Incorporating Human Knowledge in Automated Celiac Disease Diagnosis

Recently, computer-aided celiac disease diagnosis has been promoted to provide an objective opinion besides histological examination of biopsies and visual assessment of macroscopic mucosal tissue. State-of-the-art techniques, however, are not accurate enough to provide incentive for clinical deployment. In this work, we answer two questions: Do computers and human experts make similar classification errors and can expert knowledge be utilized to increase the accuracy of computer-aided methods. Three experts were asked to perform visual classification of a large number of images. The experts decisions were combined with nine different state-of-the-art image representations. Experimentation showed that the correlations between two computer-based methods were higher than the correlations between an expert and a computer-based method. Furthermore, the inclusion of expert knowledge led to statistically significant (p < 0.05) improvements in 69 out of 108 investigated settings.