Open Access

LBPV descriptors-based automatic ACR/BIRADS classification approach

  • Alima Damak Masmoudi1Email author,
  • Norhen Gargouri Ben Ayed1,
  • Dorra Sellami Masmoudi1 and
  • Riad Abid2
EURASIP Journal on Image and Video Processing20132013:19

https://doi.org/10.1186/1687-5281-2013-19

Received: 28 November 2012

Accepted: 11 March 2013

Published: 17 April 2013

Abstract

Mammogram tissue density has been found to be a strong indicator for breast cancer risk. Efforts in computer vision of breast parenchymal pattern have been made in order to improve the diagnostic accuracy by radiologists. Motivated by recent results in mammogram tissue density classification, a novel methodology for automatic American College of Radiology Breast Imaging Reporting and Data System classification using local binary pattern variance descriptor is presented in this article. The proposed approach characterizes the local density in different types of breast tissue patterns information into the LBP histogram. The performance of macro-calcification detection methods is developed using FARABI database. Performance results are given in terms of receiver operating characteristic. The area under curve of the corresponding approach has been found to be 79%.

Keywords

MammogramBreast tissueTextureClassificationFeature extractionMacro-calcification detectionLBPLBPVROCACR/BIRADSCADANN

1 Introduction

Female breast cancer is a common cause of cancer-related deaths in women, especially in western countries and where statistics are available. Mammographic images are hard to interpret because of the textural morphology information complexity of the breast and the number of image parameters that affect the acquisition of mammograms [1].

In the evaluation of mammogram images, CAD (Mammographic Computer-Aided Diagnosis) systems are aimed at assisting radiologists [2, 3]. Studies in CAD systems tend to concentrate on the detection and classification of mammographic masses and micro-calcification [4]. In addition, recent research has shown that the sensitivity of these systems to detect masses in mammograms is significantly decreased as the density of the tumors increases [5]. These classification is based on both severity for the disease and image properties.

Classification with textural mammogram information can be based on a number of categories that might not explain the same mammographic features [48]. ACR/BIRADS classification [7] is becoming a standard in the assessment of mammogram images, which are classified in fourth categories according to their density (Figure 1).

  • ACR/BIRADS I: the breast is almost entirely fatty.

  • ACR/BIRADS II: there is some fibro-glandular tissue.

  • ACR/BIRADS III: the breast is heterogeneously dense.

  • ACR/BIRADS IV: the breast is extremely dense.

Figure 1

Four samples in the ACR/BIRADS classes with respect to density: (a) ACR/BIRADS I, (b) ACR/BIRADS II, (c) ACR/BIRADS III, (d) ACR/BIRADS IV.

It is well known that there is a strong correlation between textural density in mammographic image and the risk of developing tumors [8]. Figure 1 shows four samples in the American College of Radiology’s Breast Imaging Reporting and Data System (ACR/BIRADS) classes with respect to density.

ACR/BIRADS classification will be beneficial, to characterize what the mammogram density is since it is a parameter criterion in Breast Imaging Reporting and Data System classification, as well as to create an optimal approach to follow if, for example, detecting masses in mammographic image of tissue abnormality detection (marcocalcification/microcalcification). An early step of density classification will switch on for each class, this step is more efficient for segmentation approach. For example, detecting a macrocalcification in surrounding darker pixels as in the first class could be done differently if neighbors are brighter as in the third or four class.

In this article, a novel approach to automatic breast tissue classification is investigated. The first step of the proposed approach is a pre-processing denoising module. In the next step, texture features are extracted using a novel descriptor named local binary pattern variance (LBPV). Performance evaluation is based on testing the algorithm on images from a new Tunisian database located at the radiology center EL FARABI Sfax Tunisia. Then, an artificial neural network (ANN) is used for classifying the breast density tissue. To highlight the interest of such classification, a brief summary on methods in segmentation masses region for each density class is provided.

In Section 2, we briefly describe our image processing and classification system. Then, we present the proposed textural feature extraction using LBPV descriptor and we provide a brief summary of the breast segmentation approaches. In Section 3, we present the performances results followed by conclusion in Section 4.

2 System module description

The functionality of the proposed system is provided by means of four modules: (a) the preprocessing module, (b) the feature extraction module, (c) the ANN classification module, and (d) the mass segmentation (Figure 2).
Figure 2

The proposed method of macro-calcification detection.

