Home Search Collections Journals About Contact us My IOPscience

Breast composition measurements using retrospective standard mammogram form (SMF)

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

2006 Phys. Med. Biol. 51 2695

(http://iopscience.iop.org/0031-9155/51/11/001)

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

Download details:

IP Address: 142.132.1.147
This content was downloaded on 26/08/2015 at 06:37

Please note that terms and conditions apply.

INSTITUTE OF PHYSICS PUBLISHING PHYSICS IN MEDICINE AND BIOLOGY

Phys. Med. Biol. 51 (2006) 2695–2713 doi:10.1088/0031-9155/51/11/001

Breast composition measurements using retrospective

standard mammogram form (SMF)
R Highnam1,6 , X Pan1, R Warren2, M Jeffreys3, G Davey Smith4
and M Brady5
1 Siemens Molecular Imaging Ltd, Hythe Bridge Street, Oxford, UK
2 Addenbrooke’s Hospital, Cambridge, UK
3 Massey University, Wellington, New Zealand
4 Clinical Epidemiology, University of Bristol, UK
5 Engineering Science, University of Oxford, UK

E-mail: rphighnam@aol.com

Received 5 January 2006, in final form 30 March 2006

Published 9 May 2006
Online at stacks.iop.org/PMB/51/2695

Abstract
The standard mammogram form (SMF) representation of an x-ray mammogram
is a standardized, quantitative representation of the breast from which the
volume of non-fat tissue and breast density can be easily estimated, both of
which are of significant interest in determining breast cancer risk. Previous
theoretical analysis of SMF had suggested that a complete and substantial
set of calibration data (such as mAs and kVp) would be needed to generate
realistic breast composition measures and yet there are many interesting trials
that have retrospectively collected images with no calibration data. The main
contribution of this paper is to revisit our previous theoretical analysis of SMF
with respect to errors in the calibration data and to show how and why that
theoretical analysis did not match the results from the practical implementations
of SMF. In particular, we show how by estimating breast thickness for every
image we are, effectively, compensating for any errors in the calibration data.
To illustrate our findings, the current implementation of SMF (version 2.2β)
was run over 4028 digitized film–screen mammograms taken from six sites
over the years 1988–2002 with and without using the known calibration data.
Results show that the SMF implementation running without any calibration data
at all generates results which display a strong relationship with when running
with a complete set of calibration data, and, most importantly, to an expert’s
visual assessment of breast composition using established techniques. SMF
shows considerable promise in being of major use in large epidemiological
studies related to breast cancer which require the automated analysis of large
numbers of films from many years previously where little or no calibration data
is available.

6 Address for correspondence: #9, 2 Hood Street, Mt Victoria, Wellington, New Zealand.

0031-9155/06/112695+19$30.00 © 2006 IOP Publishing Ltd Printed in the UK 2695

2696 R Highnam et al

1. Introduction

There is increasing acceptance of a causal relationship between measures of breast tissue

