In a retrospective case-control study on 46 metastasized and 34 non-metastasized primary tongue carcinomas, the nuclear morphology and chromatin pattern were assessed in 3 microns thick, formalin-fixed, paraffin-embedded, and Feulgen-stained tissue sections of surgical resection specimens, by means of high-resolution computer-assisted image analysis. The aim of this study was to disclose differences in karyometric features, such as nuclear size-, shape-, and chromatin-pattern features, between these groups, with a view to developing a discriminant function that can predict the occurrence of metastasis for the individual patient. In addition, the lymph node metastases of 31 patients and the normal tongue epithelium of 21 patients were also assessed, to study the possible differences between these two groups and primary tumours. In the metastasized tumours, the chromatin was significantly more condensed (P = 0.01) and exhibited significantly less variation in chromatin condensation (P < 0.001) than in the group of non-metastasized carcinomas. Comparison of lymph node metastases with their primary tumours disclosed only minor differences in chromatin pattern. These findings suggest that only minor genetic differences exist between primary tongue carcinomas and their metastases. Tumour cells of tongue carcinomas showed highly significant differences from cells of normal tongue mucosa for most karyometric features. Logistic regression analysis resulted in a classifier, based on the circularity of the nucleus (CIRC) and the standard deviation of the chromatin condensation (SD COND), to predict the occurrence of lymph node metastases. After cross-validation, the percentages of correct classifications in the group of metastasized and non-metastasized tumours were 72 and 62 per cent, respectively. These results are comparable to the classification results obtained from a classifier based on the clinical T-stage, but our karyometric classification results show a much more equal distribution between the sensitivity and specificity. Karyometric features appeared to be more appropriate to predict metastases than biomarkers such as p53, bcl-2, and Ki-67.