Automated Domain-Specific Feature Selection for Classification-based Segmentation of Tomographic Medical Image Data

Classification-based segmentation is an approach to establish generic analysis of medical image data. Significant feature sets covering different characteristics of regions to segment allow for robust discrimination of topologically defined classes. In this work a method for automated domain-specific feature selection to achieve a higher level of predictability is presented, incorporating multivariate feature analysis. For calculation of the probability density function, different approaches, like histogram analysis, enumeration of the entire feature space or umbrella Monte Carlo Integration are investigated. Furthermore, meta features calculated on entire classification results rather than on particular regions are introduced. Predictability of both, single local and meta features, is evaluated for different medical datasets as well for simulated intensity volumes, allowing testing and evaluating specific classification problems. The automated feature selection proofs to be accurate for classification-based segmentation utilizing well-known machine learning approaches.