composition and increased risk of breast cancer (Boyd et al 1998). This has encouraged the
conjecture that such measures of breast tissue composition may serve as biomarkers for breast
cancer risk (Heine et al 2002).
To date, the composition measure most studied has been the percentage of dense breast
tissue in the breast. This is typically estimated as the ratio of the area of the mammogram judged
to be of a dense nature and the area of the whole (projected) breast. These areas may either
be estimated visually or using a computer-based thresholding technique (Boyd et al 1998).
The resulting ratio can be made into and presented as a percentage; alternatively, it may be
classified into one of a small number of categories, such as the six category classification (SCC,
Byng et al (1994)), or the BI-RADS classification (ACR 1998). Less frequently nowadays,
categorizations such as the Wolfe system (Wolfe 1976), which map qualitative, perceptual
descriptions of parenchymal patterns into composition measures are used.
Despite their reported success in predicting breast cancer risk, there are widely recognized
shortcomings with using measures of breast density computed from mammographic images,
including the following: the projective nature of mammography; the confounding factor in such
measurements that stems from the substantial variation in the image that results from any given
breast by changing the imaging parameters (kVp, mAs, etc); the subjectivity introduced by
human interaction; the requirement of using trained (and expensive) human readers; the belief
that it is absolute volume of dense tissue that might be more important than the percentage.
All these reasons have inspired increasing interest in automated volume-based measures
(Pawluczyk et al 2003, Hufton et al 2006) rather than area-based measures.
This paper reports findings concerning the estimation of volumetric measures of breast
composition, based on an estimate of the amount of non-fat tissue in the breast, using standard
mammogram form (SMF) (Highnam and Brady 1996, 1999).
To understand SMF, let us first consider how it might be possible to ‘standardize’ a
mammogram’s appearance in such a way that it removes variations due to image acquisition
and brings to the fore quantitative information about the breast tissue. Since a mammogram is
a projective two-dimensional (2D) view of the compressed three-dimensional (3D) breast, the
optimal representation of the breast anatomy that can possibly be gleaned from a mammogram,
at least from a single view, would comprise measurements of the amounts of different tissue
types in the relatively tall thin volume of tissue above each pixel. What tissue types might
be distinguished in this way? Highnam and Brady (1996, 1999), citing the work of Yaffe,
Dance and others, suggest a taxonomy comprising three tissue classes: calcification, adipose
tissue and ‘interesting’ tissue (stroma, tumour, etc). They note that the volume of calcified
tissue is, in almost all cases, tiny relative to the other two and thus proposed the (quantitative)
representation of the thickness of ‘interesting’ tissue hint (x, y) where (x, y) is an image
location, and where hint (x, y) is measured in centimetres, the maximum value being denoted
by H , the distance between the compression plates or breast thickness.
It is one thing to postulate such a mathematical representation of projected anatomy; it is
entirely another thing to develop a computational procedure to generate it from a mammogram.
The presentation of an algorithm to generate hint (x, y) is the main focus of the Highnam and
Brady monograph (1999). In their original scheme, the algorithm requires as inputs a digitized
mammogram, as well as certain parametric information: tube kVp, exposure mAs and H .
Since the latter is either not recorded on many film–screen systems, or is done so inaccurately,
Highnam and Brady subsequently developed a method to estimate H from the mammographic
image. Of course, in the six years that have elapsed since Highnam and Brady’s monograph
Breast composition using retrospective SMF 2697

was published, several refinements to the approach have been presented, and continue to be
developed. Nonetheless, it is reasonable to claim that a representation such as hint (x, y) can
be computed, with relatively small error bars at each image location. Nevertheless, we
continue to work to improve both the process of generating the hint (x, y) representation and of
the representation itself. Note that the numerical sensitivity of the hint (x, y) representation to
minor errors in parameter values was analysed by Highnam and Brady (1999) using a Taylor’s
series analysis. Though that analysis provides some interesting insights into the stability of the
representation, it does assume that the parameter values can be varied independently. While
this is true from the standpoint of physics, in practice the implementation details mean that
the relationships between parameter values constrain the results of varying any given one.
This leads directly to the main observation in this paper and, as we note below, enables the
retrospective application of hint (x, y).
The SMF or hint (x, y) representation enables computation of the amount of dense breast
tissue, either as an absolute volume (SMF volume) equal to the sum of the hint (x, y) values
over all image points lying within the breast region, or as the ratio of the volume of dense
tissue divided by the total volume of the breast (SMF%), which may naively be computed as
H times the area of the contributing breast region (it is naive because this simple product does
not take account of breast edge effects). Even though the SMF model of the breast includes
all dense tissues into its breast composition measures, it has been shown that when most of the
calibration data are available, SMF% correlates strongly with an expert’s visual assessment of
breast density using the SCC (Jeffreys et al 2006), and both SMF measures referred to above
show a small but significant association with age as well as correlation with important known
breast cancer risk measures such as body mass index (Jeffreys et al 2003b).
To date, each succeeding version of the GenerateSMF software has assumed the
availability of certain calibration parameters, most notably tube voltage and exposure time.
While this is an increasingly reasonable assumption, consequent upon the introduction of
digital mammography, it appears to rule out the retrospective application of SMF to data for
which the requisite calibration data were either not recorded, or are lost, which is the case (for
example) for many studies looking at the effects of drugs over time where they seek to include
mammograms from many years prior.
The main contribution of this paper is to revisit our previous theoretical analysis of
SMF with respect to errors in the calibration data and show how and why that theoretical
analysis did not match the results from the practical implementation of SMF. Our starting
point has been the observation that the current implementation demonstrates a remarkable
ability to overcome errors in the calibration data. Indeed, it even works surprisingly well
if no calibration parameter values are entered. We call the resulting computational process
‘calibration parameter compensation’, abbreviated to CPC. Critically, CPC enables SMF to
be applied retrospectively.
We begin with a brief overview of SMF, then with reference to estimating breast
thickness H show how our previous theoretical analysis of SMF does not correspond to the
implementations of GenerateSMF. This leads us to propose the constraint propagation process
we call calibration parameter compensation (CPC), and we demonstrate its ability to constrain
the transfer functions relevant to mammography. We then move to the problem of retrospective
estimation of hint (x, y) and show results from a database of 4028 film–screen mammograms.

