Segmentation

The principle and applicability of an Active Appearance model based on the analysis of 3D deformation fields for segmentation has been demonstrated. Segmentation speed is much higher than ANIMAL or expert manual segmentation. Its accuracy, measured by statistics of overlap with respect to manual segmentation, is virtually equal to that of ANIMAL. A $\kappa$ of 0.7 is usually considered to represent good agreement between labellings. While the values found here are marginally lower, one must take into account the fact that the hippocampus has a large surface-to-volume ratio, so that even small errors of surface agreement adversely affect $\kappa$. For example, a single hippocampus, compared with itself after a displacement of only 1 mm, results in $\kappa=0.80$. It should also be mentioned that the inter-observer variability (5 observers) for the manual segmentation of the left hippocampus of the same subjects, was $\kappa = 0.86$ [5]. Hence, early results from the non-optimized AB method indicate that it measures favorably against established automated techniques and human observers, while being significantly faster.

The AB model could theoretically be more accurate than ANIMAL, since it incorporates prior information in its model, and is given a larger field for convergence. Improvements in the choice of training parameters, for example in the finding of matrix ${\bf A}$ or using a better iterative algorithm, should increase accuracy. Robustness remains to be determined. Of particular interest will be the segmentation results of patients with neuro-degenerative disorders using such a model-based technique trained on normal subjects. It is conceivable that additional training sets may be necessary to properly address each structural/pathological situation.

For each model, before performing PCA, the mean element (mean grey-level image, mean x,y,z warp) is substracted from each element. Analysis is therefore performed on the difference vectors, and the principal directions explain the variations around the mean. Consequently a large number of eigenvectors needs to be kept in each model in order to reach the desired percentage of explained variation (f).

Features of this successful novel approach include (1) its speed compared to locally available segmentation methods; (2) its reliance on all grey-level voxels and deformation vectors as ``landmarks'' and hence maximum use of information; (3) 3D and automated; and (4) flexibility in the choice of SOI/VOI. The major constraint is the restriction to the domain of structure neighborhoods whose non-linear registration is achieved using ANIMAL for the creation of the 3D Warp Model.