Model-based segmentation techniques incorporate a-priori knowledge (shape, intensity, etc. – together with their allowed variations) about the structures to segment in the segmentation process. This makes the segmentation of heterogeneous or incomplete objects possible.
My focus lies on explicit dynamic models, especially on Stable Mass Spring Models (SMSMs), which were invented by our group. SMSMs are three-dimensional models, which are stable and flexible at the same time. Stability is achieved by means of a torsion force. Before, the lack of stability was the major problem of three-dimensional dynamic models for segmentation. SMSMs are also explicit, physically-based models, which make interaction with them an easy and natural process.
Because of efficient implementation, simulation of the model dynamics takes place at high step rates allowing for a fluent visualization of the optimization process and a fast automatic segmentation. Furthermore, fast simulation and visualization enable on-the-fly interaction with the model.
But before any model-based segmentation can take place, a model itself is needed. Model generation is a complex task that should be done automatically. Therefore, demands for such models must be known. Until now, we can generate models automatically from a sample object in a dataset using heuristic techniques.
SPECTSegmenter is an experimental software initially developed for the model-based segmentation of the left ventricle in 3D SPECT data in cooperation with MIRG. It is a full-fledged experimental environment for Stable Mass Spring Models (SMSMs). Besides automatic segmentation, a lot of interaction at all stages of segmentation is possible. Also, automatic model generation from sample segmentations is possible.
One of the most important features is the ability to debug models. So, the behavior of single model parts can be observed over the time allowing for straight improvement of the model performances. The complex 2D and 3D visualization possibilities facilitate a natural understanding of the model behaviors.