The concentration of Fos, a protein encoded by the immediate-early gene c-fos, provides a measure of synaptic activity that may not parallel the electrical activity of neurons. Such a measure is important for the difficult problem of identifying dynamic properties of neuronal circuitries activated by a variety of stimuli and behaviours. We employ two-stage statistical pattern recognition to identify cellular nuclei that express Fos in two-dimensional sections of rat forebrain after administration of antipsychotic drugs. In stage one, we distinguish dark-stained candidate nuclei from image background by a thresholding algorithm and record size and shape measurements of these objects. In stage two, we compare performance of linear and quadratic discriminants, nearest-neighbour and artificial neural network classifiers that employ functions of these measurements to label candidate objects as either Fos nuclei, two touching Fos nuclei or irrelevant background material. New images of neighbouring brain tissue serve as test sets to assess generalizability of the best derived classification rule, as determined by lowest cross-validation misclassification rate. Three experts, two internal and one external, compare manual and automated results for accuracy assessment. Analyses of a subset of images on two separate occasions provide quantitative measures of inter- and intra-expert consistency. We conclude that our automated procedure yields results that compare favourably with those of the experts and thus has potential to remove much of the tedium, subjectivity and irreproducibility of current Fos identification methods in digital microscopy.