Texture-Adaptive Active Contour Models

Unsupervised segmentation is a key challenge for automated quantification of medical images. Although a balloon model is able to detect arbitrarily shaped objects in images, it requires careful adjustment of parameters prior to segmentation. Based on global texture analyses, our method allows to set these parameters automatically for heterogeneous images such as MRI, ultrasound, or microscopy. Cooccurrence matrices are extracted from prototype images and used as feature vectors to train a synergetic classifier. These matrices are computed likewise for all other images. To control segmentation, similarity measures for these features are applied to weight the linear combination of the prototype parameters. The method was tested on 81 synthetic images and applied to a set of 1616 heterogeneous radiographs. Setting the parameters of active contour models by the proposed method improves the acceptance rate of unsupervised segmentation from 31% up to 71%.