Abstract
The interpretation of medical images is one of the most difficult tasks in computer vision, largely because of the high degree of variability associated with normal and abnormal appearances. This thesis introduces a systematic method for the detection of microcalcifications as one of the most important signs of early breast cancer. It involves a four step procedure. The first step is blob detection to detect regions of microcalcification size range. The second step involves a specially designed directional region growing method to find the best fitting boundaries for each blob region. A newly developed combination of classifiers is then applied to label each region as a microcalcification or background. The final processing step involves a search for the existence of clusters of microcalcifications using a hierarchical nearest mean clustering method. The contributions of the work to the field of image processing are; a new blob detection system; a novel region growing method and a theoretical framework for combining classifiers which use a combination of shared and distinct representations. Here specifically, we present a blob detection method with the capability of detecting any suspected blob of specific size range. Then a new region growing method is developed based on a unique directional growing process providing predictable behaviour for the method. The application of two discontinuity measures is considered for the extraction of two fitting boundaries representing information about the region and its local background. The information conveyed by the boundaries and their associated regions is used to compute reliable representations for labelling each blob region. The robustness of the region growing method to the choice of a starting point and to Gaussian noise is examined on real images. We demonstrate that commonly used classifiers provide reliable results in labelling the suspected regions. In spite of achieving an acceptable performance using different individual classifiers, a decision fusion rule involving a weighted combination of classifiers is developed and its performance on the problem is investigated. The combination rule is applicable when mixed mode representations (some shared and some individual features) are used. A comparative study of the individtial classifiers and also of conventional classifier combination techniques with the weighted combiner is performed on independent test sets. The results achieved with the presented algorithm are very promising and approaching a level where a clinical pilot evaluation for screening purposes would be warranted.