. However, accurate automatic segmentation is by no means an easy feat and remains an active area of research in computer vision. There are more segmentation methods than can be mentioned in this thesis, with abundant literature on most of them. Segmentation methods range from earlier intensity based approaches such as thresholding and region growing to pattern recognition approaches such as neural networks and model based approaches such as Active Shape and Appearance models. Thresholding approaches segment images by creating a binary partitioning of the image intensities. A typical thresholding approach attempts to determine an intensity value called the threshold which can separate the image into desired classes. This method of segmentation is simple yet effective when the image that is being segmented contains structures with contrasting intensities. Region growing extracts an image region that is connected to a point called the seed point – usually manually selected – based on some predefined criteria which can be based on intensity and/or edge information [22]. Pattern recognition techniques seek to partition a feature space derived from an image by using data with known labels. These techniques are supervised methods since they require training data that are manually segmented and then used as references for segmenting new data. Clustering algorithms perform segmentation without training data and are termed unsupervised methods. Deformable models are model based techniques for segmenting images by using closed parametric curves that deform under the influence of internal and external forces. To delineate an object boundary in an image, a closed curve must first be placed near the desired boundary and allowed to undergo an iterative “deformation” process. Deformable models include the Active contour model (snakes) and the Active Shape Model (Smart snakes). In this thesis however, our focus will be on Active Shape Model.

Department: Computer Science (M.Sc Thesis)

Format: MS Word

Chapters: 1 - 5, Preliminary Pages, Abstract, References, Appendix.

No. of Pages: 67

Price: 20,000 NGN

In Stock

Buy Now

Our Customers are Happy!!!

See What Some of Our Customers are Saying About Us