How to improve a neural network for early detection of hepatic cancer

Abstract

Radiologists perform differential diagnoses of hepatic (liver) masses using ultrasonography (US), computed tomography (CT), magnetic resonance imaging (MRI), and laboratory tests, but interpretation is often difficult. In our earlier research, a backpropagation neural network was designed to diagnose five classifications of hepatic masses: metastatic carcinoma, hepatoma, cavernous hemangioma, abscess, and cirrhosis. After being trained using ultrasonographic data and laboratory tests, the network classified hepatic masses correctly in 51 of 72 cases. That accuracy of 71% 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. What do we need to do to increase that accuracy and make the network friendly enough that radiologists will use it in their diagnoses? We have reviewed the literature, discussed alternatives and developed a plan to improve the diagnostic accuracy of the networks. That plan consists of: (i) get many more patient cases and more data variables, including MRI and CT data, so the network can be more highly trained. A shortage of enough patient cases to properly train the network is the key problem; (ii) use genetic algorithms and other techniques to preprocess the data; (iii) the network should have an optical interface to read images directly; and (iv) build a user-friendly interface using the C programming language on a 486 microcomputer. Continued research along the guidelines in this study should provide a sophisticated neural network for early detection of hepatic cancer that hopefully will exceed the diagnostic abilities of most radiologists.