Home Search Collections Journals About Contact us My IOPscience

Comparing measurements of breast density

This content has been downloaded from IOPscience. Please scroll down to see the full text.

2007 Phys. Med. Biol. 52 5881

(http://iopscience.iop.org/0031-9155/52/19/010)

View the table of contents for this issue, or go to the journal homepage for more

Download details:

IP Address: 131.94.16.10
This content was downloaded on 03/09/2015 at 13:21

Please note that terms and conditions apply.

IOP PUBLISHING PHYSICS IN MEDICINE AND BIOLOGY

Phys. Med. Biol. 52 (2007) 5881–5895 doi:10.1088/0031-9155/52/19/010

Comparing measurements of breast density

R Highnam1, M Jeffreys2, V McCormack3, R Warren4, G Davey Smith5
and M Brady6
1 Highnam Associates Limited, Wellington, New Zealand
2 Massey University, Wellington, New Zealand
3 Department of Epidemiology and Population Health, London School of Hygiene and Medicine,

London, UK
4 Addenbrooke’s Hospital, Cambridge, UK
5 Department of Clinical Epidemiology, University of Bristol, Bristol, UK
6 Department of Engineering Science, University of Oxford, Oxford, UK

E-mail: rph@highnam.com

Received 11 January 2007, in final form 13 July 2007

Published 14 September 2007
Online at stacks.iop.org/PMB/52/5881

Abstract
Breast density measurements can be made from mammograms using either
area-based methods, such as the six category classification (SCC), or volumetric
based methods, such as the standard mammogram form (SMF). Previously, we
have shown how both types of methods generate breast density estimates which
are generally close. In this paper, we switch our attention to the question of why,
for certain cases, they provide widely differing estimates. First, we show how
the underlying physical models of the breast employed in the methods need to
be consistent, and how area-based methods are susceptible to projection effects.
We then analyse a set of patients whose mammograms show large differences
between their SCC and SMF assessments. More precisely, 12% of 657 patients
were found to fall into this category. Of these, 2.7% were attributable to errors
either in the SMF segmentation algorithms, human error in SCC categorization
or poor image exposure. More importantly, 9.3% of the cases appear to be due
to fundamental differences between the area- and volume-based techniques.
We conclude by suggesting how we might remove half of those discrepancies
by introducing a new categorization of the SMF estimates based on the breast
thickness. We note however, that this still leaves 6% of patients with large
differences between SMF and SCC estimates. We discuss why it might not
be appropriate to assume SMF (or any volume measure) has a similar breast
cancer risk prediction capability to SCC.

1. Introduction

There is an increasing interest in the use of quantitative measures derived from breast images,
for a growing range of applications including the estimation of breast cancer risk, radiation
0031-9155/07/195881+15$30.00 © 2007 IOP Publishing Ltd Printed in the UK 5881
5882 R Highnam et al

Table 1. Six category classification (SCC): a visual, area-based assessment of breast density is
made and then the breast categorized as follows.

