Classification Based on Human Perceptual Features

One way to obtain features is to ask radiologists to read a set of mammograms and to describe the appearance of the lesions [53-55,60-63]. Radiologists are specialists who are trained to recognize the normal or abnormal appearance of radiographs. Therefore, it is only reasonable to classify breast lesions based on their perceptual descriptions. The advantage of this approach is that it makes direct use of radiologists' expertise. However, a standard language must be developed for a computer technique to interpret radiologists' description. Radiologists may often use somewhat different terminology when describing breast lesions. Occasionally they may also use the same words to convey slightly different meanings. This ambiguity in terminology is generally not a problem for a radiologist, since he or she can make good use of their own description in making the malignant versus benign diagnosis. However, it is problematic for computer classification because it is not possible for an automated technique to interpret a particular radiologist's language unless a standard set of terms are used.

Getty et al. have used this approach to classify breast lesions based on radiologist-described features [53]. They developed two diagnostic aids to help radiologists interpret breast lesions. The first aid was a checklist containing a set of features that a radiologist must assess and record. The second aid was an estimated probability of malignancy provided by a computer classifier that analyzed the features reported by radiologists. The process of developing the checklist of features was remarkably involved. It started with interviews with five mammography specialists. These interviews generated several dozen features. Then, this extensive set of features was reduced to 29 through mathematical multidimensional scaling and hierarchical clustering, and through a series of consensus-seeking group discussions. Finally, the 29 features were reduced to 13 by means of linear discriminant analysis. The final set of 13 features was believed to be a necessary and important set of features for classifying malignant and benign breast lesions. The final set included six features describing masses, three features describing clustered microcalcifications, and two features describing architectural distortion and skin thickening. The complete set of features is included as an appendix in their seminal paper in 1988 [53] and is also listed in Table 1. This process illustrates the difficulty in identifying important features and in developing a common language for the features. Remarkably, however, once such a feature set is developed, it may be a useful aid to radiologists even without computer classification — hence, the first of their two diagnostic aids for interpretation of mammograms. That diagnostic aid helps radiologists by guiding them to interpret breast lesions in a methodical way through consideration of each of the important features. We will come back to this work later in the chapter.

FIGURE 2 Photographically enlarged regions of mammograms depicting microcalcifications.

FIGURE 2 Photographically enlarged regions of mammograms depicting microcalcifications.

In an effort to standardize mammogram interpretation in a much larger scope, the American College of Radiology (ACR) has developed the Breast Imaging Reporting and Data System (BI-RADs) [29]. The purpose of the BI-RADS is to serve as a quality assurance tool to standardize mammographic reporting, reduce confusion in breast imaging interpretations,

(a) a malignant cluster of microcalcifications and (b) a benign cluster of

(a) a malignant cluster of microcalcifications and (b) a benign cluster of and facilitate outcome monitoring. BI-RADS contains a breast imaging lexicon that is a list of breast-lesion features and standardized terms for describing these features. It is meant to be adopted by radiologists so that there can be uniformity in the interpretation and description of breast lesions. The BIRADS lexicon can be a useful source for computer analysis as

FIGURE 3 Schematic diagram of a typical computer technique for the classification of malignant and benign breast lesions.

well. Baker et al. have developed a classification technique based on radiologist-reported BI-RADS features [55]. They used a total of 18 features. Ten of them describe breast lesions — calcifications, mass, and any associated findings — using only BI-RADS descriptive terms. The rest of the features are descriptions of patient history. The BI-RADS features are attractive because they eliminate the need for other investigators to duplicate the work of identifying important and comprehensive features and because, at least in principle, it could be readily adopted by others once BI-RADS becomes widely used by radiologists.

However, reporting features is a time-consuming process for radiologists. In addition, the results of the computer classification are prone to be influenced by variability in radiologists' interpretation of mammograms [68]. Requiring radiologists to interpret and describe 10 or 13 features may reduce their productivity in reading mammograms and diminish their enthusiasm for performing this task. The use of BI-RADS features can potentially alleviate this problem. The Mammography Quality Standard Act (MQSA) now requires radiologists to use BI-RADS assessment categories in reporting mammography findings. Although not specifically required by MQSA, BI-RADS recommends that descriptions of features of breast lesions be included in mammography reports. If this is

TABLE 1 Examples of human perceptual features and computer-extracted features that can be used to classify malignant and benign breast lesionsa

Feature type or number

Feature description

Reference

Human perceptual features 1 2

0 0

Post a comment