The Breast 21 (2012) 503e506

Contents lists available at SciVerse ScienceDirect

The Breast
journal homepage: www.elsevier.com/brst

Original article

A rst evaluation of breast radiological density assessment by QUANTRA software

as compared to visual classication
Stefano Ciatto a, *, Daniela Bernardi a, Massimo Calabrese b, Manuela Durando c, Maria Adalgisa Gentilini d,
Giovanna Mariscotti c, Francesco Monetti b, Enrica Moriconi e, Barbara Pesce e, Antonella Roselli e,
Carmen Stevanin f, Margherita Tapparelli f, Nehmat Houssami g
a
U.O. Senologia Clinica e Screening Mammograco, Dipartimento di Radiodiagnostica, APSS, Trento, Italy
b
S.C. Senologia Diagnostica, IST, Istituto Nazionale per la Ricerca sul Cancro, Genova, Italy
c
Radiologia diagnostica 4 - U, A.O.U. San Giovanni Battista, Torino, Italy
d
Servizio Osservatorio Epidemiologico, Direzione promozione ed educazione alla salute, APSS, Trento, Italy
e
Diagnostica Senologica, C.d.C. Paideia, Roma, Italy
f
Servizio di Radiologia, Comprensorio Sanitario di Bolzano, Bolzano, Italy
g
Screening and Test Evaluation Program, School of Public Health, Sydney Medical School, University of Sydney, Sydney, Australia

a r t i c l e i n f o a b s t r a c t

Article history: Breast radiological density is a determinant of breast cancer risk and of mammography sensitivity and
Received 31 July 2011 may be used to personalize screening approach. We rst analyzed the reproducibility of visual density
Received in revised form assessment by eleven experienced radiologists classifying a set of 418 digital mammograms: reproduc-
22 December 2011
ibility was satisfactory on a four (BI-RADS D1e2e3e4: weighted kappa 0.694e0.844) and on a two
Accepted 11 January 2012
grade (D1-2 vs D3-4: kappa 0.620e0.851), but subjects classied as with dense breast would range
between 25.1 and 50.5% depending on the classifying reader. Breast density was then assessed by
Keywords:
computer using the QUANTRA software which provided systematically lower density percentage
Mammography
Radiological density
values as compared to visual classication. In order to predict visual classication results in
Computer assessment discriminating dense and non-dense breast subjects on a two grade scale (D3-4 vs, D1-2) the best
tting cut off value observed for QUANTRA was 22.0%, which correctly predicted 88.6% of D1-2, 89.8% of
D3-4, and 89.0% of total cases. Computer assessed breast density is absolutely reproducible, and thus to
be preferred to visual classication. Thus far few studies have addressed the issue of adjusting computer
assessed density to reproduce visual classication, and more similar comparative studies are needed.
2012 Elsevier Ltd. All rights reserved.

Introduction In the present study a set of digital mammograms was classied

Radiological breast density has been associated with breast
cancer risk on an individual1e4 or familial basis5e7 and to
mammography sensitivity, a signicant association of interval
cancer risk being associated with dense breast.8e12
The method most commonly used to classify mammography
density is quantitative classication, proposed by the American
College of Radiology in the BI-RADS system,13 commonly deter-
mined on a visual basis.3,5,9,14 Examples for the four D1, D2, D3 and
D4 categories are provided in Fig. 1. As visual assessment is
subjective and associated with suboptimal reproducibility15,16 its
replacement with absolutely reproducible computerized assess-
ment has been suggested.6,7,17,18
* Corresponding author. Tel.: 39 3486540748.
E-mail address: stefano.ciatto@gmail.com (S. Ciatto).
according to BI-RADS quantitative density classication by a panel
of radiologists currently involved in mammography reporting, and
by a dedicated software (QUANTRA 1.3),19,20 with the aim of
comparing the two methods and providing criteria to adjust
numerical computerized density assessment to currently used BI-
RADS categories.
Material and methods
A set of 418 two view digital mammograms acquired using
a Hologic Selenia Dimensions equipment was used for the study
purpose. Cases were randomly selected among women self refer-
ring to mammography either for subjective symptoms, or follow-up
after breast surgery or spontaneous screening at A.O.U. San Gio-
vanni Battista Radiology Department, Turin, from May 31st to July
21st 2010. Average age of subjects was 53 years (range 30e90).

0960-9776/$ e see front matter 2012 Elsevier Ltd. All rights reserved.
doi:10.1016/j.breast.2012.01.005
504 S. Ciatto et al. / The Breast 21 (2012) 503e506

Fig. 1. Examples of the four BI-RADS quantitative categories of breast density: (A) D1 (0e25% of breast volume occupied by broglandular densities); (B) D2 (26e50%); (C) D3
(51e75%); (D) D4 (76e100%).