SCC Class Breast density

1 0%
2 1–10%
3 11–24%
4 25–49%
5 50–74%
6 75–100%

dose monitoring and the analysis of the effects of hormone replacement therapy (Boyd et al
1998, Heine and Malhotra 2002).
Such quantitative measures of breast composition can be derived using a variety of visual,
computer-assisted and fully automated methods and are typically designed to estimate what is
known generically as a ‘breast density’ percentage, although estimates and the use of actual
volumes of tissue are now attracting increasing interest.
The various methods for estimating breast composition measures are at differing stages
of development, so comparisons between more recent and more established ones can provide
useful insights into the advantages of each. In this paper, we examine and explore a number
of different breast composition measures and see how and why they interrelate as they do.
There are two broad classes of breast composition measures—area and volumetric. The
area measures work directly from the image of the compressed, projected breast and attempts
to quantify breast composition by segmenting (automatically or semi-automatically) certain
areas of the mammographic image (Byng et al 1994, Karssemeijer 1998, Zhou et al 2001,
Sivaramakrishna et al 2001, Zwiggelaar and Denton 2006). For example, breast density might
be estimated as the ratio of the projected area of segmented tissue regions deemed to be ‘dense’
to the projected area of the whole breast. Area methods, in effect, make a binary decision at
each pixel about whether or not the column of tissue above it is fatty or dense. This rather
simplistic assumption, in combination with the knowledge of the compressed breast thickness,
enables a rough estimate of the volume of dense tissue; but, to date, we have not found any
publications that use this approach, presumably because of concerns about partial volume
effects and breast thickness accuracy.
In this paper, we focus on the area-based six category classification (SCC) system (Byng
et al 1994) which categorizes the breast into one of six classes based on a visual assessment;
see table 1.
Volumetric measures attempt to quantify breast composition by using the pixel value to
derive information about the x-ray attenuation properties of the column of breast tissue above
that pixel. Based on those properties and on published x-ray attenuation measurements, most
volumetric techniques reduce the x-ray attenuation measure into a thickness of ‘interesting’
(‘non-fatty’ or ‘dense’) tissue (hint (x, y)) and a thickness of adipose tissue (hfat (x, y)) at each
pixel (x, y) (Byng et al 1994).
In this paper, we focus on the volumetric technique based on the standard mammogram
form (SMF), as implemented in version 2.2β of the GenerateSMF software (Siemens
Molecular Imaging, Oxford). GenerateSMF uses a model of the imaging process to
automatically map the pixel value to the thicknesses of the relevant tissues (Highnam and
Brady 1999, Highnam et al 2006). It is then straightforward to work out estimates of the
volumes of interesting tissue and breast density; see equations (1) and (2):


SMF volume = volume of interesting tissue = hint (x, y) dx dy (1)
A
Comparing measurements of breast density 5883

where A is the projected area of the breast (note that the user can select whether or not the fatty
breast edge is included). A measure of breast density, which we term SMF%, then follows
naturally, as in equation (2):
SMF% = SMF volume/Breast volume. (2)
Previous publications have shown that SMF% is often closely correlated with SCC
(Jeffreys et al 2006). However, whereas SCC has a known relationship with breast cancer
risk, the same cannot be assumed for SMF%, at least not without further investigation, which
is currently underway. In this paper, we analyse cases where SMF% and SCC do not agree.
Area and volumetric measures have a number of obvious differences but they also embody
more subtle differences which relate to the actual physical definitions of breast density on which
they are based. The breast is generally considered to consist of a volume of heterogeneous
breast tissue (sometimes called the ‘glandular disk’) which is encapsulated by a layer of
adipose tissue (we call this ‘the breast edge’). Different measures may use the breast edge in
different ways, or indeed they may ignore it altogether. We use results from SCC, SMF and
also ‘glandular disk density’ (Dance et al 2000) to highlight the effects of using, or ignoring,
the breast edge.
In this paper, we proceed from a theoretical analysis of physical breast density definitions,
as well as area versus volumetric estimation techniques. We then analyse both the images and
the results of applying the various techniques to a database of 4096 digitized mammograms;
then we use our theoretical analysis to understand why SMF% and SCC differ in certain cases.
Finally, we end with a suggestion as to how we could make SMF and SCC correlate even
better than they already do.

