Segmentation of magnetic resonance images of the thorax by backpropagation

Segmentation of images-especially medical images-is an important problem for which automatic solutions are urgently sought. In this paper, we report on work in which neural networks are trained by backpropagation to segment magnetic resonance (MR) images of the thorax, by classifying pixels as either boundary (pixels on the boundary between lung interior and surrounding tissue) or non-boundary. Networks trained on part of a single image slice from a particular patient produce an output for the whole slice in which the lung outline is considerably enhanced. They are also able to generalise successfully to other slices from the same patient