R2 Quantra is a software application intended to assess volu-
metric breast density. The algorithm is based on a physical model of
the X-ray imaging chain that relates breast tissue X-ray attenuation
to the digital mammography images provided to the radiologist.
Assessments are based on published physical parameters for the
breast and the imaging system, as well as information about indi-
vidual X-ray exposures, including attenuation coefcients for breast
tissue21 and X-ray spectra for the target material.22
QUANTRA estimates the amount of broglandular tissue that an
X-ray must have penetrated in order to deposit a measured amount
of energy at the detector and provides a result in centimetres of
broglandular tissue penetrated at each pixel in the image. After
completing analysis on a pixel-by-pixel basis, it aggregates single
pixel values into the volume of broglandular tissue, given in cubic
centimetres. Breast density is calculated as a ratio of broglandular
tissue and total breast volume estimates. Individual density was
calculated for the study purpose as the average value of right and
left breasts numerical values.
Visual assessment of breast density was performed indepen-
dently by eleven radiologists, currently reporting mammography,
according to four-category (D1-4) BI-RADS classication quantita-
tive criteria. Radiologists were not all using BI-RADS density cate-
gories on a routine basis but were provided with original BI-RADS
classication criteria. Interobserver agreement in classifying
breast density was assessed according to kappa statistic, adjusting
for chance agreement. Conventionally, kappa values of 0.00e0.20,
0.21e0.40, 0.41e0.60, 0.61e0.80 and 0.81e1.00 indicate minimal,
fair, moderate, substantial and almost perfect agreement,
respectively.23
The majority report (mode) was used to compare visual to
QUANTRA classication for the study purpose. We calculated the
cut off of QUANTRA numerical breast density percentage which
allowed a separation of the study set into dense (BI-RADS D3e4)
and non-dense (BI-RADS D1e2) best tting with visual classica-
tion, in order to provide a practical criterion to translate QUANTRA
measures into conventional BI-RADS values for current clinical use.
Results
Table 1 shows the distribution of reports by the eleven readers
and the resulting majority report. Majority visual classication was
D1, D2, D3 or D4 in 25.8% (range 18.7e40.2%), 39.0% (range
28.7e47.6), 28.9% (range 21.5e36.6%), and 6.2% (range 2.9e22.0%)
of cases, respectively. Distribution on a two grade scale (D1-2 vs
D3-4) was 64.8% (range 49.5e72.0%) and 35.2% (range 25.1e50.5%),
respectively.
Interobserver reproducibility of readers as compared to majority
report was rather high, weighted kappa statistic on a four grade
scale (D1/D2/D3/D4) ranging from 0.694 to 0.844, with substantial
or almost perfect agreement being recorded for 6 or 5 readers,

Table 1
Distribution of visual density reports by the eleven readers, resulting majority report and kappa statistic of agreement with majority report (row percentages and 95%
condence limits are reported in brackets).

Reader D1 (%) D2 (%) D3 (%) D4 (%) Weighted kappa vs majority (D1/D2/D3/D4) Kappa vs majority (D1e2/D3e4)
A 137 (32.8) 164 (39.2) 105 (25.1) 12 (2.9) 0.799 (0.758e0.840) 0.802 (0.741e0.863)
B 87 (20.8) 120 (28.7) 153 (36.6) 58 (13.9) 0.727 (0.683e0.771) 0.685 (0.619e0.752)
C 168 (40.2) 130 (31.1) 94 (22.5) 26 (6.2) 0.765 (0.722e0.807) 0.841 (0.786e0.896)
D 113 (27.0) 181 (43.3) 91 (21.8) 33 (7.9) 0.838 (0.800e0.876) 0.788 (0.725e0.851)
E 124 (29.7) 157 (37.6) 117 (28.0) 20 (4.8) 0.840 (0.803e0.877) 0.851 (0.796e0.904)
F 132 (31.6) 100 (23.9) 94 (22.5) 92 (22.0) 0.694 (0.654e0.734) 0.777 (0.717e0.838)
G 78 (18.7) 188 (45.0) 130 (31.1) 22 (5.3) 0.804 (0.763e0.846) 0.818 (0.760e0.876)
H 92 (22.0) 169 (40.4) 136 (32.5) 21 (5.0) 0.844 (0.807e0.881) 0.835 (0.780e0.890)
J 91 (21.8) 154 (36.8) 132 (31.6) 41 (9.8) 0.830 (0.793e0.868) 0.808 (0.751e0.866)
K 114 (27.3) 168 (40.2) 118 (28.2) 18 (4.3) 0.763 (0.718e0.807) 0.738 (0.670e0.807)
L 114 (27.3) 199 (47.6) 90 (21.5) 15 (3.6) 0.750 (0.704e0.796) 0.652 (0.575e0.729)
Majority 108 (25.8) 167 (39.9) 117 (28.0) 26 (6.2) Ref. e
 S. Ciatto et al. / The Breast 21 (2012) 503e506 505