2. Breast density physical definitions

Consider a highly simplified imaging system with a 2D breast and parallel x-ray beams as
shown in the top-left of figure 1, where the inner, dotted ellipse represents the border between
the fatty breast edge and the interior glandular disk, and the dark circle represents an area of
interesting tissue, which in this example consists of 100% glandular tissue.
We define the following physical breast density measures which are illustrated in
figure 1:
• BD1. Percentage of glandular tissue within the glandular disk. This is the percentage
investigated primarily by physicists (Dance et al 2000), and, in our 2D example, would
be the area of the dark circle divided by the area of the shaded inner ellipse.
• BD2. Percentage of glandular tissue within the breast, but excluding the ‘side parts’
of the breast edge. This definition takes into account the observation that when the
breast is projected onto a mammogram, the layer of fat at the sides of the breast is often
invisible to the naked eye, so visual area estimates typically ignore that area. In our 2D
example, breast density would be the area of the dark circle divided by the area of the
shaded rectangle. We contend that the visually assessed SCC is probably such a measure.
Furthermore, GenerateSMF generates a similar measure when it is run with the program
option telling it to ignore tissue in the breast edge for composition calculations.
• BD3. Percentage of glandular tissue within the whole breast including the breast edge.
In our 2D example, this would be the area of the dark circle divided by the area of the
shaded outer ellipse, corresponding to the whole breast. GenerateSMF generates such a
measure when it is run with the program option telling it to use a model of the breast edge
to generate a total breast volume.
5884 R Highnam et al

Figure 1. Simplified imaging system: the top-left shows the model used in this section, notably
2D, with parallel x-ray rays and a glandular disk surrounded by a fatty breast edge. The other three
figures illustrate various potential breast density measures (BD1–BD3, please refer to the text for
a fuller description), and the shaded areas show the different breast areas that they would use to
compute breast density.

Although these estimates of breast density are evidently related, the estimated values that
result can be quite different. For example, BD1 excludes the whole layer of fat surrounding
the entire breast and thus tends to lead to higher estimates of breast densities. Similarly, BD2
excludes the breast edge but includes the fatty tissue layer above and below the main breast
area. Finally, BD3 includes the entire breast and so tends to lead to the lowest estimates of
breast density.
As a specific, idealized example, let us assume a 5 cm thick rectangular breast, a 0.5 cm
layer of fat and a 15 cm breast width in our 2D example. Further assume a circle of glandular,
or dense, tissue of radius 2.0 cm and thus an area of 12.6 cm2.

• BD1 would give a breast density of 100 × 12.6/(14 × 4) = 22.5%.

• BD2 would give a breast density of 100 × 12.6/(14 × 5) = 18.0%.
• BD3 would give a breast density of 100 × 12.6/(15 × 5) = 16.8%.
The consequence is that, even before we attempt to account for projection effects, the above
reasoning predicts that the different measures tend to lead to estimates that vary substantially
depending upon the physical model used. Evidently, such variations become more important
when they are applied to small breasts, since in such cases, including/excluding the breast
edge layer becomes measurably more significant.
However, it appears that SMF running without including the breast edge uses the same
effective breast density model as SCC assessed visually (BD2), which can be quite different
from glandular disk density (BD1), and different again from SMF running with a model of the
breast edge (BD3).
Comparing measurements of breast density 5885

4L
The volumetric breast density is:
= 100.0 x (4/3 π r3 )/(4L x 2L x 2L)
= 100.0 x (2.39 L3)/16L3
2L = 15.0%

r = 0.83L 2L

4L
The projected image generated. Using an area based
breast density we get:
= 100.0 x ( π r2)/(4L x 2L)
= 100.0 x (2.16 L2)/8L2
= 27.0%
2L

r = 0.83L

Figure 2. Spherical dense tissue example: the top picture shows a fatty, rectangular breast with
a sphere of dense tissue inside. The bottom picture shows the projection of that sphere as per
mammogram. On the right, we show the workings to generate the volumetric and area-based
breast density estimates.