2. Standard mammogram form (SMF)

A complete technical description of hint (x, y)/SMF is presented in Highnam and Brady 1999.
We sketch the main features here, in order to draw out the salient points for CPC. We focus
2698 R Highnam et al

on film–screen systems but the model is easily adapted to raw or ‘for processing’ digital
images.
The SMF model for primary energy imparted to the film–screen in the area of
the intensifying screen corresponding to a pixel (x, y) is summarized mathematically in
equation (1):

Emax
E (x, y) = φ(Vt , x, y)Ap ts
imp
N0rel (Vt , ε)S(ε)G(ε) e−µluc (ε)hplate e−hµ(ε) ε dε (1)
0
where φ is the photon flux, Vt is the tube voltage, Ap is the area of the pixel, ts is the time
of exposure, Emax is the maximum photon energy, N0rel is the relative number of incident
x-ray photons at the specific energy ε, S is the screen absorption, G is the grid transmission,
µluc is the linear attenuation coefficient of Lucite, hplate is the thickness of the Lucite breast
compression plate and hµ is the linear x-ray attenuation coefficient due to breast tissues in the
column of tissue above (x, y).
While some of these values are fixed for the particular x-ray machine and do not vary,
for example hplate , and others are specific to the particular detector system, for example S, the
values for several of the parameters are specific to the particular image, for example, the time
of exposure ts . Some of the required parameter values can be set from the data displayed by
the x-ray system when the mammogram is taken. These include mAs, tube voltage and filter
and target information. Yet others may be set simply by knowing the name of the component
and using values built into the software (for example, ‘Philips Grid’). Finally, as a last resort,
if the value is still not known then either the most commonly occurring component type is
assumed, or we may attempt to estimate the value from the image. For example, as we discuss
in the next section, we have shown that we can estimate breast thickness using the information
around the breast edge.
Note that equation (1) refers only to the primary component. It does not refer to the scatter
contribution, which adds to the total energy imparted to each pixel. Subsequent GenerateSMF
processes have computed this as first described in Highnam et al (1994). Likewise,
equation (1) simplifies the physics in that it also does not show reciprocal law failure, anode
heel correction, compression plate slant correction, all of which are all detailed in Highnam
and Brady (1999).
As we noted in the introduction, to develop the hint (x, y) representation we assume that
we can ignore calcification, though we have demonstrated (Yam et al 1999) that we can use the
significantly higher attenuation coefficient of calcium to detect calcifications. It follows that
the total linear attenuation coefficient for any cone of breast tissues above a pixel (x, y) can be
written as shown in equation (2) where hint (x, y) is the thickness of ‘interesting’ tissue (hence
the name of the representation), hfat (x, y) is the thickness of fat and the linear attenuation
coefficients are the averages for each class as described by Johns and Yaffe (1987):
hµ(ξ ) = hint (x, y)µint (ξ ) + hfat (x, y)µfat (ξ ). (2)
If the compressed breast thickness H is known for this cone of tissue, then we can rewrite
equation (2) to include just one unknown (see the following equation):
hµ(ξ ) = hint (x, y)(µint (ξ ) − µfat (ξ )) + H µfat (ξ ). (3)
Observe that if the attenuation coefficients are known with reasonable accuracy, and if we
can estimate H , then there is only one unknown in this equation, namely hint (x, y); we return
to this below.
In the GenerateSMF implementations to date, the energy imparted is mapped to a film
density by an assumed sigmoid-shaped film–screen characteristic curve, equation (4), where
the origin is the point of symmetry of the curve, a and b are the sigmoid parameters and the
Breast composition using retrospective SMF 2699