respectively. Reproducibility on a two grade scale (D1e2/D3e4)
was also high, ranging from 0.62 to 0.851, with substantial or
almost perfect agreement being recorded for 5 or 6 readers,
respectively.
QUANTRA classied cases with density values comprised
between 7.0 and 50.5%. Compared to majority visual report,
QUANTRA density ranged between 7.0 and 19.0% for D1 (mean 12.2,
median 12.0), 10.0 and 32.0% for D2 (mean 18.0, median 17.5), 16.0
and 47.0% for D3 (mean 28.5, median 27.5), and 24.5 and 50.5% for
D4 (mean 34.5, median 33.0) categories, respectively.
Table 2 shows the proportion of cases correctly classied as
dense (D3e4) or non-dense (D1e2) as compared to majority visual
classication when using different cut offs of QUANTRA density
values. The best accuracy was reached with a cut off value of
22.0%, which classied correctly 88.6% of D1-2, 89.8% of D3e4,
and 89.0% of total cases as compared to majority visual report.
the same QUANTRA algorithm was used, we presently have no
explanation for such a difference but for a substantial difference in
visual classication results by the two panels.
Computer assessed breast density is absolutely reproducible,
and thus to be preferred to visual classication. Still, all that we
know about the value of breast density as an individual risk factor
or as a determinant of mammography lower sensitivity is based on
visual classication. Finding the optimal QUANTRA cut off value to
reproduce visual classication looks a much more practical option
that repeating retrospective studies of the association of QUANTRA
assessed density and breast cancer risk or mammography sensi-
tivity. The above mentioned differences observed in the only two
studies addressing this aspect (the present study and that of
Rafferty et al.,19 to our knowledge), suggests that other similar
studies should be carried on comparing QUANTRA and visual
classication.