3. Projection effects

Next, consider the 3D example depicted in figure 2, and, for the moment, ignore the different
definitions of breast density discussed in the previous section. To emphasize that breast density
is a dimensionless parameter (a percentage), the same example is now presented in terms of
an undefined length parameter L. The figure is a representation of a fatty breast, in which there
is a small, spherical dense tissue component. The breast density estimated volumetrically is
15%, whereas when estimated based on the projected area it is 27%, based on BD3.
A second simplified example reinforces this conclusion, as shown in figure 3. In this case
there is a narrow band of dense tissue (0.2L thick). The breast density estimated volumetrically
is 10% whereas the estimated density based on the projected area is 100%. Note that the
differences due to a projection between an area and a volume estimate are greater than those
shown in the previous 3D example.
As additional examples of how area- and volume-based measures do not necessarily
agree, consider that one may add layers of fat at the bottom and/or the top of any breast and
still obtain the same area-based breast density, whereas the volumetric breast densities will
change accordingly. Evidently also, the effects of projection and density estimation can be
compounded.
Of course, perfectly spherical and perfectly horizontal areas of dense tissue would never
be seen in practice! Nevertheless, such examples highlight the need for care in comparing
estimates of breast density, especially when they derive from very different methods.
Note that in the examples presented above, it does not matter whether the projection
angle is CC or MLO; in both cases, the flaws in the area-based techniques are the same.
This observation is consistent with the numerous papers on this subject which report general
agreement between breast densities computed from CC and MLO views.
5886 R Highnam et al

4L

The volumetric breast density according to BD3 is:

2L =100.0 x (0.2L x 2L x 4L) / (4L x 2L x 2L)
= 10%.

0.2 L thick 2L

4L The projected image generated. Using an area based

breast density we get:
= 100.0 x (4L x 2L) / (4L x 2L)
= 100%
2L

Figure 3. Planar dense tissue example: the top picture shows a fatty, rectangular breast with a
plane of glandular tissue inside. The bottom picture shows the projection of that plane as per
mammogram. On the right, we show the volumetric and area-based breast densities.

In this section, we have demonstrated that area-based measures are intrinsically different
from volume-based measures based on simple phantoms. In the remainder of this paper, we
use the above analysis as a guide as we investigate actual differences between the area-based
measure, SCC and the volume measure SMF.

4. Data and software used

The women included in the study are part of the Glasgow alumni cohort (McCarron et al
1999). The cohort was assembled from students at the University of Glasgow (1948–68) who
attended a medical examination at the student health service. Surviving cohort members were
contacted by postal questionnaire in 2001. The project received ethical approval from the
Multi-centre Research Ethics Committee (Scotland).
There were 3556 women in the original Glasgow alumni cohort, of whom 657 women
consented to help in this project. As a result, 4028 images were digitized using a Canon
FS3000 at a resolution of 100 µm with a 8-bit precision. Of the 657 women in the study,
23 had cancer. Whereas in other publications by (some of) the present authors using this
database, the cancer cases were excluded, in this document, all the available data which have
full calibration information are used (namely, kVp, target and filter materials, mAs). This ruled
out 161 images, leaving us with 3867 images. We retained the cancer data because we wanted
to see if the cancers generated more discrepancies between SCC and SMF; see section 3
for more detail.
The images were taken at six sites throughout Scotland over a 15 year period (1988 to
2002). The women included were aged between 40 and 76 with a median age of 57. The SMF
algorithm (and its attendant parameters) was left unchanged for all six sites, even though they
had different equipment and imaging practices.
The software used is SMF version 2.2β from Siemens Molecular Imaging (Oxford, UK).
When SMF is run, it automatically delineates the ‘breast edge’ around the breast image which
is predominantly comprised of fat and has relatively uncompressed breast tissue above it (Pan
et al 2006). The user can then request GenerateSMF to generate a breast volume either
including, or ignoring, that breast edge. If requested to include the breast edge, GenerateSMF
Comparing measurements of breast density 5887

60

50

SMF% Without Breast Edge

40

30

20

10

0
0 10 20 30 40 50 60
SMF% With Breast Edge

Figure 4. SMF% with and without ‘side’ breast edge: the plotted points are the corresponding
SMF% values for the same image when including, or ignoring, the breast edge. The line represents
the identify function. Ignoring the breast edge (the y-axis), produces higher SMF% than when
including the breast edge (the x-axis).

uses a simple half-sphere model with a radius of the breast edge to compute the relevant
volumes. In this way, GenerateSMF is able to generate either a BD2 or BD3 breast density.
For the SCC visual assignment of area density categories, scanned images were displayed
at 300 µm resolution on a flat-panel display system. At this resolution, the images were
approximately the same size as a mammogram film. All the images were displayed to appear
as if viewed on a light box. No other adjustment or image post-processing was applied
during the reading period. We have previously reported on the similarity of density measures
obtained when these assessments are made from the digitized image compared to that from
the original film (Jeffreys et al 2003). Visual density measures were made by one radiologist
experienced in density assessment (RW) who has previously reported good agreement with
other radiologists in assigning visual density categories to mammograms (Atkinson et al 1999,
2004).

