In the past decade, computer-assisted diagnosis systems involving ANNs and BBNs have become increasingly popular in the fields of medical image processing or other medical applications. A large number of application systems has been reported, but most of them have not been successful in practice, especially in the clinical environment [3]. Although it has been well known that the bias in the training database or validation method and the noise in selected features have great impact on the ultimate performance of an ANN and a BBN in clinical environment, many problems related to the optimization of database and feature selection remain unsolved. In database selection, there is no agreement upon (1) how to assess adequacy of a database, (2) how to compare the difficulty between different databases, and (3) how to establish a common database in the field. In feature optimization, because of a large number of features that may be initially extracted by some researchers, current methods are not always efficient for exploring the feature spaces. The problem becomes more serious when using an ANN than using a BNN, because the weights in the ANN need to be trained for each selection of feature set, while the weights in the BBN only need to be computed once. For example, although GA is considered as one of the most efficient computing algorithms available to date, it cannot practically fulfill a task to optimize feature selection for an ANN in a large feature space. In one study where the initial feature space contained 587 features, the researchers could only use GA to perform feature selection on a subset of the entire feature set, because of the limitation on power and speed of the computer [23]. Thus, searching for more effective and efficient methods in optimization of database and feature selection will still be one of the major research trends to improve the performance and generalization of ANNs and BBNs used in medical imaging processing. Researchers are continuously exploring different methods to establish standards for building common databases and to set up standard procedures for measuring and reporting the adequacy and difficulty of databases [21]. Meanwhile, improving computation efficiency will also be an important research topic for feature optimization in a large feature space. One possibility is to take advantage of multiple platforms and develop distributed or parallel GAs to accelerate computation speed. Other studies focus on the explorations of multilevel evolution and self-adapting mechanisms with GAs to control parameters of internal representation, mutation, crossover, and population size [9].

Since a significant research effort continues to refine present methods and introduce new techniques for database optimization and feature selection, the obstacles to the practical implementation of computer-assisted diagnosis systems in medical images will probably be overcome in time. As this occurs, patients will experience the benefit of a reliable and cost-effective diagnosis from a physician aided by a computerassisted diagnosis system, since evidence supports that independent double-reading can significantly improve the accuracy of medical diagnosis [25].

0 0

Post a comment