The application of image analysis and neural network technology to the study of large-cell liver-cell dysplasia and hepatocellular carcinoma.

Abstract

Liver cell dysplasia (LCD) is considered a preneoplastic lesion, whose characterization and differentiation from hepatocellular carcinoma (HCC) and from the reactive changes seen in cirrhosis has been controversial. We studied 12 cases of LCD (large cell type) with image analysis techniques (IA) and compared the findings with those of HCC (n = 40), and a spectrum of non-neoplastic hepatic lesions including normal liver and cirrhosis (n = 49). A minimum of 200 Feulgen-stained nuclei were measured from each lesion with the CAS 200 image analysis system. The data were collected with the aid of CellSheet software. Thirty-four variables were measured, including geometric, textural, and photometric nuclear features and DNA ploidy. The data were analyzed with multivariate statistics and a backpropagation neural network (NN). Stepwise statistical analysis selected 22 variables that were statistically significant in the three groups with P values <.05. Various NN architectures were developed using these variables. The best NN architecture included a sigmoidal transfer function, 14 input, 16 hidden, and 3 output neurons. It trained to completion after 8,887 runs using 90% of the lesions. This NN yielded a 100% cross-validation rate for unknown cases. These data support the concept of LCD (large cell type) as a lesion that can be objectively distinguished from HCC and non-neoplastic liver. Our study also demonstrates the potential usefulness of IA for the evaluation of difficult histopathological problems.