A computational intelligence and machine learning-based framework for improving case-based computer-aided medical decision systems with application to mammography.

摘要

In case-based computer-aided decision systems (CB-CAD), previously acquired examples associated with different disease states are used to classify new cases. The case-based paradigm is becoming increasingly popular in computer-aided diagnosis. The primary reason is that large digital medical databases are now available, mainly of medical images. This incurs a need for efficient storage and use of the cases for medical decision making. Case-based systems are up to the task since new examples can be simply added to the system without the need for retraining. Further, the operating principles of case-based systems resemble the decision making process of physicians. Thus, case-based CAD systems are accepted more easily in the medical environment. Such systems, however; face certain limitations. Since all examples are stored in the system's database, the system induces large storage requirements. Furthermore, individual comparisons of a query case to all case-base examples increase the system's response time per query.;In this dissertation, a comprehensive computational intelligence and machine learning-based framework is proposed for optimization of case-based medical decision systems. It applies to two crucial components of case-based systems: decision algorithm and case base. The study is performed in the context of a computer-aided decision system for detecting breast cancer in screening mammograms that has been previously published. Although the study is based on the specific CAD system, efforts were made to ensure that the proposed techniques will be easily applicable to other case-based CAD systems.;In the first stage of the research, an improvement of the decision function is proposed for the CAD system. The study hypothesis at this stage is that differentiating the importance of each case in the knowledge database may improve the system's performance. A problem of finding an optimal vector of importance weights is formalized as an optimization problem and a genetic algorithm is applied to solve it. The initial experimental results show that the proposed technique results in a statistically significant improvement of the classification performance.;In the second stage of the research some computational intelligence and machine learning techniques were used to optimize the case base of the CAD system by removing superfluous/detrimental examples. This type of optimization is of great significance since it can reduce storage requirements of the system, decrease response time of the system, and possibly improve system classification performance by removing misleading patterns. The results show that using computational intelligence and machine learning techniques allows for the database of examples to be significantly reduced while increasing performance of the system.;The third part of the dissertation research is devoted to building ensemble classifiers for improving the classification performance and reducing storage requirements of case-based systems. Two methods are proposed that automatically adapt the ensemble size to the problem. The new methods are compared to more traditional approaches. Experimental results show that the ensemble techniques provide a significant improvement, in the classification performance of the CAD system while at the same time reducing the total number of examples used for classification.;The last part of this dissertation is devoted to comparison of all the proposed techniques, an extension of the ensemble approach and application of the ensemble approach to evaluate case-specific reliability of classifier decisions. Some insight into combining the proposed techniques to further improve performance is also offered.