ABSTRACT
In this section we present a general-purpose neural architecture for segmenting two-dimensional and three-dimensional medical images. The architecture is based on a continuous Hopfield neural network including one or more sets of two-dimensional layers of neurons with local connections. This architecture can be specialized to perform the segmentation of two-dimensional images, the multiscale segmentation of two-dimensional images and the segmentation of three-dimensional images by simply changing the number of such sets and/or the size of the component layers. By changing synaptic weights the architecture can adapt to the differences existing between tomographic and radiographic images. The segmentation produced by this architecture is optimum with respect to a ‘goodness’ criterion which establishes the tradeoff between sensitivity and robustness. The section describes the derivation of the architecture and some experimental results obtained with synthetic and real medical images.
