Knowledge Representation with Multistage Neural Networks

An alternative to a hard bounded grouping scheme such as those proposed above is in the use of soft decision boundaries. These knowledge-based components utilize a mixture-of-experts paradigm. Lai and Fang [11, 12] proposed the use of a hierarchical neural network to model the optical transformation of a 12-bit magnetic resonance image (MRI) into an 8-bit representation for display on a computer monitor. This optimal optical transformation is crucial if the expert radiologist is to effectively interpret the displayed image. The authors cite a major obstacle to the implementation of a simple linear solution as the difference in optimal parameters for different types of MR images. For example, 3D angiographic images and T1- or T2-weighted images should be parameterized differently. To account for such differences, the authors propose a hierarchical arrangement of neural networks trained to provide accurate estimation for a wide range of images. To achieve the mapping of optimal transformation parameter values with individual images, a variety of histogram, wavelet and spatial features are used to characterize each image. The decision of each network module in the hierarchy is combined using a weighted averaged fusion scheme. Evaluating their framework on a dataset of more than 2400 images the authors report that their methodology gives satisfactory results and it is robust to unknown images.

Similarly, Pitiot et al. [13] proposed an automated method for extracting anatomical structures in MRI based on textural classification. The authors hypothesize that performing a pretextural classification prior to segmentation will lead to a more accurate definition of the anatomical boundary. The pretextural classification is based on a mixture-of-experts paradigm, such that each expert is trained on a particular grouping of textural features extracted from a moving window within the image. A second-stage multiscale neural network is trained on equally drawn numbers of random samples from correctly and incorrectly classified pixels from the first stage. The network arrangement of stage two is trained on local morphology and texture features from a wider pixel neighborhood in the task of detecting anatomical structures. Evaluating their framework on a small dataset of 10 testing images, the authors report an increase in classification rates as a result of the two-stage hybrid neural classification.

Sha and Sutton [14] proposed the use of a network-of-networks paradigm first discussed by Guan et al. [17] for dynamically reconfiguring a test image for enhancement and segmentation. Under the proposed framework, the image is connected on a pixel-by-pixel basis by weights in a manner analogous to an attractor neural network. Pixels are updated based on local variances obtained from weight connections with neighbours in an iterative manner until convergence of the network architecture. In their study, the authors present only qualitative results on the enhancement and segmentation of MRI images.

0 0

Post a comment