2.1 Preprocessing

In this module, input images are prepared for the processing steps that will follow. The basic need for preprocessing in mammographic images is to increase the contrast, especially for dense breasts.

To remove the background noise [9], preserving the local information of suspicious areas can enhance mammograms. This approach was proposed by Lai et al. [10], who used modified median filtering and four selective averaging schemes called selective median filter. The output of this filter within a window W(n,m), centered at image coordinates (n,m) [11] is given by
I n , m = median I r , s : ( r , s ) N ( n , m ) I r , s I n , m < I th
(2.1)

where I n,m is the image intensity at (n,m), N(n,m) is the area in the image covered by the window W(i,j), and I th is the threshold.

The set of pixels, in computing the median, is restricted to those with a difference in gray level not greater than a some threshold I th. Adjusting the parameter I th allows to control the amount of edge smearing and to remove the background noise [12].

2.2 A novel textural feature extraction using LBPV descriptor

LBP [13] is a gray scale operator which characterizes the textural structure in the image. A pattern number is computed by comparing a central pixel in the gray image value with its neighborhoods:
LBP P , R = P = 0 P 1 Sign ( g p g c ) 2 P
(2.2)
Sign ( x ) = 1 , x 0 0 , x < 0 ,
(2.3)

where g c is the gray value of the central pixel, g p is the value of its neighbors, P is the number of neighbors, and R is the radius of the neighborhood.

If the coordinates of g c are (0,0), then the coordinates of g p are given by R sin ( 2 Πp P ) , R cos ( 2 Πp P ) . The gray values of neighbors which do not fall exactly in the center of pixels are estimated by interpolation.

Suppose the texture image is NM. After identifying the LBP pattern of each pixel (i,j), the whole texture image is represented by building a histogram:
H ( k ) = i = 1 N j = 1 M f ( LB P P , R ( i , j ) , k ) , k [ 0 , K ]
(2.4)
f ( x , y ) = 1 , x = y 0 , otherwise
(2.5)
To remove the effect of rotation in local image the LBP P , R ri (rotation invariant local binary pattern) is defined by:
LBP P , R ri = min { ROR ( LBP P , R , i ) | i = 0 , 1 , , P 1 }
(2.6)

where ROR(x,i) is a circular bit-wise right shift on the P-bit number.

The rotation invariance with uniform patterns LBP P , R riu 2 is proposed by Ojala et al. [13].
LBP P , R riu 2 = p = 0 p 1 sign ( g p g c ) , if U ( LBP P , R ) 2 p + 1 , otherwise
(2.7)
where superscript r i u 2 reflects the use of rotation invariant uniform patterns, U introduces the uniformity measure
U ( LBP P , R ) = sign ( g p 1 g c ) sign ( g 0 g c ) + p = 1 p 1 sign ( g p g c ) sign ( g p 1 g c ) .
(2.8)
A rotation invariant variance measures (VAR)of the local image can be defined as follows:
VAR P , R = 1 p p = 0 p 1 sign ( g p μ ) 2
(2.9)
where
μ = 1 p p = 0 p 1 sign ( g p )
(2.10)

LBP P , R riu 2 VAR P , R is the powerful descriptor of local contrast information because it exploits the complementary characteristics of local spatial patterns and local contrast [13].

Usually high-frequency textural information regions will have higher variances and contribute more to the discrimination of images. Therefore, LBPV is proposed to characterize the local contrast information into a one-dimensional LBP histogram where VAR P,R can be used as an adaptive weight to adjust the contribution of LBP descriptor in histogram calculation. The local binary variance code histogram is computed as follows:
LBPV P , R = i = 1 N j = 1 M w ( LBP P , R riu 2 ( i , j ) , k ) , k [ 0 , K ]
(2.11)
w ( LBP P , R riu 2 ( i , j ) , k ) = VAR P , R ( i , j ) , LBP P , R riu 2 ( i , j ) = k 0 otherwise
(2.12)
Though locally operated, LBPV operator has an excellent ability to summarize the different global or local densities in the image. Figure 3 illustrates that each density in the image can be illustrated by a respective peak in the occurrence histogram. Therefore, LBPV can be seen as a good descriptor candidate that could eventually mirror and handle sensitively different densities in mammographic images.
Figure 3

Perfect ability of LBPV operator to describe global density in the image: (a) Binary image, (b) Histogram LBPV of binary image, (c) Gray level image, (d) Histogram LBPV of gray level image.

