Pergamon Expert Systems With Applications, Vol. 11, No. 3, pp. 351-360, 1996 Copyright © 1996 Elsevier Science Ltd Printed in Great Britain. All rights reserved 0957--4174/96 $15.00+0.00 PII: S0957-4174(96)00050-4 Self-Integrating Knowledge-Based Brain Tumor Diagnostic System CHING-HUNG WANG'~ Institute of Computerand Information Science,NationalChiao-Tung University, Hsin-Chu30050,Taiwan,R.O.C. TZUNG-PEI HONG Departmentof Information Management,KaohsiungPolytechnic Institute,Kaohsiung84008,Taiwan,R.O.C. SHIAN-SHYONG TSENG Institute of Computerand Information Science, NationalChiao-Tung University, Hsin-Chu30050,Taiwan,R.O.C. Abstract--In this paper, we present a self-integrating knowledge-based expert system for brain tumor diagnosis. The system we propose comprises knowledge building, knowledge inference and knowledge refinement. During knowlege building, an automatic knowledge-integration process, based on Darwin's theory of natural selection, integrates knowledge derived from knowledge-acquisition tools and machine- learning methods to construct an initial knowledge base, thus eliminating a major bottleneck in developing a brain tumor diagnostic system. During the knowledge inference process, an inference engine exploits rules in the knowledge base to help diagnosticians determine brain tumor etiologies according to computer tomography pictures. And, a simple knowledge refinement method is proposed to modify the existing knowledge base during inference, which dramatically improves the accuracy of the derived rules. The performance of the brain tumor diagnostic system has been evaluated on actual brain tumor cases. Copyright © 1996 Elsevier Science Ltd 1. INTRODUCTION RECENTLY, EXPERT SYSTEMS have been successfully applied to many fields and have shown excellent performance. Expert systems provide sound expertise in the form of diagnosis, instruction, prediction, consulta- tion and so on. They can also be used as training tools to help new personnel interpret data and monitor observa- tions (Waterman, 1986). Developing a successful expert system requires, however, effectively integrating knowl- edge from a variety of sources, such as that from domain experts, historical documentary evidence, or current records, to construct a complete, consistent and unambi- guous knowledge base (Baral, 1991; Gragun, 1987). For large-scale expert systems that generally cannot rely on a single knowledge source, the use of multiple knowledge "~Author for correspondence. Also Directorate General of Tele- communication Laboratories, Ministry of Transportation and Communications, Chung-Li, Taiwan 32617, Taiwan, R.O.C. inputs from many knowledge sources is especially important to ensure comprehensive coverage. Thus, integrating multiple knowledge sources plays a critical role in building successful expert systems. In this paper, we present a brain tumor diagnostic system that can integrate multiple knowledge sources to quickly build a prototype knowledge base. This prototype knowledge base then adapts itself according to inference results from the expert system, consequently improving the accuracy of the rules it derives. The brain tumor diagnostic system (BTDS) consists of three main functional units: knowledge building, knowl- edge inference and knowledge refinement (Wang & Tseng, 1995). The knowledge-building unit includes three modules: machine learning, knowledge acquisition and knowledge integration. The machine-learning mod- ule maintains a variety of machine-learning strategies (Cendowska, 1987; Michalski, 1980; Mitchell, 1982; Quinlan, 1986) to induce knowledge from actual instances. The knowledge-acquisition module maintains 351