5. Results

We first consider the effect of the breast density definition on the results, before investigating
our database for those cases for which SCC and SMF% do not match. We then attempt to
account for those discrepancies in terms of the breast density definitions and projection effects.

5.1. The effect of breast density definition

5.1.1. Effects of including and excluding the breast edge (BD2, BD3). We begin by
considering the differences between a BD2 breast density (ignoring the breast edge around
the sides) and a BD3 breast density which includes all the fatty side breast edge. Figure 4
shows the two sets of values presented as a scatter plot. As detailed in the earlier section,
ignoring the breast edge in the SMF% calculation (‘Without Breast Edge’ on the graph) leads
to substantially higher estimates of breast density: an additional 2–3% at SMF% = 10 (on
5888 R Highnam et al

45
SMF% with Breast Edge
40
SMF% without Breast Edge

35

30
SMF%
25

20

15

10

0
0 1 2 3 4 5 6 7
SCC

Figure 5. SCC versus SMF% (with and without ‘side’ breast edge): the squares represent SMF%
with a breast edge estimate and the triangles are without. The values shown are the medians
along with the first and third quartiles. Both SMF estimates have a similar relationship with SCC
although the absolute values change markedly.

the x-axis) and approximately an additional 10% at SMF% = 40%. Thus, it is critical to
understand which definition is being used when comparing breast densities.
The greater effect at higher breast densities appears to be due to denser breasts tending
to be smaller, since the majority of the difference in the breast volumes between women
corresponds to an additional fat content. For this reason, the breast edge has more of an effect
when included.
A similar relationship between SMF% and SCC exists with and without the breast edge,
but the absolute range of the SMF% values for each category rises, as shown in figure 5. This
indicates that we can compare SCC to either, but also that if we attempt to categorize future
SMF% results into SCC categories we need to be certain which breast density definition is
being used; indeed, it might be better to output just one result to avoid confusion.

5.1.2. SMF% and glandular disk density (BD2, BD1). Dance et al (2000) estimated the
average glandular disk density for a range of breast thicknesses, by comparing the exposure
factors selected by the automatic exposure control (AEC) of the x-ray sets for compressed
breasts with those selected by the AEC for tissue equivalent materials of various thicknesses
and compositions.
Figure 6 shows Dance’s BD1 measure results (in the solid line with square markers) for
women of age 50–64 versus the breast thickness. We also show their results with a 0.5 cm
fatty breast edge added back to make it more like a BD2 measure and thus roughly comparable
to SMF% (dotted line, triangle markers). The grey background dots are the SMF% (without
side breast edge) distribution and the solid line with no markers is a cubic least-squares fit to
those data.
Without the addition of a breast edge estimate, the average glandular disk density reported
by Dance et al is very different from the SMF% in our database, especially at average and lower
breast thicknesses (<6 cm). With the correction, there is an improved relationship although
Comparing measurements of breast density 5889

140
SMF% versus Breast Thickness
120 Glandular Disk% versus Breast Thickness
Glandular Disk% Converted to SMF% versus Breast Thickness
100

80
SMF %

60

40

20

0
0 2 4 6 8 10
Breast Thickness, H (cm)

Figure 6. SMF% versus glandular disk density: the solid line with square markers is from Dance
et al (2000) showing average glandular disk density, the dashed line with triangle markers shows
the glandular disk density converted to include a fatty breast edge, the grey dots represent the
SMF% distribution for each breast thickness and the solid line with no markers is a cubic best-fit
through those dots. The relationship between glandular disk density and SMF% is improved with
the conversion to make the breast density definitions more consistent.

still substantial differences, some of which are to be expected, since we are comparing patient-
specific densities to average glandular disk density derived from phantoms.

