In order to perform this fully automatic segmentation on individual patient data sets an “Articulated Atlas” model is trained from a set of training data first.
Here, the shapes and sizes of the individual structures (bones, soft tissue and tubular structures) are learned. Therefore, statistical shape models are used, which contain an average model of the structure as well as 95% of the seen training data sets. That means the model is highly variable but on the other hand limits the structures to cover only natural deformations.
In addition, the positions of the structures to each other are learned. This ensures that all of the structures in the final segmentation are located correctly and prevents that any structure is completely misplaced.
Finally, the “Articulated atlas” is transferred to an individual patient data set to perform an automatic segmentation. Here, the previously trained knowledge of the shapes and sizes of the individual structures and the relative positions to one another is used to achieve an accurate fully automatic segmentation.