Medical image segmentation using a tree model

A model-driven, multiscale medical image segmentation system is presented. A tree representation is calculated for the image, using a modification of the immersion algorithm used for watersheds calculation. Segmentation is carried out by a matching process between the obtained tree and a tree model, which embeds the prior knowledge about the images. Tree matching is done in a multilevel way, processing different tree levels sequentially. For each level, an optimization process is performed, in which an error function, obtained from differences between the model and the segmented tree, is minimized. 13 parameters, concerning gray level, shape, position and connectivity, are used to characterize the objects. The model is obtained from a set of training images, assigning manual labels to tree nodes with a user interface designed especially for this purpose. Three-dimensional, multicomponent images can be processed by adapting gradient and parameter calculation. The system has been tested for intracranial cavity segmentation in magnetic resonance images, giving accurate results.