5.2. Discrepancies between SCC and SMF%

Next, we investigated those patients for whom the estimate of SMF% (without the breast edge)
was substantially different to the SCC readings for more than one image. More precisely,
we analysed the data for those patients who had more than one image which was reported as
having
• SMF% < 25% and SCC > 4
or
• SMF% > 30% and SCC < 5.
These values were chosen on the basis of the values shown in figure 5 to highlight
obviously different values.
We found that, of 657 cases, 80 had two or more mammograms that matched one of these
criteria (12%).
We analysed each set of patient images by utilizing DICOM conversion and patient
consolidation software (EasyScan) kindly provided by BreastScreen Aotearoa (New Zealand),
and we identified the following trends amongst the data with large differences between SMF%
and SCC.
• Small breast thickness, under 3 cm thick, this invariably leads to high SMF%, but not
necessarily high SCC.
• Large breast thickness, over 6.5 cm thick, invariably leads to low SMF%, but not
necessarily low SCC.
5890 R Highnam et al

Figure 7. SMF%, SCC discrepancy, small breast thickness: this breast had a very high SMF%
of 44.0%, whilst the SCC category was just 3. The recorded breast thickness was 2.3 cm and the
estimated thickness was 1.6 cm, indicating that this is a very thin breast.

• Dense block of tissue, a very dense area in an otherwise very fatty breast, this makes an
SCC assessment difficult, but not SMF%.
• Patchy dense breasts, SMF% quantifies each of the dense areas, whereas SCC visual
assessment tends to combine them into one big area.
• High cloud appearance as in the planar example given above illustrates, where the breast
appears covered with a thin layer of dense tissue.
• Human error in the visual reading of SCC.
• Noise issues with SMF, in particular the segmentation of the breast failing either fully or
partially leading to erroneous areas and breast thicknesses (Highnam et al 2006).
• Underexposed images, making the visual reading very hard.
Some actual examples from our database of these discrepancies are shown in figures 7–10.
After studying all 80 for which there are significant discrepancies, table 2 shows the
distribution of explanations, as described in the previous section. The human error/noise
issues with SMF and underexposed images can evidently be improved upon to get better
correspondence. More importantly, however, the other discrepancies are much more difficult
to reconcile, since they appear to reflect fundamental differences between area and volumetric
techniques.

5.3. Making SMF% more similar to SCC

Some of the effects noted earlier are directly related to the effects of projection and the
automatic exposure control. For example, adding homogeneous layers of fat to the top and
bottom of a breast would (in principle) produce no change in the image and so no change in
Comparing measurements of breast density 5891

Figure 8. SMF%, SCC discrepancy, patchy dense breasts: this breast had a moderate SMF% of
25.0%, whilst the SCC category was a high 5. Visually, the breast is very patchy with some large
volumes of fat which appear not to be considered in the visual assessments.
Table 2. Discrepancy distribution: cases where SCC and SMF differ by a wide margin were
investigated and categorized into various discrepancies for all 80 patients.

Percentage of
Cause of discrepancy total discrepancies
Small breast thickness 13.75%
Large breast thickness 13.75%
Dense block of tissue 11.25%
Patchy dense breasts 21.25%
High cloud appearance 17.5%
Human error 2.5%
Noise issues with SMF 17.5%
Underexposed images 2.5%

Total 100.0%

SCC; but it would lead to major changes in SMF%. To investigate this further, we looked at
SCC versus SMF% for different ranges of the estimated breast thickness.
Figure 11 shows the relationship for those images with—estimated breast thicknesses
between 2 and 3 cm (186 images); the entire database; and those images with estimated breast
thicknesses between 6 and 7 cm (512) images.
For each breast thickness range, there appears to be a similar relationship between SMF%
and SCC, albeit with the absolute range of SMF% values shifted. This suggests that one
could, where it deemed necessary, achieve even better agreement between SMF% and SCC by
categorizing the images based upon SMF% and the corresponding breast thickness estimation.
For example, if we were told the breast had SMF% = 35% and a breast thickness of 2.5 cm,
we would not use the overall results (circles) to say this would correspond to SCC 6; rather,
we would use the 2–3 cm breast thickness information to say that this more likely corresponds
to SCC 5.
5892 R Highnam et al

