Intelligent Strategies for Image Segmentation (ISIS)

description

In many of our collaborations, the need for identifying and measuring anatomical structures arises frequently. The Segmentation Group is involved in developing and applying new techniques to the segmentation problems driven by our collaborations. The range of segmentation problems that we address is wide and includes

Statistical Segmentation Methods

An important class of statistical methods for image segmentation relies on Markov random fields (MRFs). Typical MRF-based strategies for biomedical image segmentation model the local/Markov dependencies on the labels, parametrically, which often fail to effectively capture the structure and variability in the intensity data. We apply an adaptive MRF model to the intensity image by nonparametrically inferring the Markov statistics from the input image itself. In this way, we avoid imposing ad hoc models that may not conform well with the data. Subsequently, we segment the data based on the inferred model. This strategy helps us automatically deal with the noise, inhomogeneity, and partial voluming in MR images. We leverage this approach for segmenting tissues or structures in adult and neonatal, MR and DT images.

Finding the best segmentation (in terms of Bayesian measures) for the N-label case can become computationally expensive. However, constraining the segmentation scenario to the 2-label problem conduces a quick solution by means of a minimizing graph cut on the appropriately constructed image graph. Graph construction occurs via spatial continuity priors derived from MRF modelling as well as likelihood measurements from probabilistic atlas construction.

Template-Based Methods

Anatomical templates make it possible to incorporate complex expert knowledge in automatic segmentation algorithms. Our work includes developing advanced techniques for population-specific template generation and image normalization that allow the entire brain to be segmented as a whole, making it possible to estimate the locations of subcortical structures and Brodmann areas. In addition, we are developing segmentation techniques that focus on individual structures and encode the statistical model of the structures' shape and appearance in the template. These structure-specific techniques leverage our medial modeling framework.

Support for Manual Expert-Driven Segmentation

We are the main development site for the user-friendly tool ITK-SNAP that brings the complexity of level set methods to the fingertips of neuroanatomists and provides a simple but powerful interface for manual segmentation in three dimensions.

bibliography