energy has been made relative to the log energy at the origin. The reader is referred to books
on ‘serpentine curves’ for a full derivation. These values can often be determined from the
quality control books at each breast care centre for any particular day:
imp
b2 Log Erel (x, y)
D(x, y) = Dorigin + imp . (4)
Log2 Erel (x, y) + a 2
Finally, we assume a linear film digitizer with known linear parameters (α, β) and thus
represented by the following equation:
P (x, y) = αD(x, y) + β. (5)
Highnam and Brady (1999) demonstrate that if sufficient calibration data are available, it
is possible to map from pixel value to film density and then back to energy imparted (removing
scatter and other spatial variant effects) to get the left-hand side of equation (1). Because of
the fundamental SMF assumption, the right-hand side of equation (1) has just one unknown,
hint (x, y), which we can now vary until the right-hand side matches the value generated from
the image for the left. That, in essence, is how GenerateSMF processes work.

3. Estimating H

We noted in the previous section that current GenerateSMF implementations estimate breast
thickness directly from the image. They do so by assuming the existence of a fatty breast edge
and therefore a set of smooth, homogeneous curves which start at the breast skin line and end
at the edge of the uniformly compressed breast region. This last curve is assumed to have
hint (x, y) = 0 and hfat (x, y) = H and we term this the hint (x, y) = 0 curve.
In earlier implementations of GenerateSMF, the hint (x, y) = 0 curve was estimated as
part of the generation process so that we would compute the hint (x, y) values for a certain H ,
determine where the hint (x, y) = 0 curve was and how smooth it was, and then iterate until
the smoothness criteria were met. This worked surprisingly well, giving estimates to within
2.2 mm for CC and 4.4 mm for MLO images as reported in Highnam et al (1998).
Figure 1 shows the inner and outer most smooth, homogeneous curves for a typical breast
as white, dotted lines. When the pixel value rises sufficiently, the curves quickly become less
smooth, as shown by the continuous line, and we have used a variety of measures to quantify the
departure from smoothness (e.g. fractal measures, Lipschitz). As stated previously, we assume
the hint (x, y) = 0 curve to be the innermost smooth curve. In the current implementation of
SMF the segmentation algorithms and the method of finding the innermost smooth curve are
greatly improved in terms of both speed and accuracy by estimating the hint (x, y) = 0 curve
directly from the image itself without needing to generate SMF each time.
By finding the hint (x, y) = 0 curve, we are effectively computing ‘ground truth’
information from the image. We are asserting, in effect, that the average pixel value under the
innermost smooth curve should be mapped to an average hint value of 0.0.
Our previous Taylor-series-based theoretical analysis of hint (x, y)/SMF (see, for example,
Highnam et al (1996), Highnam and Brady (1999)) led to the conclusions that both SMF% and
SMF volume were highly susceptible to errors in the calibration data. As a specific example, a
change in breast thickness H → H ± δH of just 0.1 cm could change SMF% by approximately
5%. Evidently, were this to be the case in practice, it would constitute a major obstacle to
using GenerateSMF in calculations of SMF%/SMF volume, hence of assessing breast cancer
risk.
Fortunately, documented evidence about the errors in breast thickness readings from most
mammography machines (Burch and Law 1995) and the lack of breast thickness readings for
2700 R Highnam et al

Figure 1. Breast edge: the breast edge is mostly fat and displays mammographically as a set
of smooth curves with the same pixel value (the dotted lines); as the pixel value rises, the curve
becomes rough and the breast edge has finished (the continuous line).

