Image data can be categorized as two-dimensional (2D) and three-dimensional (3D). The data in 2D image is called pixel and voxel in 3D image otherwise. 2D image has being insensitively studied in the image processing community, with many algorithms proposed to mine the pixel data. As contrast, the medical image data often appear in the form of 3D. However, the voxel data mining algorithms are not as abundant as pixel data mining. Hence, we attempt to propose new methods that are capable of mining useful features from voxels. Based on the mined features, we can use them for medical image classification that can serve as computer-aided diagnosis (CAD) tools.
Positron emission tomography (PET) is a modern 3D imaging technique, containing discriminative information in diagnosing dementia disease. However, mining the high-dimensional image is a great challenge, since the discriminative information is only a very small amount. We develop a Gaussian mixture model (GMM) with model selection approach to automatically extracting useful features, which can be further used to classify different groups of images. The voxels can be clustered via a GMM, and the model selection method finds the optimal number of clusters. The resulting clusters contain relevant information that is known to be related to dementia. Besides, we also attempt to improve the clustering approach by allowing the number of clusters to be chosen without a priori. To this end, the infinite Gaussian mixture model, a non-parametric model, is an alternative to GMM with model selection method. In summary, the statistical machine learning methods facilitate us in deriving useful voxel patterns.
Figure: 2D view of a PET image with discovered discriminative voxels.