The development and performance evaluation of a computer expert system for the histopathologic diagnosis of salivary gland neoplasms
Abstract
The design, development, implementation and testing of a prototype, interactive histopathologic expert system (called "SAGA") capable of diagnosing 15 types of primary epithelial neoplasms of salivary glands is described in this study. SAGA incorporates a multiple subprogram modular design architecture and makes use of multiple reasoning methodologies, including: data-driven and goal directed rule-based reasoning, linear pattern recognition, and Bayesian classification. The system's user interface incorporates both a "hypertext" context-sensitive information assistance facility and the video display of scanned photomicrographic histopathologic images. SAGA can report a differential diagnosis of its findings with an associated assessment of its confidence in its diagnosis. The system's diagnostic performance was evaluated in a series of tests. The results of a weighted Kappa analysis of SAGA's diagnoses versus those of four human expert oral pathologists for a set of 20 salivary gland neoplasm test cases indicate no statistical difference in diagnostic performance between the system and the human experts, and each of the experts in relation to the others in the group, as demonstrated by the use of Wilcoxon rank sums test. The results of a modified version of Turing's test of artificial intelligence demonstrated no statistically significant difference in the number of cases a judge disagreed with SAGA's diagnoses versus the number of cases they disagreed with the diagnosis of four human expert pathologists for a set of twenty salivary gland neoplasm test cases, as indicated by use of Fisher's exact test on the data obtained from three experimental trials. Furthermore, it was demonstrated that an experimental group (diagnoses aided by SAGA) of novice subjects' diagnostic performance and capabilities can be significantly augmented over those of a control group (diagnoses aided by textbooks/atlases) through the use of the histopathologic expert system. An over 300% increase in the number of correct answers obtained from subjects of the experimental group (n = 26) over those of the control group (n = 26) in the diagnosis of a set of 8 salivary gland neoplasm test cases was demonstrated to be statistically significant through the use of Student's t-test and Chi-square test.Description
University of Maryland, Baltimore. Pathology. Ph.D. 1994Keyword
Health Sciences, DentistryHealth Sciences, Pathology
Artificial Intelligence
Computer Science
Diagnosis, Computer-Assisted--instrumentation
Salivary Gland Neoplasms--diagnosis
Salivary Gland Neoplasms--pathology