many mammograms of interest led to a design in which GenerateSMF always estimates breast
thickness directly from the image and from known calibration data using the method sketched
above.
The upshot is that, though subtle, this apparent implementation detail has a profound
consequence: errors in the calibration data are used in the estimation of breast thickness, so
the calibration parameters are not independent, so that the Taylor’s series analysis is massively
overly pessimistic. Instead, we realize that the GenerateSMF process embodies a set of mutual
constraints between parameter values, and these have the welcome property of automatically
correcting for errors in the calibration data by the use of ‘ground truth’ from the image. This
constraint propagation process we call calibration parameter compensation (CPC).

4. Calibration parameter compensation

We follow Tromans (2006) in regarding GenerateSMF as a transfer function from image

brightness values P (x, y) to interesting tissue values hint (x, y). Figure 2 shows an example
transfer function generated from equation (1) for a 12-bit image. In this particular case, the
film–screen characteristic curve used, as described by equation (4), has a = 0.84, b = 1.67,
imp
and the point of symmetry is at log Eorigin = 5.75 and, for reference, a pixel value of 1500
represents a film density of 2.57.
Suppose that Ph int 0 is the average pixel brightness value along the image curve determined
to correspond to hint = 0 as in the previous section. No matter what are the errors in the
calibration data, we have now fixed one point of the SMF transfer function. In figure 2,
a pixel value of 1500 is fixed to map to hint = 0. Thus, even if there are errors in the
calibration data, that pixel value will still be mapped to that thickness of interesting tissue
by the SMF implementation. In short, when we estimate breast thickness, that thickness is
automatically adjusted to compensate for calibration data errors so that the transfer function
remains reasonably correct.
Breast composition using retrospective SMF 2701

6.00

5.00

4.00

hint (cm)
3.00

2.00

1.00

0.00
0 500 1000 1500 2000 2500 3000 3500 4000 4500
Pixel Value (P)

Figure 2. Typical SMF transfer function: a typical SMF transfer function between pixel value in
the original image and h value in the SMF representation for a breast thickness of 5.5 cm, 25 kVp,
Mo/Mo target/filter and 190 mAs. The graph is linear in nature for the fatty areas of the breast but
non-linear in the denser parts suggesting that conversion to SMF before further processing, such
as computer-aided detection, will have greater impact in the denser parts of the breast.

7.00

6.00

5.00

40mAs
h int (cm)

4.00
61mAs, True Values

3.00 100mAs

40mAs, CPC: H=4.2cm

2.00
100mAs, CPC: H=5.95cm

1.00

0.00
0 1000 2000 3000 4000 5000
Pixel Value (P)

Figure 3. SMF with incorrect mAs: transfer function with and without CPC and with incorrect
mAs. The ‘true’ values are 29 kVp, Mo/Mo filter/target, H = 5.0 cm and 61 mAs; see the wide
black curve. The two continue curves are for incorrect mAs (40, 100) but with using the true H =
5.0 cm and the two dotted curves are for when CPC is used with those incorrect mAs. Note the
breast thicknesses estimated in the legend.

Of course, if we have only one piece of ‘ground truth’, errors in the calibration data that
affect the shape of the transfer function (such as tube voltage) may still have significant impact
and the fact that the breast thickness is being ‘adjusted’ to absorb errors raises concerns about
the accuracy of the quantitative measures such as SMF%, which use the breast thickness to
estimate breast volume as part of the process. With this in mind, we next consider a number
of transfer functions and the impact of deliberate errors in the key calibration data.
First, suppose that an incorrect mAs is used; see figure 3. This shows in a wide black
line the ‘true’ transfer function with the correct values of 61 mAs and H = 5.0 cm. It also
shows the transfer functions (in thin continuous black with markers) for the case when the
mAs is incorrect (40 and 100 mAs) but breast thickness estimation (CPC) is not used. Finally,
2702 R Highnam et al

Table 1. Incorrect mAs and breast composition. The mAs used is shown in the left column and
then we have the breast thickness (H) and breast composition measures, Av hint is closely related
to SMF volume and Av hint /H is closely related to SMF%.

