To accurately identify the site of origin of a tumor is crucial to cancer diagnosis and treatment. With the emergence of DNA microarray technologies, constructing gene expression profiles for different cancer types has already become a promising means for cancer classification. In addition to binary classification, the discrimination of multiple tumor types is also important semi-supervised ellipsoid ARTMAP (ssEAM) is a novel neural network architecture rooted in adaptive resonance theory suitable for classification tasks. ssEAM can achieve fast, stable and finite learning and create hyper-ellipsoidal clusters inducing complex nonlinear decision boundaries. Here, we demonstrate the capability of ssEAM to discriminate multi-class cancer through analyzing two publicly available cancer datasets based on their gene expression profiles.