The automated detection of lesions in the breast is important. The area of computer-aided detection (CAD) in digital mammography is devoted to developing sophisticated image analysis tools that can automatically detect breast lesions. The whole process can be viewed as a pipeline of subprocesses that are aimed at finding regions of interest (ROI) and classifying them in breast images. These processes (layers) are common to most medical imaging applications and involve image preprocessing, enhancement, segmentation, feature extraction, classification, and postprocessing for reducing false positives. There is a variety of algorithms for these processes available in medical imaging literature but little to guide their selection. There are only a few comparative studies that exhaustively compare different algorithms on large datasets and correlate the success of the algorithm with the type of data used. Most clinical studies use a preselected set of image analysis algorithms that are uniformly applied to all images. In our opinion, this practice is not good. In this chapter we demonstrate the use of a knowledge-based framework that integrates the various layers of analysis under an adaptive scheme. The main emphasis is to have at our disposal more than one algorithm per layer to produce the same type of output, and then

1 Pann Research, Department of Computer Science, University of Exeter, Exeter EX4 4QF,

2 Met Office, Fitzroy Road, Exeter EX1 3PB, UK

based on the properties of the image under consideration, predict the single best algorithm to be applied at each layer from this set. We demonstrate that this scheme of work has significant advantages over a nonadaptive structure (where only one algorithm is available per layer and it is fixed for all images in the dataset).

We aim to answer the following questions: (a) What is a knowledge-based framework? We discuss the components of this framework in section 11.2 putting it in the context of previous research. (b) How does the image enhancement layer work in this framework? This is detailed in section 11.3 where we discuss measures of image viewability based on enhancement, and demonstrate the role of good enhancement in image segmentation. We also propose two new mapping schemes that can map the image features to chosen enhancement methods. (c) How does the image segmentation layer work within the knowledge-based framework? In section 11.4 we detail the implementation of sophisticated Gaussian mixture models in both supervised and unsupervised modes, with an expert combination framework and compare them on overlap measures. (d) What are the different strategies for reducing false positives? In section 11.5 we discuss several postprocessing steps that are aimed at reducing the number of false positives per image. (e) Is the adaptive knowledge-based framework superior to a nonadaptive scheme that uses the same algorithms across all images uniformly? We discuss our results on this issue in section 11.6 where we show the relative superiority of the adaptive framework.

0 0

Post a comment