11 November 1991 Neural-network-based image processing of human corneal endothelial micrograms
Author Affiliations +
This report presents an application of a learning network to the detection of cell membranes in human corneal endothelial micrograms. Our neural network model is a multilayered feed- forward network, and units in any single layer are divided into clusters. Every unit in the higher layer is connected with some of the units in each cluster of the lower layer. Units in the same layer have the same size of receptive field. In order to perform space-invariant processing in the same cluster, units in the same cluster have the same pattern of connectivity, but units in the different clusters have a different one. Such a network has been shown to be robust against distortions of input patterns and to match well with optical implementations. The neural network is trained by small parts of a microgram to extract the boundaries of the endothelial cells using the supervised learning algorithm. Desired output images are their cell membrane images that are traced by hand. After training, the network showed good performance with the whole microgram, which contained non-experienced parts. The final membrane image was obtained with the help of additional processing by a conventional digital filter based on mathematical morphology and linear filtering. The approach for shortcut learning and the internal representations of the network are studied.
© (1991) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.
Akira Hasegawa, Akira Hasegawa, Wei Zhang, Wei Zhang, Kazuyoshi Itoh, Kazuyoshi Itoh, Yoshiki Ichioka, Yoshiki Ichioka, } "Neural-network-based image processing of human corneal endothelial micrograms", Proc. SPIE 1558, Wave Propagation and Scattering in Varied Media II, (11 November 1991); doi: 10.1117/12.49647; https://doi.org/10.1117/12.49647

Back to Top