We applied a learning network to a cell's boundary detection of human corneal endothelium photomicrographs measured by specular microscopy. Interconnections between units in our model are constrained to be locally space invariant to meet space-invariant processing. The neural network was trained to extract the cell's boundary by showing part of the photomicrograph and its subjective boundary image, which is an outline drawing made by hand. After training, the network showed good performance with the microphotograph that was not trained. Internal representations of the network were also studied.