Using an artificial neural network to diagnose hepatic masses

Abstract

Using abdominal ultrasonographic data and laboratory tests, radiologists often find differential diagnoses of hepatic masses difficult. A computerized second opinion would be especially helpful for clinicians in diagnosing liver cancer because of the difficulty of such diagnoses. A back-propagation neural network was designed to diagnose five classifications of hepatic masses: hepatoma, metastatic carcinoma, abscess, cavernous hemangioma, and cirrhosis. The network input consisted of 35 numbers per patient case that represented ultrasonographic data and laboratory tests. The network architecture had 35 elements in the input layer, two hidden layers of 35 elements each, and 5 elements in the output layer. After being trained to a learning tolerance of 1%, the network classified hepatic masses correctly in 48 of 64 cases. An accuracy of 75% is higher than the 50% scored by the average radiology resident in training but lower than the 90% scored by the typical board-certified radiologist. When sufficiently sophisticated, a neural network may significantly improve the analysis of hepatic-mass radiographs.