Fast and accurate, non-linear autoassociators perform well in the face of unbalanced data sets, where few to no positive examples are present. In cancer diagnosis, for example, this can be convenient if only benign data is available, or if only a very small proportion of malignant data is available. As proof of concept, we apply a non-linear autoassociator to breast tumor data to predict the presence of cancer using only benign examples to train the autoassociator. Our results indicate that the non-linear autoassociator approach to automated breast cancer diagnosis is convenient and yields accurate results with minimal overhead.