Image classifiers for the cell transformation assay: a progress report

The Cell Transformation Assay (CTA) is one of the promising in vitro methods used to predict human carcinogenicity. The neoplastic phenotype is monitored in suitable cells by the formation of foci and observed by light microscopy after staining. Foci exhibit three types of morphological alterations: Type I, characterized by partially transformed cells, and Types II and III considered to have undergone neoplastic transformation. Foci recognition and scoring have always been carried visually by a trained human expert. In order to automatically classify foci images one needs to implement some image understanding algorithm. Herewith, two such algorithms are described and compared by performance. The supervised classifier (as described in previous articles) relies on principal components analysis embedded in a training feedback loop to process the morphological descriptors extracted by "spectrum enhancement" (SE). The unsupervised classifier architecture is based on the "partitioning around medoids" and is applied to image descriptors taken from histogram moments (HM). Preliminary results suggest the inadequacy of the HMs as image descriptors as compared to those from SE. A justification derived from elementary arguments of real analysis is provided in the Appendix.