A Neural Architecture for Potentially Classifying Cytology Specimens by Machines
Robert L. Harvey, P.N. DiCaprio, K.G. Heinemann, M L Silverman, J M Dugan · PubMed Central · 1990
This paper describes a general purpose vision system. We have applied the system to classifying cytology specimens. The system uses neural network and conventional processing modules to model biological vision systems. The modules make up a locating channel and a classifying channel. The locating channel finds individual cells in the field-of-view. The classifying channel learns and recognizes the cells. Learning is by example. We tested the classifying channel on 156 cell images from human cervical smears. Results suggest one can drive the false negative and false positive rates below a few percent for initial screening. Training would require several hundred cells of normal and abnormal types.