Purpose: The purpose of this work is to train a backpropagation neural network to make correct diagnosis of thyroid diseases and to compare its performance with human practitioners of medicine.Method: An 84-14-12 neural network was implemented using signs and symptoms of thyroid diseases as input and the twelve kinds of thyroid related illness as output. The training took place first by varying the number of hidden nodes (until the optional 14 was obtained), then by varying the number of hidden layers, the learning coefficient, the momentum coefficient and the tolerance between output and targets. Sensitivity and noise injection were also considered. The training was terminated when all diseases could be diagnosed accurately by the neural network. Test patterns consisting of some symptoms omitted or surreptitiously added were inputted into the "virtual" physician and the result was compared with the diagnosis of human physicians. A sign test was performed to determine whether there were significant mean differences between pair members.Results: Six sets of input were given to four human physicians and to the neural network. None of the human physicians were able to make correct diagnosis for six test patterns. The neural network, on the other hand, was able to make the correct diagnosis in all cases. No significant difference was found between pairs of human physicians or between pairs of human and virtual physician.New Concept of the Work: This study represents the first realistic, field tested investigation of the performance of a "virtual" physician as implemented by a neural network.Conclusion: We have shown that it is possible to train a neural network to make correct diagnosis of diseases on the basis of signs and symptoms. Its usefulness is limited to an environment where no laboratory tests are available and a quick diagnosis is mandatory.