The LBPV-based descriptor operator is a good contrast distribution method. Two occurrence histograms are, respectively, generated. Taking R=1 and P=8, the size of the vector is fixed to 512. This size is unique and independent of minimum number or image size. Such advantage overcomes limitations of direct matching discrepancies. Figure 4 illustrates a mammogram sample image with the corresponding histogram.
Figure 4

Perfect ability of LBPV operator to describe global densities in the image.

2.3 Breast tissue density classification module with ANN

Usually, ANNs can be considered as an information of system which is composed of interconnected elements constituting a network. Inspired by the biological nervous system, ANN adjusts weights between neurons (Table 1).
Table 1

Output of ANN classifier

Output

y3

y2

y1

ACR/BIRADS I

0

0

1

ACR/BIRADS II

0

1

0

ACR/BIRADS III

0

1

1

ACR/BIRADS VI

1

0

0

Let x=(x 1,x 2,...,x d ) T be an input vector and w=(w 1,w 2,...,w d ) T the weight vector, the output is determined as follows:
y = g w T x b = g i = 1 d w i x i b
(2.13)
where g(.) is namely a sigmoidal activation function defined by
g ( x ) = ( 1 + e x ) 1
(2.14)
For each mammogram samples, LBPV vectors are computed. The description based on these features is used in the recognition step as neural network inputs. The neural network architecture is represented in Figure 5.
Figure 5

The applied network architecture.

To be able to evaluate the effectiveness of the training, one can measure the relative error as follows:
E = i = 1 n ( I N I T ) Number of samples ,
(2.15)

where I N is the image resulting from ANN output and I T is the target.

After training steps, generalization error was evaluated for different features and network conditions. Figure 6 shows the evolution of training and generalization errors by incrementing the number of hidden layers.
Figure 6

Evolution of training and generalization errors by increasing the number of hidden layers.

2.4 Breast segmentation

In general, masses in low-density breasts are better detected than masses in high-density breasts, although each algorithm performs differently with regard to this.

2.4.1 Detection of concentric layers

In [14], the segmentation of masses by detection of concentric layers, using progressively lower average intensity, is proposed. This approach with a region granulation (i.e., a grey-level transformation) reduces the large number of intensity levels.

The grey-level transformation step starts by linearly normalizing the intensity between levels 0 and 1. Next, the pixels are assigned a grey level. This is done by sequentially visiting each granule pixel and examining its local neighborhood. If all neighbors are within 98% of the granularity, they are assigned to the same grey level, else they will be assigned to a different granulation. After this transformation, a morphological opening is performed to decrease scattered grey levels.

The segmentation of suspicious masses regions is based on the inspection of the granularity. Thus, all the regions with similar or higher levels are grouped. For each level, a set of features, including area, eccentricity, solidity, and dispersion, are computed. This procedure is repeated for the brightest levels. Therefore, the region growing is established, and the suspicious regions are those containing at least three developing concentric layers. All the parameters used have empirically been adjusted to FARABI database. One of these parameters is related to the minimum distance between possible macro-calcifications. This distance will be used later to obtain probability images.

2.4.2 Thresholding approach

This approach is proposed by Kom et al. [15]. The corresponding algorithm is based on the thresholding mammographic image obtained by subtracting from the mammogram a linear filtered representation of itself.

The enhanced image I EN is given for each coordinates (x,y) by:
I EN ( x , y ) = a × log ( 1 + m × I m ( x , y ) ) , I m ( x , y ) < α
(2.16)
I EN ( x , y ) = exp ( I m ( x , y ) a ) 1 b , I m ( x , y ) > α
(2.17)

where I m (x,y) is the original mammogram, m is its maximum grey level, b = 1 exp ( m a ) m , a and α are two parameters fixed experimentally; in this study, a=10 and α=0.3.

Consequently, the subtracted image I Sub is given for each pixel (x,y) by:
I Sub ( x , y ) = I m ( x , y ) I EN ( x , y )
(2.18)

At last, the subtracted image is thresholded by using an adaptive local threshold to obtain suspicious macro-calcifications.

