Contrast Enhancement Mixture of Experts Framework

In CAD of breast lesions, one aim of contrast enhancement is to improve the performance in image segmentation. Therefore, the optimal contrast enhancement method for a mammogram is the one maximizing the sensitivity of the detection of a breast lesion following image segmentation. The proposed knowledge-based enhancement component will predict the optimal contrast enhancement method, or expert, for a test mammogram using knowledge learnt from a set of training mammograms. Figure 11.3 summarizes the enhancement component framework. Individual enhancement experts (1,..., n) are used in training and testing, shown on the left and right of Fig. 11.3, respectively. Experts are grouped together in training and testing for particular mammogram types. During training, an optimal enhancement expert is identified (say for example expert 2) and the mapping between a global characteristic of the training mammogram and the enhancement expert is captured as component knowledge. For the testing mammogram, the optimal expert can be predicted (which was expert 2) based on an image feature vector. Where possible, the a priori knowledge of the mammogram breast type will be used. Different parameterized versions of the knowledge-based contrast enhancement component are constructed for each breast type grouping.

TRAINING TESTING

TRAINING TESTING

Breast grouping knowledge

Figure 11.3:

Breast grouping knowledge

Figure 11.3:

This section describes the mixture of experts framework and it is laid out as follows. Section 11.3.2.1 reviews the contrast enhancement experts used to build the framework. Then the segmentation algorithm used to evaluate the enhanced images is briefly described together with quantitative measures of segmentation performance. In section 11.3.2.2 results are presented when applying the different image enhancement on DDSM images and the resulting segmentation from them. Section 11.3.2.3 discusses the features that can be extracted from the mammograms to be fed into a mapping scheme (e.g., neural networks) that maps features to optimal enhancement methods. Finally, section 11.3.2.4 discusses a machine learning system for this mapping. A neural network is used in two different modes: double network mapping and a single direct mapping scheme.

0 0

Post a comment