Three-dimensional image classification using hierarchical spatial decomposition: a study using retinal data

This thesis describes research conducted in the field of image mining especially volumetric image mining. The study investigates volumetric representation techniques based on hierarchical spatial decomposition to classify three-dimensional (3D) images. The aim of this study was to investigate the effectiveness of using hierarchical spatial decomposition coupled with regional homogeneity in the context of volumetric data representation. The proposed methods involve the following: (i) decomposition, (ii) representation, (iii) single feature vector generation and (iv) classifier generation. In the decomposition step, a given image (volume) is recursively decomposed until either homogeneous regions or a predefined maximum level are reached. For measuring the regional homogeneity, different critical functions are proposed. These critical functions are based on histograms of a given region. Once the image is decomposed, two representation methods are proposed: (i) to represent the decomposition using regions identified in the decomposition (region-based) or (ii) to represent the entire decomposition (whole image-based). The first method is based on individual regions, whereby each decomposed sub-volume (region) is represented in terms of different statistical and histogram-based techniques. Feature vector generation techniques are used to convert the set of feature vectors for each sub-volume into a single feature vector. In the whole image-based representation method, a tree is used to represent each image. Each node in the tree represents a region (sub-volume) using a single value and each edge describes the difference between the node and its parent node. A frequent sub-tree mining technique was adapted to identified a set of frequent sub-graphs. Selected sub-graphs are then used to build a feature vector for each image. In both cases, a standard classifier generator is applied, to the generated feature vectors, to model and predict the class of each image. Evaluation was conducted with respect to retinal optical coherence tomography images in terms of identifying Age-related Macular Degeneration (AMD). Two types of evaluation were used: (i) classification performance evaluation and (ii) statistical significance testing using ANalysis Of VAriance (ANOVA). The evaluation revealed that the proposed methods were effective for classifying 3D retinal images. It is consequently argued that the approaches are generic.