A hierarchical neural network assembly for classification of cervical cells in automated screening.

Experiments are described using artificial neural networks to classify cells imaged in cervical smears according to their degree of abnormality. This problem of classification was broken into 3 subtasks, each of which an independent back-propagation neural network was trained to solve. Input patterns were fed to the first network for classification of the cells as essentially normal or abnormal and then, depending on the outcome of the classification, a second stage was invoked for classifying the cell as (i) normal or mildly dysplastic, or (ii) moderately or severely dysplastic. It is shown that the correct choice of normalization of input data, as well as the use of a hierarchy of neural networks, each optimised for a specific subtask of the whole classification process, yields a predictive value hitherto unattained by automated systems.