Supervised dimensionality reduction and contextual pattern recognition in medical image processing

The past few years have witnessed a significant increase in the number of supervised methods employed in diverse image processing tasks. Especially in medical image analysis the use of, for example, supervised shape and appearance modelling has increased considerably and has proven to be successful. This thesis focuses on applying supervised pattern recognition methods in medical image processing. We consider a local, pixel-based approach in which image segmentation, regression, and filtering tasks are solved using descriptors of the local image content (features) based on which decisions are made that provide a class label (in case of image segmentation) or a gray value (in case of filtering or regression) for every pixel. The basic probabilistic decision problem, underlying---implicitly or explicitly---all the methods presented in this thesis, can be stated in terms of a conditional probability optimization problem u = argmax_y P(y|x) in which x is a d-dimensional vector of measurements, i.e., a feature vector, describing the local image content and y is a quantity that takes values from a set Y. Typically, in a classification task, Y is a discrete set of labels and in case of regression, Y equals R. Based on the maximization in the previous equation, to every vector x (which is associated to a pixel in an image), a particular u from Y is associated. This approach is---because of its local nature---quite different from the shape and appearance methods mentioned in the beginning of this chapter which try to solve image processing tasks in a more global way. A recent comparative study [B. van Ginneken, M. B. Stegmann, and M. Loog. Segmentation of anatomical structures in chest radiographs using supervised methods: a comparative study on a public database, Medical Image Analysis, accepted, 2005] shows that in image segmentation, pixel-based approaches can compete with shape and appearance models, providing an interesting alternative to the latter. The principal methodological part of the thesis consists of three dimensionality reduction methods that can aid the extraction of relevant features to be used for performing image segmentation or regression. Furthermore, an iterative segmentation scheme is developed which draws from classical pattern recognition and machine learning methods. Finally, two applications of these techniques in two problems related to computer-aided diagnosis (CAD) in chest radiography are presented. Firstly, the task of segmenting the posterior ribs is considered. Secondly, a regression framework is presented, which aims at suppressing bony structures in chest radiographs.