الفهرس | Only 14 pages are availabe for public view |
Abstract Case based reasoning (CBR) suggests a model of reasoning that depends on experiences and learning. CBR solves new cases (or examples) by adapting solutions of retrieved similar cases. However, despite of its importance, adaptation process in CBR is a very difficult knowledge intensive task, especially for medical diagnosis. This is due to the complexities of medical domains, which may lead to uncertain diagnosis decisions. Recent studies show that there is no structured adaptation model that has been yet developed for medical diagnosis. revealed that reusing diagnostic solutions is a very difficult unsolved problem. In this thesis two studies on adaptation algorithms have been done. The first is a study on conventional adaptation algorithms as substitutional and transformational rules. It points out the appropriateness and the limitations of conventional algorithms for different CBR tasks as planning, design and diagnosis. from this study, we concluded that the use of conventional adaptation algorithms for case adaptation is the main bottleneck for CBR particularly in planning, design and diagnosis tasks. The second is a study on recent adaptation approaches as neural networks and induction adaptation algorithms. It points out recent directions for overcoming the adaptation burden. It also points out the main advantages and limitations of each recent approach for CBR tasks as planning, design and diagnosis. from this study, we concluded that recent adaptation approaches have shown significant results, when applied in planning and design tasks. However, case adaptation is still the main bottleneck for diagnosis tasks. To test the applicability of different adaptation methods in a real medical domain two CBR-based expert systems prototypes have been developed for thyroid cancer diagnosis; namely CANCER-T and CANCER-C. Their main goal is to test the applicability of different adaptation methods in a real medical domain, which is thyroid cancer diagnosis. In CANCER-T, a case- memory of 820 real thyroid cancer patient cases is built; these cases are obtained from the expert doctors in the National Cancer Institute of Egypt. In the retrieval phase, the nearest-neighbor retrieval algorithm is implemented for case retrieval. While in the adaptation phase, five adaptation methods have been implemented. These are Reinstantiation, Parameter Adjustment, Local Search, Transformational and Hanney Inductive Adaptation approach. These methods have shown very low accuracy rates, which are: 0% for Reinstantiation, Parameter Adjustment and Local Search, this is because no disease can be substituted for another. 1.578% for Transformational, this is because an intractable number of adaptation rules is required in order to handle all feature differences, which is estimated as 244. Also, Hanney Inductive adaptation approach gives an accuracy rate of only 3.157%, this is because an intractable number of adaptation rules is still required. In CANCER-C, the same case-memory of 820 cases is used but each case in the case-memory was decomposed into three sub- cases at three phases of cancer diagnosis; these are the Suspicion, the To-Be-Sure and the Stage phases for diagnosing thyroid cancer suspicion, type and stage. In the retrieval phase, the nearest- neighbor algorithm is implemented for case retrieval. While in the adaptation phase, two recent adaptation methods have been implemented: Hierarchical combined with Transformational, and Hierarchical combined with Hanney Induction adaptation approach. These methods have also shown very low accuracy rates as follows: Hierarchical combined with Transformational gives 6.3 %, 2.89% and 14.4 % at the Suspicion, the To_Be_Sure and the Stage phases receptively. This is because an intractable number of adaptation rules is still required at each phase; these are estimated as 218, 215 and 211 at the Suspicion, the To-Be-Sure and the Stage phases respectively. Also, Hierarchical combined with Hanney Inductive adaptation approach gives 4.2%, 8.7% and 26.25% at the Suspicion phase, the To-Be-Sure phase and the Stage phase respectively. Moreover a new hybrid adaptation model for cancer diagnosis has been developed [90, 91]. It combines transformational and hierarchical adaptation techniques with certainty factors (CF’s) and artificial neural networks (ANN’s). The model consists of a hierarchy of three phases that simulates . |