A Survey on Automatic Breast Density Classification

Two main trends have been followed in breast density classification: mammographic density quantification and mammographic tissue classification. Mammographic density quantification is related to obtaining a single value in order to quantify the breast dense tissue. This value can be obtained using the common 2D views [70,154,160,165,168,173] or otherwise using 3D information, in which case, it is called a volumetric measure [15,74,183]. However, in medical practice, such quantitative analysis seems unnecessary. In fact, radiologists mainly estimate the breast density by visual judgment of the mammographic imaged tissue. According to this judgment, and using a determined classification, the breast is assigned to their corresponding class. Thus, automatic tissue classification methods try to imitate such visual judgment, learning from the radiologists experience.

This qualitative description of the breast density introduces large intraobserver and interobserver variations in the estimated classification, thus obtaining in general different qualitative descriptions. Although there exist a number of different lexicon/scales for breast tissue classification [123], nowadays, the commonly used is the BIRADS lexicon [2].

In the literature, different approaches based on the use of only histogram information have been proposed for classifying breast tissue [83,196]. However, in our experience and using public databases, it is clear that histogram information alone is not sufficient to classify mammograms according to BIRADS categories [128,202]. To illustrate this, the third row of Figure  shows the respective histograms of four different mammograms, each belonging to a different BIRADS class. Note that although the mammograms belong to different classes, the four histograms are quite similar both in the mean grey-level value and the shape of the histogram.

Figure 3.1: From top to bottom the four columns show the original mammogram, the segmented breast area, and the associated histogram, respectively. From left to right this shows four similar histograms, each of increasing BIRADS category: from BIRADS I (first column) to BIRADS IV (last column).

Thus, several researchers have focused their attention on the use of texture features to describe breast density. Miller and Astley [121] investigated texture-based discrimination between fatty and dense breast types applying granulometric techniques and Laws texture masks. Byng et al. [22] used measures based on fractal dimension. Bovis and Singh [17] estimated features from the construction of spatial grey level dependency matrices. Recently, Petroudi et al. [137] used textons to capture the mammographic appearance within the breast area. Zwiggelaar et al. [199,200] segmented mammograms into density regions based on a set of co-occurrence matrices and the subsequent density classification used the relative area of the density regions as the feature space.

Table 3.1: Table summary of the reviewed work on breast density classification. The upper block shows works which quantify the density of the breast. The works of bottom block classify the breasts according to the lexicon shown.

	Author	Year	Segmenting Features
	Taylor [173]	$1994$	Fractal, grey-level
		Suckling [168]	$1995$	Grey-level
		Heine [70]	$2000$	Grey-level
		Sivaramakrishna [165]	$2001$	Grey-level
		Saha [154]	$2001$	Grey-level
		Blot [15]	$2005$	Volumetric
		Highnam [74]	$2006$	Volumetric
		Selvan [160]	$2006$	Grey-level
		Van Engeland [183]	$2006$	Volumetric
	Author	Year	Extracted Features	Classifier	Lexicon
	Magnin [109]	$1986$	Co-occurrence	N$/$ A	Wolfe ($4$ )
		Caldwell [23]	$1990$	Fractal Analysis	Bayesian
		Tahoces [171]	$1995$	Grey-level, Fourier	LDA
		Boyd [18]	$1995$	Histogram	Bayesian
		Byng [22]	$1996$	Fractal Analysis	Bayesian
		Zhou [196]	$2001$	Histogram	Rule-based
		Bovis [17]	$2002$	Co-occurrence	kNN
		Petroudi [137]	$2003$	Textons	kNN
	Miller [121]	$1992$	Granulometric	Bayesian	Wolfe ($4$ )
		Byng [22]	$1996$	Histogram	Bayesian
		Karssemeijer [83]	$1998$	Histogram	kNN
		Blot [14]	$2001$	Co-occurrence	kNN
		Zwiggelaar [200,199]	$2003$	Co-occurrence	kNN
		Gong [59]	$2006$	Textons	Rule-based
		Martin [115]	$2006$	Histogram	Rule-based

Table  shows different proposals for breast classification. In the table, the works are classified according to their objective: breast density quantification or classification. Moreover, some other characteristics of the works, as the year, features, the type of classifier, and the number of categories used are shown. Note that among all the previous analyzed works, only the ones developed by Bovis and Singh [17] and Petroudi et al. [137] classified breasts according to BIRADS categories. Moreover, the classification algorithms are further separated into approaches that extract the features treating the global breast as a single region, and approaches that extract features segmenting the breast according to some parameters, for example, the distance to the skin-line [83].