Translation-invariant aorta segmentation from magnetic resonance images
William T. Katz, Merickel · 1989
A backpropagation neural network has been applied to the problem of segmentation of the aorta from MRI images. In order to achieve translationally invariant classification the network's input units are loaded from a small receptive field which is moved across the images in a uniformly random manner. Three-dimensional graphic analysis of the input space and hidden-unit output space provides some insight into the network's progress during training and the evolved internal representation. Because of overlapping tissue signatures, spatial as well as statistical information is required for segmentation. Results indicate that a network with a receptive field of comparable size to the target region can compensate for the intermixing of aorta and nonaorta points by using contextual information. Neural networks that achieve perfect translationally invariant segmentation of training set images and up to 98% generalization to novel images from the same subject have been developed. Generalization across patients, however, is not currently possible owing to interpatient variations in tissue MRI signatures.>