Without CPC With CPC

mAs H Av hint Av hint/H H Av hint Av hint/H
40 5.0 cm 0.2 cm 4% 4.2 cm 1.13 cm 27%
61 5.0 cm 1.25 cm 25% 5.0 cm 1.25 cm 25%
100 5.0 cm 2.25 cm 45% 5.95 cm 1.35 cm 23%

we show (dotted lines with markers) the resulting transfer function when CPC is used. The
legend notes the breast thicknesses found.
Clearly, with CPC the transfer function is much closer to the ‘true’ case. For example,
for a pixel value of 3000 the true h value is 2.0 cm, with CPC we get 1.95 cm, without CPC
we get 1.0 cm. Note that this analysis also illustrates the sensitivity of the SMF algorithm to
errors in mAs when CPC is not used.
We note that the previous theoretical analysis (Highnam and Brady 1999) suggested that
errors in mAs would translate the h values up or down when acting independently of all other
errors. Similarly, this analysis suggested that errors in breast thickness would also translate
the values up or down, rather than be multiplicative or non-linear, and thus it is perhaps no
surprise that the transfer functions with CPC are almost identical no matter what the mAs is.
Consider next the quantitative values of breast composition for our example pseudo-
breast. Let us assume that the average pixel value in the image is 2500, which equates to a film
density of 1.5 and is a reasonable target film density for many automatic exposure controls.
Let us further assume that those average values are a fair representation of the overall breast
composition in our pseudo-breast.
From figure 3, the h value at a pixel value of 2500 in the ‘true’ case is 1.25 cm. Since we
are assuming that this value is a fair representation of the whole breast we are also assuming
that it is directly proportional to SMF volume which is the sum of the h values over the whole
breast. In table 1 we use Av hint to be this representative hint and therefore SMF volume value.
Further, dividing Av hint by the known breast thickness (5 cm in this case) provides us
with an estimate of SMF% (breast density), which in this case is 25%. See table 1 for further
values for this example.
Table 1 shows that with CPC, the changes in the average h are much smaller than without
CPC. For example, the true value is when we have 61 mAs and we get 1.25 cm for Av hint,
without CPC and with an mAs of 100 we get 2.25 cm, whilst with CPC we get 1.35 cm.
Likewise, CPC shows a remarkable ability to compensate for errors especially in the
pseudo-SMF% values. For example, the true values are when we have 61 mAs and we get
25% for Av hint/H. Without CPC, and with an mAs of 100, we would get 45% whilst with
CPC we get 23%, a dramatic improvement.
Now consider using an incorrect tube voltage, which was previously shown to have
a non-linear effect on hint values when the calibration data values were considered to be
independently changeable. Figure 4 shows the transfer functions with the true value of 29 kVp,
and for erroneous values of 25 and 31 kVp, again with and without CPC. Once more, CPC
demonstrates a remarkable ability to preserve the correct h values, especially in the lower pixel
values (1500–2200) which correspond to fattier tissue.
Table 2 shows the breast composition measures, again using a pixel value of 2500 as the
average for our pseudo-breast.
Breast composition using retrospective SMF 2703

6.00

5.00

4.00 25kVp
29kVp,TrueValues

h int (cm) 3.00

31kVp
25kVp,CPC:H=3.65cm
31kVp,CPC: H=5.6cm
2.00

1.00

0.00
0 500 1000 1500 2000 2500 3000 3500 4000 4500

Pixel Value (P)

Figure 4. SMF with incorrect kVp: the true values are 29 kVp, H = 5.0 cm, 61 mAs and the
transfer function is shown with a wide black line. The two continuous thin black lines with markers
show different kVp without CPC, and the two dotted lines show the results with CPC; note the
estimated breast thicknesses in the caption.