Figure 9. SMF%, SCC discrepancy, high cloud appearance: this breast had a moderately low
SMF% of 20.0%, whilst the SCC category was 4. Visually, the whole projected area is covered by
glandular tissue, but it looks very thin.

Figure 10. SMF%, SCC discrepancy, dense block of tissue: this breast had a moderate SMF% of
24.0%, whilst the SCC category was 5. Visually, the breast includes a large, very dense area and an
equally large fatty area which appears not to be taken into consideration in the visual assessment.

From the previous section, such a re-categorization could remove up to 27.5% of the total
number of discrepancies by removing those discrepancies due to the small and large breast
thicknesses. However, even if we can remove them, and resolve the 22.5% due to human
error, SMF noise issues and underexposure that still leaves 50% of the discrepancies or 6% of
all patients where we might still expect SCC and SMF to differ.
Comparing measurements of breast density 5893

50.0

40.0

30.0
SMF %

20.0
H 6 - 7cm
All Images
10.0
H 2 - 3cm

0.0
0 1 2 3 4 5 6
SCC

Figure 11. SMF% versus SCC for various ranges of breast thicknesses: we use squares to show
the relationship for breast thicknesses between 6 and 7 cm thick, triangles for breast thicknesses
between 2 and 3 cm thick and circles for all the images. The points show the median value and the
lines show first and third quartiles.

6. Discussion

This study included images from 657 women, 23 of whom had cancer. Of those, just three
showed up as having a large discrepancy between the SMF and SCC readings: one ‘dense
block’, one ‘high cloud’ and one underexposed image. We may safely conclude that including
the cancers does not significantly impact the results.
Version 2.2β of the GenerateSMF software was tuned specifically to mammograms
digitized on a Canon scanner. It may be that the discrepancy distribution would be very
different for images scanned on other scanners, but our experience leads us confidently to
predict similar broad patterns. Note in particular that automated breast density measures are
considerably easier to implement on digital systems and so some of the discrepancies noted
in table 2 can be expected to be largely irrelevant for digital images.
Instead of comparing SMF to SCC, and whilst data are still being collected to examine the
relationship between SMF and breast cancer risk directly, we could compare SMF to the breast
density measures being derived from 3D imaging techniques such as CT/MRI or even digital
breast tomosynthesis. Each technique has its own issues and tuning requirements, but there
should be a correlation if not exact equivalence. This is the approach taken by van Engeland
et al (2006).

7. Conclusions

In this paper, we have explored why the results from a volumetric-based breast density measure
(SMF) should not always be expected to correspond closely to those from an area-based
technique (SCC). Reasons include different physical models of breast density and projection
effects.
We have suggested how we can make SMF more compatible with SCC by modifying
the categorizations based upon the breast thickness. However, even with that modification,
as well as the removal of obvious error cases, approximately 50% of the discrepancies are
likely to remain. This suggests that for around 6% of patients, we may expect to see large
5894 R Highnam et al

differences between SMF and SCC. Thus, even though SMF and SCC correlate well, we
cannot assume that SMF (or other volumetric methods) will have similar breast cancer risk
predication abilities. In fact, given the obvious flaws in the area-based methods, we might
expect volumetric methods to perform better.
Of course, the ultimate test for all the breast density measures requires an understanding
of how they are to be used and in that context most measures should not be forced to correlate
perfectly with each, but should be measured against their ability to predict breast cancer risk.