2.4.3 Laplacian edge detector approach

This method is proposed by Petrick et al. [16] who used an optimal Laplacian Gaussian edge detector (LGED) with the aim of finding closed regions in the enhanced version of the mammographic image. This approach begins by preprocessing the mammogram using a density-weighted contrast enhancement (DWCE) filter, which is based on two filtered mammograms of the original image I m (x,y): the first is the density image I F D ( x , y ) , which is a smoothed version of the image, obtained by using a Gaussian filter. The next is the contrast image I F C ( x , y ) , obtained by subtracting the original mammogram from a second smoothed version of the image.

The density of mammograms is filtered all over again with a nonlinear filter K M and used to define a multiplication factor which adjusts the related pixel in the contrast image. Thus, the density parameter of each pixel is weighted by local contrast. The intermediate image I F KC can be expressed by
I F KC ( x , y ) = K M ( I F D ( x , y ) ) I F C ( x , y )
(2.19)
where I F KC ( x , y ) is used to define a next multiplication value using another nonlinear filter K N L , which is multiplied by the weighted contrast of the related pixels
I F E ( x , y ) = K NL ( I F KC ( x , y ) ) I F KC ( x , y )
(2.20)

where I F E ( x , y ) is the output of the DWCE filter. The output of this filtering process is a mammogram where the potential masses are highlighted.

To detect macro-calcifications, an edge detector is used. An LGED was applied. It is defined by
LGED ( x , y ) = 2 G ( x , y ) I F E ( x , y ) ,
(2.21)

where G(x,y) is a two-dimensional Gaussian smoothing function.

2.4.4 Classifier approach

This method is proposed by Karssemeijer and te Brake [17, 18]. The classification approach makes possible the detection of macro-calcifications using second-order Gaussian derivative operators. If a line-like structure is present at a given site, this algorithm provides an estimation of the orientation of textural mammogram tissue. With this information, two new features are built. The first one characterizes the total number of pixels pointing the center, while the next feature estimates whether these directions are circularly oriented. With both features, and a set of classified mammograms, this approach trains a binary decision tree. Afterward, the decision tree can be used for macro-calcification detection in medical image.

3 System performances

In order to test the proposed method, images from EL FARABI database are used. Images in this database have their density classified according to ACR/BIRADS categories.

3.1 EL FARABI database

The mammograms used in this study were collected from EL FARABI radiologic center. They are acquired using different mammographic screen/film system and settings (all with molybdenum anode) and in the framework of different applications. The obtained images including both clinical routine carried out on symptomatic women and screening programs addressed to asymptomatic women. All the images were digitized with a CCD scanner at a pixel size of 85×85 μ m2 with a 12-bit resolution. Each image is thus 3560×4640 pixels with G=212=4096 gray-level tones. No normalization was applied to images. The database consists of 2052 mammograms from 342 analyzed subjects. Some of the mammograms show different views (craniocaudal, lateral, oblique) of the same subject and are considered as different samples in our analysis. The diagrams reported in Figure 7 show the partition of the database in left/right breast images (left) and craniocaudal/oblique/ lateral views (right).
Figure 7

Database partition. Left: number of left/right breast images. Right: number of craniocaudal/oblique/lateral views.

3.2 Receiver operating characteristics (ROC)

ROCs are usually used in many fields for decision making to validate a given classification method. In this study, we use it for a validation of ACR/BIRADS mammogram density classification.

3.2.1 ROC curves

An ROC curve is a graphical visualization of the TPR (True Positive Rate) as a function of the FPR (False Positive Rate) of mammogram classifier systems.

3.2.2 The ROC convex hull method

The ROCCH (ROC convex hull) method accommodates both binary and continuous ROC curve. Binary recognitions are represented by individual points in ROC space. Continuous ROC produces numeric outputs of thresholds that can be applied, yielding a series of (FPR,TPR) pairs forming an ROC curve. Each point may or may not contribute to the ROC convex hull.

