METHODS IN LABORATORY INVESTIGATION - IMAGE-ANALYSIS AND DIAGNOSTIC CLASSIFICATION OF HEPATOCELLULAR-CARCINOMA USING NEURAL NETWORKS AND MULTIVARIATE DISCRIMINANT FUNCTIONS

BACKGROUND: Hepatocellular carcinoma (HCC) is often difficult to diagn ose in cytologic material and small tissue biopsies since histomorphol ogic information is minimal or absent. The potential for misdiagnosis is greatest in attempting to discriminate well-differentiated HCC from dysplastic hepatocytes in cirrhosis. We investigated the feasibility of developing artificial intelligence classification methods based on nuclear image analysis data for use as adjuncts to the morphologic dia gnosis of HCC. EXPERIMENTAL DESIGN: Ninety hematoxylin-eosin stained h istologic slides including 56 with well- to poorly differentiated HCC and 34 showing a morphologic continuum from normal to markedly dysplas tic benign hepatocytes were assembled from four laboratories. A relati vely inexpensive PC-based image analysis system was used to measure 35 nuclear morphometric and densitometric parameters of 100 nuclei in ea ch specimen. The data were randomized into classification training and testing sets containing equal numbers of benign and HCC samples. Obje ctive diagnostic classification criteria for HCC based on neural netwo rks and multivariate discriminant functions (DFs) were developed for t he most discriminatory subsets of morphometric, densitometric, and com bined morphometric/densitometric variables as selected by stepwise dis criminant analysis of training data. RESULTS: Morphometric parameters provided the best results with the following testing data positive and negative predictive values (PV+ and PV-) for HCC classification: 86.2 % PV+ and 81.3% PV- for a linear DF, 85.7% PV+ and 76.5% PV- for a qua dratic DF and 100% PV+ and 85.0% PV- for a neural network. CONCLUSIONS : Our results demonstrate that nuclear image analysis-based objective classification criteria for HCC can be developed using artificial inte lligence methods and that histologic material prepared at different in stitutions can be reliably classified. Neural networks for HCC classif ication were superior to linear and quadratic DFs. Morphometric data y ielded the best results compared with densitometric or combined morpho metric/densitometric data.