Acknowledgments

We are grateful to the women who participated in this study for allowing access to their
mammograms, Pat Forrest who performed the digitization and for the assistance of all
the Scottish Breast Screening Centres. This work was undertaken while Dr Jeffreys (née
Okasha) was employed at the University of Bristol. We are grateful for the financial support
provided by the Breast Cancer Campaign, Breast Cancer Research Trust and World Cancer
Research Fund International. The Centre for Public Health Research, Massey University
is supported by a programme grant from the Health Research Council. Lastly, we are
very grateful to BreastScreen Aotearoa (New Zealand) for providing the EasyScan image
conversion and visualization program for the use in this research. Michael Brady and Ruth
Warren acknowledge support from the EU Mammogrid project, and from the EPSRC-MRC
Interdisciplinary Research Consortium MIAS.

References

Atkinson C, Warren R, Bingham S and Day N 1999 Mammographic patterns as a predictive biomarker of breast
cancer risk: effect of tamoxifen Cancer Epidemiol. Biomarkers Prev. 8 863–6
Atkinson C, Warren R, Sala E, Dowsett M, Dunning A, Healey C, Runswick S, Day N and Bingham S 2004 Red
clover-derived isoflavones and mammographic breast density: a double-blind, randomized, placebo-controlled
trial Breast Cancer Res. 6 R170–9
Boyd N, Lockwood G, Byng J, Tritchler D and Yaffe M 1998 Mammographic densities and breast cancer risk Cancer
Epidemiol. Biomarkers Prev. 7 1133–44
Byng J, Boyd N, Fishell E, Jong R and Yaffe M 1994 The quantitative analysis of mammographic densities Phys.
Med. Biol. 39 1629–38
Dance D, Skinner C, Young K, Beckett J and Kotre C 2000 Additional factors for the estimation of mean glandular
breast dose using the UK mammography dosimetry protocol Phys. Med. Biol. 45 3225–40
Heine J J and Malhotra P 2002 Mammographic tissue, breast cancer risk, serial image analysis, and digital
mammography: part 1. Tissue and related risk factors Acad. Radiol. 9 298–316
Highnam R and Brady M 1999 Mammographic Image Analysis (Dordrecht: Kluwer)
Highnam R, Pan X, Warren R, Jeffreys M, Davey Smith G and Brady M 2006 Breast composition measurements
using retrospective standard mammogram form (SMF) Phys. Med. Biol. 51 2695–713
Jeffreys M, Warren R, Davey Smith G and Gunnell D 2003 Breast density: agreement of measures from film and
digital images Br. J. Radiol. 76 561–3
Jeffreys M, Warren R, Highnam R and Davey Smith G 2006 Initial experience of using an automated volumetric
measure of breast density: the standard mammogram form (SMF) Br. J. Radiol. 79 378–82
Karssemeijer N 1998 Automated classification of parenchymal patterns in mammograms Phys. Med. Biol. 43 365–78
McCarron P, Davey Smith G, Okasha M and McEwen J 1999 Life course exposure and later disease: a follow-up
study based on medical examinations carried out in Glasgow University (1948–68) Public Health 113 265–71
Pan X, Brady M, Highnam R and Declerck J 2006 The use of multi-scale monogenic signal on structure orientation
identification and segmentation Proc. of the IWDM 2006 ed S Astley et al (Berlin: Springer) pp 601–8
Sivaramakrishna R, Obuchowski N A, Chilcote W A and Power K A 2001 Automatic segmentation of mammographic
density Acad. Radiol. 8 250–6
Comparing measurements of breast density 5895

van Engeland S, Snoeren P, Huisman H and Karssemeijer B 2006 Volumetric breast density estimation from full-field
digital mammograms IEEE Med. Imaging 25 273–82
Zhou C, Chan H-P, Petrick N, Helvie M, Goodsitt M, Sahiner B and Hadjiiski M 2001 Computerized image analysis:
estimation of breast density on mammograms Med. Phys. 28 1056–69
Zwiggelaar R and Denton E R E 2006 Texture based segmentation Proc. of the IWDM 2006 ed S Astley et al (Berlin:
Springer) pp 433–40