3.2.3 Area under ROC curve

The area under the ROC curve (AUC) is common metric that can be used for performance evaluation of a decision-making system. It constitutes an efficient way to compare different tests (indicator variables). For performance evaluation, it is convenient to reduce the ROC curve to a single scalar value representing expected performance. The easiest possibility is to calculate the AUC which is a part of the area of the unit square. Consequently, the value of AUC will always satisfy the following inequalities
0 AUC 1
(3.1)
It is clear that if the AUC is close to 1 (area of unit square), AUC indicates very good diagnostic test. However, as the random guessing produces the diagonal line between the points [0,0] and [1,1], which has an area of 0.5, reasonable tests should have
0 . 5 AUC 1
(3.2)

The AUC has an important statistical property: the AUC of an ROC relative to a recognition system is equivalent to the probability that the recognition will evaluate randomly chosen positive instance higher than a randomly chosen negative instance.

3.3 ACR/BIRADS automatic classification results

The method was applied to a set of 400 image mammographic taken from the FARABI Digital Database. This database provides for each mammogram additional information, including the density of the breast determined by an expert according to BIRADS categories. In order to simulate the real world, our database is formed by 50 mammograms with ACR/BIRADS I, II, III, and IV, so 200 images are used for training. However, 200 mammographic image are used for generation test.

Figure 8 shows that the AUC of this approach has a value of 0.97.
Figure 8

ROC curve for ACR/BIRADS automatic classification.

3.4 Breast tissue influence

The breast density evaluation experiment is related to the ability of each method to detect macro-calcification in all FARABI database images. This evaluation mimics the radiologist in identifying the presence of tumors.

The accuracy of the approach is depending on breast tissue density classification and is summarized in Table 2. This table shows the AUC for each method detailed for each ACR/BIRADS density. Generally, macro-calcification are detected in low-density images than in high-density images.
Table 2

Influence of the breast density based on AUC

 

Detection of

Thresholding

Laplacian edge

Classifier

 

concentric layers

approach

detector approach

approach

ACR/BIRADS I

0.81

0.71

0.70

0.72

ACR/BIRADS II

0.64

0.79

0.69

0.65

ACR/BIRADS III

0.59

0.61

0.74

0.63

ACR/BIRADS VI

0.52

0.58

0.59

0.78

For instance, looking at experimental results, algorithms for detection of concentric layers have better performance on fatty breasts tissues on ACR/BIRADS I compared to other density classes. Therefore, thresholding and Laplacian edge detector approaches get the best accuracy for mammographic images belonging to ACR/BIRADS II and ACR/BIRADS III, respectively. Classifier approach performs better for the most dense tissue (ACR/BIRADS IV).

The reason for such different behaviors is related to different factors. For example, macro-calcifications in fatty mammograms frequently have a more delineated boundary than in denser medical images. Moreover, one can see a set of circumscribed layers around the macro-calcifications that are exploited in the granularity algorithm (i.e., detection of concentric layers). Thresholding and Laplacian edge detector approaches seem beneficial for medical images belonging to intermediate ACR/BIRADS II and ACR/BIRADS III classes, where macro-calcifications are highlighted with respect to the normal tissue. To end with classifier approach which performs better for the dense tissues (ACR/BIRADS IV). In fact, it uses the contour information for masse detection and it has better performances for increased intensity changes in mammograms.

4 Conclusion

A novel automatic ACR/BIRADS classification for segmentation of mammographic masses is presented. To exploit the local and global textural information in mammographic images, the LBPV was proposed. This descriptor characterizes globally rotation invariant matching with locally variant LBP features for mammogram texture classification. This approach is tested on 342 pairs of patient mammograms.

As performances metric, we get for the EL FARABI database an AUC which has value of 0.79.

A segmentation technique has also been done, describing several methods and pointing out their specific features. These approaches have fully been evaluated using ROC curve analysis and tested using a digitized database. Annotations used as the gold standard were provided by expert radiologists who read mammograms routinely.

Segmentation results depend on the breast density. Based on our testing of these algorithms on the FARABI database, abnormal mammograms belonging to ACR/BIRADS I tend to show improved detection over abnormal mammograms belonging to other ACR/BIRADS category. This is related to the increase of the contrast in parenchymal tissue, which is mistaken for abnormal regions.

