Segmentation of Contrast Enhanced Digitized Mammograms

The aim of the knowledge-based contrast enhancement component is to predict the optimal contrast enhancement for a given mammogram. The optimal contrast enhancement is the one maximizing the segmentation of the enhanced image. Segmentation performance is measured by the sensitivity in the detection of true-positive regions within the segmentation image. Section 11.3.2.1.1 identifies the contrast enhancement methods used in this configuration of the adaptive knowledge-based model. Following this, sections 11.3.2.1.2 and 11.3.2.1.3 describe the segmentation method used to evaluate the performance of the contrast-enhanced image and quantitative evaluation of segmentation quality. Finally, section 11.3.2.1.4 identifies the optimal contrast enhancement that exists for each mammogram, providing evidence for the construction of a knowledge-based enhancement component to predict the optimal enhancement method.

11.3.2.1.1 Contrast Enhancement Experts. The utility of six contrast enhancement methods are evaluated: histogram equalization (HISTOEQ), fuzzy enhancement (FUZZY), density weighted contrast enhancement (DWCE), adaptive contrast enhancement (ACE), adaptive contrast enhancement with local entropy (ACELE), and adaptive contrast enhancement with local fractal dimension (ACELFD). Each of these methods will be used as enhancement experts within this configuration of the knowledge-based framework. Full details on the their algorithms is available in [18].

11.3.2.1.2 Segmentation Methods. The aim of image segmentation is to label a pixel in an image as belonging to one of the known corresponding real world objects. In the detection of breast lesions in digitized mammograms, image segmentation results in contiguous areas or regions of pixels, labeled as normal or suspicious. For the purpose of evaluating image enhancement, we use an unsupervised Gaussian mixture model (GMM) and hidden Markov random field (HMRFu) model of image segmentation proposed by Zhang et al. [24]. For ease of referencing, this shall be referred to as HMRFU in the rest of this chapter. The HMRFU segmentation method is used to segment contrast-enhanced images so that the performance of the contrast enhancement can be determined. The HMRFU segmentation algorithm operates in an unsupervised manner. The only a priori knowledge required for the segmentation is the maximum number of classes, L, from which a pixel is labelled. By setting L = 2, HMRFU will label pixels as either normal or suspicious. The HMRFU method models each class using a single Gaussian whose parameters are defined using a maximum likelihood estimate. Followoing convergence, a maximum a posteriori (MAP) segmentation is performed by labeling each pixel with the class maximising the a posteriori probability estimates.

Table 11.2: Outcomes detected following image segmentation

Outcome

Description

Summary

A detected area is defined as True-Positive (TP) if the following two conditions are true:

1. The common area between the Actual A and Target T divided by the area of the target region is greater than or equal to a certain percentage J^n.

2. The total area of the segmented actual region must be less than a constant

Cmax.

In this evaluation Tmn = 50% and Cmax is four times the size of the image target region in the complete dataset.

SUBTP A detected area is defined as SUBTP if the two following conditions are true.

1. The overlap area between the target and the actual regions is less than T ■

2. The actual area is less than or equal to Cmax.

area

11.3.2.1.3 Quantifying Segmentation Performance. To evaluate the quality of the segmentation of a mammogram enhanced using a particular contrast enhancement method, a mechanism for quantifying the segmentation performance is required. Utilizing a simple sensitivity outcome of detection will not identify the true segmentation performance. Instead, a measure of the area of the target ROI correctly labelled as suspicious is proposed. This is achieved using an overlap methodology described by Kallergi et al. [25]. In their study, the authors describe a series of quantitative measures based on the overlap of a suspicious actual region following image segmentation with that of a target ground truth region, denoted by an expert radiologist. Based on the study by Kallergi et al. [25], two quantitative measures are selected and described in further detail in Table 11.2. Figure 11.4 illustrates diagramatically the two outcomes

Figure 11.4: Diagrammatic example of a (a) TP and (b) SUBTP.

to be detected. In each case the target region is shown as darker color and the actual region, following segmentation, is shown as lighter color overlapping. Figure 11.4(a) shows the TP outcome where the target and actual region overlap is greater than ?min = 0.5 and conversely, Fig. 11.4(b) where the overlap of the target region is less than Tlnin = 0.5, the SUBTP outcome.

0 0

Post a comment