Wavelet compression of active appearance models

Abstract

Active Appearance Models (AAMs) provide a method of modelling the appearance of anatomical structures in medical images and locating them automatically. Although the AAM approach is computationally efficient, the models used to search unseen images for the structures of interest are large - typically the size of 100 images. This is perfectly practical for most 2-D images, but is currently impractical for 3-D images. We present a method for compressing the model information using a wavelet transform. The transform is applied to a set of training images in a shape-normalised frame, and coefficients of low variance across the training set are removed to reduce the information stored. An AAM is built from the training set using the wavelet coefficients rather than the raw intensities. We show that reliable image interpretation results can be obtained at a compression ratio of 20:1, which is sufficient to make 3-D AAMs a practical proposition.