Table 2. Incorrect kVp and breast composition. The kVp used is shown in the left column and the
results shown with and without CPC.
Without CPC With CPC
kVp H Av hint Av hint H H Av hint Av hint/H
25 5.0 cm – – 3.65 cm 1.00 cm 27%
29 5.0 cm 1.25 cm 25% 5.0 cm 1.25 cm 25%
31 5.0 cm 2.20 cm 44% 5.6 cm 1.40 cm 25%

For these errors in the calibration data, CPC can be seen to be essential to getting usable
values from SMF. For example, without CPC and with using an incorrect tube voltage of
25 kVp, the average h value is resolved to be negative. With CPC, the average h value is a far
more sensible 1.0 cm.
As table 2 demonstrates, with these incorrect values for tube voltage, the breast
composition measures vary wildly without CPC and even with CPC there are still large
variations in the average h value (1.00 to 1.40 cm) but it holds the pseudo-SMF% value
remarkably constant (25–27%), due to breast thickness rising in tune with average h value and
so their ratio remaining fairly constant.
Finally, in an attempt to push CPC to its limits, we insert major errors in all the data,
including differences in the filter/target combinations used. Figure 5 shows the relevant
transfer functions and the legend explains the values used. Table 3 shows the estimated pseudo-
breast composition values. CPC again shows capability in reducing the errors introduced, but
both the average h and pseudo-SMF% show wide variations.
Earlier we indicated that previous theoretical analysis had shown that small changes in
breast thickness could cause large variations in SMF generated breast composition measures
and in particular, SMF% which is generated by dividing SMF volume by Breast Volume.
Large variations occurred because, under the independent parameter assumption, as breast
thickness rises, all the h values within the breast decrease and thus SMF volume decreases.
At the same time, as breast thickness rises, so does Breast Volume. Since SMF% is the
2704 R Highnam et al

7.00
28kVp, Mo/Mo9, 73mAs,true values
6.00

5.00
30kVp Rh/Rh9, 50mAs, no CPC

h int (cm)
4.00 25kVp, Mo/Mo9, 300mAs, no CPC

3.00 30kVp, Rh/Rh9, 50mAs, CPC H=5.5cm

2.00
25kVp, Mo/Mo9, 300mAs, CPC H=6.5cm
1.00

0.00
0 1000 2000 3000 4000 5000
Pixel Value

Figure 5. SMF transfer function with extreme calibration data; the legend shows the calibration
data used as input to the transfer function. The wide black curve shows the ‘true’ values.

Table 3. Breast composition with extreme values of calibration data.

Without CPC With CPC

H Av hint Av hint/H H Av hint Av hint/H

Mo/Mo, 28 kVp, 73 mAs 5.0 cm 1.25 cm 25% 5.0 cm 1.25 cm 25%
Rh/Rh, 30 kVp, 50 mAs 5.0 cm 2.5 cm 50% 5.5 cm 2.0 cm 37%
Mo/Mo, 25 kVp, 300 mAs 5.0 cm 3.0 cm 60% 6.5 cm 1.3 cm 20%

ratio of SMF volume over Breast Volume, we now have a rising numerator (SMF volume), a
decreasing denominator (Breast Volume) and thus a rapidly decreasing SMF%.
On the other hand, in the constraint propagation view set out in this section, SMF volume
is held steady by CPC, and, in most cases is substantially less than Breast Volume meaning
that SMF% is a fraction with (an already) much larger denominator than numerator. In such
a situation, even relatively large changes in the denominator (i.e. Breast Volume) do not
cause SMF% to change drastically. However, we could imagine a small, dense breast say with
SMF% = 50% and H = 3.0 cm, where an H increase of 1.0 cm would have major impact on
SMF%.
The experiments presented in this section strongly suggest that SMF is susceptible to
errors in the calibration data and that CPC is capable of correcting for some of the inevitable
errors in the calibration data, so long as the ground truth is reliable. However, equally, if the
calibration data that we estimate are at the extreme end to the actual values then although
CPC can limit the errors passed through to the breast composition measures, they do still exist
and can be substantial. Nonetheless, the results are encouraging in suggesting that an SMF
implementation with CPC should be able to run on images with calibration data missing. To
investigate this, we assumed the most extreme case scenario, and in the next section show
the results from running the current SMF implementation over 4028 mammograms assuming
absolutely no calibration data!
It is relevant to note that in this section we have shown the sensitivity of SMF without
CPC to errors in the calibration data. However, it is worth noting that SMF might not be
so sensitive to some calibration parameters which affect only contrast, such as film–screen
gradient. In that case, some h values would rise, some would fall, and the overall effect would
be minimized.
Breast composition using retrospective SMF 2705

