Classification of biomedical high-resolution micro-CT images for direct volume rendering

This paper introduces a machine learning approach into the process of direct volume rendering of biomedical highresolution 3D images. More concretely, it proposes a learning pipeline process that generates the classification function within the optical property function used for rendering. Briefly, this pipeline starts with a data acquisition and selection task, it is followed by a feature extraction process, to be ended with sequence of supervised learning steps. Learning comprises Gentle Boost and CRF (Conditional Random Fields) classifiers. The process is evaluated in terms of accuracy and overlap metrics so that we can measure how performance increases along the whole pipeline process. Empirical results confirm that, even though the classification of high-resolution computerized tomography volume data poses a challenging problem for single-run classifiers, it can be significantly improved by subsequent learning steps and refinements.