The diagnosis of craniofacial dysmorphic signs hasa high medical importance. Physicians specialized on human genetics need to get quick reliable information about dysmorphic syndromes characterized by facial abnormalities to give a prognosis about the risk of incidence for following children. In numerous cases of dysmorphic syndromes no chromosomal aberration can be found. Therefore the dysmorphic diagnosis is based on an extensive phenotyp analysis of craniofacial signs. In this study we investigate an artificial neural network approach to identify face images with morphological abnormalities. Different full-connected multi-layer perceptrons (MLP) are trained only with a small set of observations using back-propagation as learning algorithm. The gray level images are sampled, preprocessed and scaled to a size of 55 x 72 pixel. The best results are achieved with a network architecture of 4 hidden and 2 output units. The individual classification error of each investigated network architecture is calculated by the leaving-one-out method. The average classification error was 5 %, the best percentage of correct identifications of one network architecture reached 100 %. The presented results show that classifications generated by aritificial neural networks based on small sets of training examples are able to support a dysmorphic diagnosis.