Contextual Biomedical Image Learning S. Kevin Zhou A biomedical image characterizes rich contextual information that is defined as the interrelationships among shape, appearance, motion, geometry, imaging modality, disease, biology, etc. However, most algorithms either ignore such biomedical image context or partially use context; they resort to linear or parametric models and Gaussian noise assumption; and they need manual interaction. In this talk, I will present novel machine learning approaches that leverage biomedical image context for more effective and efficient analysis of biomedical images. In particular, I will address two methods: (i) Shape Regression Machine (SRM) for deformable shape detection and segmentation and (ii) BoostMotion for deformable shape motion estimation. With no assumption on the data distribution, these nonparametric methods, grounded on a unifying learning framework of boosting, transfer contextual knowledge from expert-annotated database into machine use. Boosting iteratively selects for the given task relevant visual features, which are fast to evaluate and hence enable real time processing. I will illustrate the benefits of context learning for biomedical image analysis using real time demonstrations.
|