Making the Distinction Between Crohn's Disease and Ulcerative Colitis by Histopathological Examination: A Comparison of Human Performance, Logistic Regression and Adaptive Resonance Theory Mapping Neural Networks (ARTMAP)

280 cases of inflammatory bowel disease were examined histopathologically and defined features were observed and recorded using a graphical user interface with digitised reference images. The outcome of each case was determined independent of the histopathological report giving 75 cases of Crohn's disease and 105 cases of ulcerative colitis. The cases were randomised and split into training and test sets, each of 140 cases. All 23 observed features were used as input data for logistic regression and adaptive resonance theory mapping neural networks (ARTMAP). The ARTMAP's were used singularly or as a voting cohort of 11 networks, the majority and unanimous decisions of the cohort were analysed separately. The vigilance parameter for the ARTMAP's was varied from high (0.9) to lower (0.5) to vary the number of clusters in ARTa. The best resukts were produced by logistic regression and the 11 high vigilance ARTMAP's with a sensitivity for Crohn's disease of 60% and a specificity of 80%. This was a significant improvement on the original human diagnoses that gave a sensitivity of 25% and specificity of 60%. Either of these technologies could form the basis of a decision support system in this domain (193)