Neural networks for second-order medical tasks

The ability of a neural network with a sigmoid output-node threshold function to simulate a purely quadratic decision function was studied. The network was applied to the diagnosis of diffuse lung disease. It was found that a minimally configured network can achieve this goal. The convergence of the network is graphically presented, and its performance was normalized to that of the ideal Bayesian decision maker. In a preliminary application, it easily distinguished between normal and pneumonic regions of the lung.<<ETX>>