Volumetric image mining based on decomposition and graph analysis: An application to retinal optical coherence tomography

In this paper, we introduce a method for classifying volumetric images using a decomposition and graph analysis based method. Volume decomposition plays an important role in simplifying Three-Dimensional (3-D) volumes so as to enhance their analysis. Hierarchical decomposition techniques incrementally divide a given 3-D volume into sub-volumes according to some critical function and then (typically) represent the decomposition as a tree. The issue is the nature of the critical function which dictates when the decomposition should be stopped. Broadly the decomposition should be stopped whenever homogeneous sub-volumes are reached. The question is how we define homogeneity in this context. In this paper a number of different critical functions are evaluated. The evaluation was conducted by considering the classification of 3-D Optical Coherence Tomography (OCT) retinal images according to whether they feature Age-related Macular Degeneration (AMD) or not. The OCT volumes were encoded using the proposed hierarchical tree decomposition coupled with a number of different critical functions. A frequent sub-graph mining algorithm was then applied to the tree representations and the resulting identified frequent sub-graphs used to define a feature vector encoding which was then fed into a standard classifier.