High Level Overview of Adaptive Knowledge Based Model

Figure 11.1 shows a diagrammatic high-level overview of the proposed adaptive knowledge-based model. The major components of a CAD pyramid are shown. They are contrast enhancement, image segmentation for the identification of suspicious a ROI identification, and false-positive reduction. The knowledge-based framework underpinning the adaptive knowledge-based model is used in the identification of an optimal pipeline for each mammogram. Additional knowledge is incorporated into the model by implementing separate parameterized versions of image segmentation and false-positive reduction components according to a mammogram grouping strategy. Each knowledge-based component presented in Fig. 11.1 is discussed in further detail below.

Mammogram grouping: By grouping mammograms on a predefined criteria, subsequent CAD components may be engineered to operate on specific mammograms types. Our study hypothesizes that a mammogram can be grouped on the basis of its parenchymal patterns. The aim of this component is to predict a mammogram's group by utilizing supervised learning techniques in conjunction with a training set of example images.

Optimal contrast enhancement: A range of contrast enhancement techniques previously used in mammographic CAD research are surveyed in [18]. The adaptive knowledge-based model aims to accommodate many of these methods in the form of enhancement experts and learn, on the basis of feature vectors from training mammograms, the optimal contrast enhancement expert for a given mammogram on test. We propose machine learning techniques such as artificial neural networks (ANN) for learning this mapping.

Optimal image segmentation: A variety of different image segmentation methods have been identifiedfor mammographic CAD in [18,19]. Adopting a similar strategy to that of the knowledge-based contrast enhancement experts, a set of segmentation experts are proposed. As opposed to different contrast enhancement experts, each segmentation expert is functionally identical. The adaptability property in the segmentation component is achieved by learning the saliency of input features used to perform the segmentation. This chapter hypothesizes that different segmentation experts operating on different input feature spaces will have a greater utility in the segmentation of different mammograms. Input features for expert construction will be drawn from a subset commonly utilized in mammographic CAD. For example, the subset may include image gray scales, contrast enhanced gray scales, textures, and edge-gradient information possibly at different scales of resolution. Each segmentation expert operates on apredefined set of features for apredefined group of mammograms. The implementation of an optimal segmentation is achieved by predicting the best blend of segmentation decisions given by the collection of experts for an individual mammogram.

Reduction of false-positive regions: The final component is the reduction of false-positive regions. This component operates by discriminating between normal and abnormal regions based on a feature vector extracted from a suspicious ROI. This component is implemented within the adaptive knowledge-based model as a modular arrangement of ANNs trained to specialize in particular groupings of mammograms.

0 0

Post a comment