Data-driven synthesis of composite-feature detectors for 3D image analysis

Most image analysis techniques are based upon low level descriptions of the data. It is important that the chosen representation is able to discriminate as much as possible among independent image features. In particular, this is of great importance in segmentation with deformable models, which must be guided to the target object boundary avoiding other image features. In this paper, we present a multiresolution method for the decomposition of a volumetric image into its most relevant visual patterns, which we define as features associated to local energy maxima of the image. The method involves the clustering of a set of predefined band-pass energy filters according to their ability to segregate the different features in the image. In this way, the method generates a set of composite-feature detectors tuned to the specific visual patterns present in the data. Clustering is accomplished by defining a distance metric between the frequency features that reflects the degree of alignment of their energy maxima. This distance is related to the mutual information of their responses' energy maps. As will be shown, the method is able to isolate the frequency components of independent visual patterns in 3D images. We have applied this composite-feature detection method to the initialization of active models. Among the visual patterns detected, those associated to the segmentation target are selected by user interaction to define the initial state of a geodesic active model. We will demonstrate that this initialization technique facilitates the evolution of the model to the proper boundary.