Discussion Conict of interest

The present study is based on a relatively large unbiased test set None declared.
and on a large reading panel, both conditions allowing for a reliable
estimate of the reproducibility of visual classication of breast References
density. The resulting high interobserver reproducibility conrms
previous observations.24 The fact that radiologists in the reading 1. Wolfe JN. Risk for breast cancer development determined by mammographic
parenchymal pattern. Cancer 1976;37:2486e92.
panel were not specically trained for classifying breast density, 2. Ciatto S, Zappa M. A prospective study of the value of mammographic patterns
adds to the value of such a high reproducibility and suggests that as indicators of breast cancer risk in a screening experience. Eur J Radiol
the resulting majority report is a reliable reference standard for 1993;17:122e5.
3. Brisson J, Diorio C, Masse B. Wolfes parenchymal pattern and percentage of the
visual classication.
breast with mammographic densities: redundant or complementary classi-
Although visual classication looks reproducible on a statistical cation? Cancer Epidemiol Biomarkers Prev 2003;12:728e32.
basis, it might have a suboptimal performance when used for 4. Vachon CM, Kuni CC, Anderson K, Anderson CE, Sellers TA. Association of
mammographic dened percent breast density with epidemiologic risk
practical purpose. For example, if breast density were to be used as
factors for breast cancer (United States). Cancer Causes Control
a determinant for personalized screening protocol (e.g. adding 2000;11:653e62.
ultrasonography or indicating yearly screening in subjects with D3- 5. Boyd NF, Dite GS, Stone J. Heritability of mammographic density, a risk factor
4 dense breast), subjects classied as with dense breast by visual for breast cancer. New Engl J Med 2002;347:886e94.
6. Ursin G, Ma H, Wu H. Mammographic density and breast cancer in three ethnic
classication would range between 25.1 and 50.5% depending on groups. Cancer Epidemiol Biomarkers Prev 2003;12:332e8.
the classifying reader, according to present study ndings. 7. Chen Z, Wuy AH, Gauderman WJ. Does mammographic density reect ethnic
Conrming previous reports19,20 QUANTRA provided systemat- differences in breast cancer incidence rates? Am J Epidemiol 2004;159:140e7.
8. Egan RL, Mosteller RC. Breast cancer mammography patterns. Cancer
ically lower values on a density percentage basis as compared to 1977;40:2087e90.
visual classication. In order to reproduce visual classication 9. Ciatto S, Visioli C, Paci E, Zappa M. Breast density as a determinant of interval
results in discriminating dense and non-dense breast subjects on cancer at mamographic screening. Br J Cancer 2004;90:393e6.
10. Peeters PH, Verbeek AL, Hendriks JH, Holland R, Mravunac M, Vooijs GP. The
a two grade scale (D3-4 vs, D1-2) the best tting cut off value occurrence of interval cancers in the Nijmegen screening programme. Br J
observed in the present study for QUANTRA was 22.0%, which Cancer 1989;59:929e32.
correctly classies 88.6% of D1-2, 89.8% of D3-4, and 89.0% of total 11. Young K, Wallis M, Blank R, Moss S. Inuence of number of views and
mammographic lm density on the detection of invasive cancers: results from
cases. This nding is substantially different from the report of
the NHS breast screening programme. Br J Radiol 1997;70:482e8.
Rafferty at al.19 based on a similar study design (264 cases test set 12. Buist DSM, Porter PL, Lehman C, Taplin SH, White E. Factors contributing to
and 15 radiologists reading panel), suggesting a 13% QUANTRA mammography failure in women aged 40e49 years. J Natl Cancer Inst
volumetric density value as the best cut-off to reproduce visual 2004;96:1432e40.
13. The ACR breast imaging reporting and data system (BI-RADS). 4th ed. Ameri-
classication. Had a 13% QUANTRA cut off being adopted in the canCollege of Radiology, www.acr.org; 2003.
present study, correctly classied cases would be 36.5% among D1- 14. Ciatto S, Bonardi R, Zappa M. Impact of replacement hormone therapy in
2, 100% among D3-4 and 58.8% among total cases. Assuming that menopause on breast radiologic density and possible complications of
mammography in the assessment of breast masses. Radiol Med
2001;101:39e43.
Table 2 15. Boyd NF, Wolfson C, Moskowitz M. Observer variation in the classication of
Proportion of cases correctly classied as dense (D3e4) or non-dense (D1e2) as mammographic parenchymal patterns. J Chronic Dis 1986;39:465e72.
compared to majority visual classication when using different cut offs of QUANTRA 16. Berg WA, Campassi C, Langenberg P, Sexton MJ. Breast imaging reporting and
density value (correctly classied cases are reported in parentheses). data system: inter- and intraobserver variability in feature analysis and nal
assessment. Am J Roentgenol 2000;174:1769e77.
QUANTRA D1e2 visual D3e4 visual Overall proportion of correctly 17. Vachon CM, Sellers TA, Vierkant RA, Wu FF, Brandt KR. Case control study of
value (n 271) (n 147) classied cases (cases) increased mammographic breast density response to hormone replacement
therapy. Cancer Epidemiol Biomarkers Prev 2002;11:1382e8.
15.5% 58.7% (159) 100.0% (147) 73.2% (306)
18. Zhou C, Chan HP, Petrick N. Computerized image analysis estimation of breast
21.5% 87.4% (237) 90.5% (133) 88.5% (370) density on mammograms. Med Phys 2001;28:1056e69.
22.0% 88.6% (240) 89.8% (132) 89.0% (372) 19. Rafferty E, Smith A, Niklason L. Comparison of three methods of estimating
22.5% 90.4% (245) 85.7% (126) 88.8% (371) breast density: BI-RADS density scores using full eld digital mammography,
23.0% 91.9% (249) 82.3% (121) 88.5% (370) breast tomosynthesis, and volumetric breast density. Proffered paper at Rad.
23.5% 92.6% (251) 79.6% (117) 88.0% (368) Soc. North Am. Chicago, USA: SsM01; 2009.
24.0% 93.7% (254) 76.2% (112) 87.6% (366) 20. Tuncbilek N, Sezer A, Ug ur U, Ulas U, Urut U, Cosar R, zdemir H. Qualitative
24.5% 95.2% (258) 73.5% (108) 87.6% (366) and quantitative analysis of broglandular tissue in the digital environment.
32.5% 100% (271) 33.3% (49) 77.5% (324) Proffered paper at 10th National Congress of Breast Diseases, Izmir, Turkey;
2009.
506 S. Ciatto et al. / The Breast 21 (2012) 503e506

21. Johns PC, Yaffe MJ. X-ray characterization of normal and neoplastic breast
tissue. Phys Med Biol 1987;32:675e95.
22. Boone JM, Fewell TR, Jennings RJ. Molybdenum, rhodium, and tungsten anode
spectral models using interpolating polynomials with application to
mammography. Med Phys 1997;24:1863e74.
23. Landis JR, Koch GG. The measurement of observer agreement for categorical
data. Biometrics 1977;33:159e74.
24. Ciatto S, Houssami N, Apruzzese A, Bassetti E, Brancato B, Carozzi F, et al.
Categorizing breast mammographic density: intra- and interobserver repro-
ducibility of BI-RADS density categories. Breast 2005;14:269e75.