Automated Multiscale Segmentation of Volumetric Biomedical Imagery Based on a Markov Random Field Model

A fully automated volumetric image segmentation algorithm is proposed which uses Bayesian inference to assess the appropriate number of image segments. The segmentation is performed exclusively within the wavelet domain, after the application of the redundant á trous wavelet transform employing four decomposition levels. This type of analysis allows for the evaluation of spatial relationships between objects in an image at multiple scales, exploiting the image characteristics matched to a particular scale. These could possibly go undetected in other analysis techniques. The Bayes Information Criterion (BIG) is calculated for a range of segment numbers with a relative maximum determining optimal segment number selection. The fundamental idea of the BIC is to approximate the integrated likelihood in the Bayes factor and then ignore terms which do not increase quickly with N, where N is the cardinality of the data. Gaussian Mixture Modelling (GMM) is then applied to an individual mid-level wavelet scale to achieve a baseline scene estimate considering only voxel intensities. This estimate is then refined using a series of wavelet scales in a multiband manner to reflect spatial and multiresolution correlations within the image, by means of a Markov Random Field Model (MRFM). This approach delivers promising results for a number of volumetric brain MR and PET images, with inherent image features being identified. Results achieved largely correspond with those obtained by researchers in biomedical imaging utilising manually defined parameters for image modelling.