Declarations

Authors’ Affiliations

(1)
Computers Imaging and Electronics Systems (CIELS), University of Sfax, Sfax Engineering School, Sfax-Tunisia
(2)
EL FARABI Radiology Center

References

  1. Byng JW, Boyd NF, Fishell E, Jong RA, Yaffe MJ: Automated analysis of mammographic densities. Phys. Med. Biol 1996, 41: 909-923. 10.1088/0031-9155/41/5/007View ArticleGoogle Scholar
  2. Birdwell RL, Ikeda DM, O’Shaughnessy KD, Sickles EA: Mammographic characteristics of 115 missed cancers later detected with screening mammography and the potential utility of computer-aided detection. Radiology 2001, 219: 192-202.View ArticleGoogle Scholar
  3. Freer TW, Ulissey MJ: Screening mammography with computeraided detection: prospective study of 01286 patients in a community breast center. Radiology 2001, 220: 781-786. 10.1148/radiol.2203001282View ArticleGoogle Scholar
  4. Tabar L, Tot T, Dean PB: Breast Cancer—The Art and Science of Early Detection With Mammography. Germany: Georg Thieme Verlag, Stuttgart; 2005.View ArticleGoogle Scholar
  5. Ho WT, Lam PWT: Clinical performance of computer-assisted detection (CAD) system in detecting carcinoma in breasts of different densities. Clin. Radiol 2003, 58: 133-136. 10.1053/crad.2002.1131View ArticleGoogle Scholar
  6. Boyd NF, Byng JW, Jong RA, Fishell EK, Little LE, Miller AB, Lockwood GA, Tritchler DL, Yaffe MJ: Quantitative classification of mammographic densities and breast cancer risk: results from the Canadian national breast screening study. J. Nat. Cancer Inst 1995, 87: 670-675. 10.1093/jnci/87.9.670View ArticleGoogle Scholar
  7. American College of Radiology: Illustrated Breast Imaging Reporting and Data System BIRADS. Philadelphia, PA: American College of Radiology; 1998.Google Scholar
  8. Wolfe JN: Risk for breast cancer development determined by mammographic parenchymal pattern. Cancer 1976, 37: 2486-2492. 10.1002/1097-0142(197605)37:5<2486::AID-CNCR2820370542>3.0.CO;2-8View ArticleGoogle Scholar
  9. Yang S-C, Wang C-M, Chung Y-N, Hsu G-C, Lee S-K, Chung P-C, Chang C-I: A computer aided system for mass detection and classification in digitized mammograms. Biomed. Eng.- Appl. Basis Commun 2005, 17: 215—228.Google Scholar
  10. Lai S, Li X, Bischof W: On techniques for detecting circumscribed masses in mammograms. IEEE Trans. Med. Imagi 1989, 8: 377-386. 10.1109/42.41491View ArticleGoogle Scholar
  11. Gonzalez RC: Digital Image Processing Using Matlab. Pearson Publication; 2005.Google Scholar
  12. Sampat MP, Markey MK, Bovik AC: Computer-Aided Detection and Diagnosis in mmammography. Amsterdam: Elsevier Academic Press; 2005.Google Scholar
  13. Ojala T, Pietikäinen M, Mäenpää TT: Multiresolution gray-scale and rotation invariant texture classification with local binary pattern. IEEE Trans. Pattern Anal. Mach. Intell 2002, 24(7):971-987. 10.1109/TPAMI.2002.1017623View ArticleGoogle Scholar
  14. Eltonsy NH, Tourassi GD, Elmaghraby AS: A concentric morphology model for the detection of masses in mammography. IEEE Trans. Med. Imagi 2007, 26(6):880-889.View ArticleGoogle Scholar
  15. Kom G, Tiedeu A, Kom M: Automated detection of masses in mammograms by local adaptive thresholding. Comput. Biol. Med 2007, 37(1):37-48.Google Scholar
  16. Petrick N, Chan HP, Sahiner B, Wei D: An adaptive density-weighted contrast enhancement filter for mammographic breast mass detection. IEEE Trans. Med. Imagi 1996, 15(1):59-67.View ArticleGoogle Scholar
  17. Karssemeijer N, te Brake GM: Detection of stellate distortions in mammograms. IEEE Trans. Med. Imagi 1996, 15(5):611-619. 10.1109/42.538938View ArticleGoogle Scholar
  18. Karssemeijer N, te Brake GM: Combining single view features and asymmetry for detection of mass lesions. Digital Mammography 1998, 25: 95-102.View ArticleGoogle Scholar

Copyright

© Damak Masmoudi et al.; licensee Springer. 2013

This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.