5. Retrospective SMF

5.1. Default values

In this section, we test the CPC process in the extreme case that no calibration data are
available (apart from the digitizer parameters, which are of course entirely independent of the
retrospectively analysed mammography machine). In this case, we make the following set of
default assumptions, most of which are those most frequently found in practice. As we shall
see, any or all of these assumptions can be changed:
• The filter/target combination is Mo/Mo.
• The tube voltage is assumed to be 28 kVp.
• An anti-scatter grid is assumed to be used.
• A ‘typical’ UK film–screen combination is used (see section 6 for more about this).
• The compression plate is assumed to slant by 0.5 cm across the width of the plate.
• The compression plate is assumed to be 0.4 cm thick and made of Lucite.
• The x-ray focus to detector distance is assumed to be 65.0 cm.
• The x-ray tube is ‘standard’ with a typical relationship between, for example, tube voltage
and tube current.
• Film reciprocal law failure happens as described in Highnam and Brady (1999).
• The anode heel effect is assumed to be ‘standard’ and we use the built-in default.
• No detector blur removal takes place, where detector blur here refers to the blurring
introduced by the intensifying screen.
Finally, we need to set a value for the unknown mAs. We do this using a very simple
method:
(1) Determine the breast area (A) in cm2 from the image segmentation, excluding the pectoral
muscle.
(2) Estimate the breast thickness according to H = 1.66 + 0.034 · A, an equation which
we derived from a set of values where a recorded breast thickness was available. This
estimate of H is only intended to be very approximate and is not used explicitly when we
later compute the breast thickness from the hint (x, y) = 0 curve.
(3) Assume that the automatic exposure control was set to give a film density of 1.6, and
assume that in the automatic exposure control measurement area we have 30% interesting
tissue and 70% fat.
(4) We thus have all the values to use equation (1) to work out what the mAs would have
been given all these assumptions.
This process provides estimates for all the calibration data required by SMF to provide an
accurate estimate of breast thickness from the determined ground truth, i.e. the hint (x, y) = 0
curve.

5.2. Data
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 (1948–1968)
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).
Those women still living in Scotland were asked to give consent for access to screening
mammograms taken under the Scottish Breast Screening Programme (1988 to 2002). Women
2706 R Highnam et al

were informed that their mammograms would be digitized. Cranio-caudal (CC) and medio-
lateral oblique (MLO) films were retrieved from the relevant breast screening centre for each
woman who consented and were digitized on site with a Canon FS300 digitizer scanner at a
resolution of 100 µm with 8-bit precision. X-ray parameters were collected, if available, from
local quality control and mammography records.
For the visual assignment of SCC and Wolfe patterns, scanned images were displayed at
300 µm resolution on a flat-panel display system. At this resolution, the images were roughly
the same size as a mammographic film. All the images were displayed to appear as if viewed
on a light-box (Highnam and Brady 1999). 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
from the original film (Jeffreys et al 2003a) Visual density measures were made by a radiologist
(RW) experienced in density assessment using the SCC scales (not semi-automated) and Wolfe
patterns. Future studies might also include BI-RADS ratings.
There were 3556 women in the original Glasgow Alumni Cohort, of whom 657 women
consented to help in this project and 4028 images were digitized. Of the 657 women in the
study, some had cancer and other publications from the authors using this database exclude
those cases whereas, in this document, all the available data are used. The images were taken
at six sites throughout Scotland over a 15 year period (1988 to 2002). Included women were
aged between 40 and 76 with the median age of 57.
Data inspection revealed several mismatches and errors in the data entry. Of the 4028
images, 3873 had mAs, 3983 had a valid kVp (between 25 and 32), and 3867 had both mAs
and a valid kVp. No separate tuning of the SMF algorithm was required or performed for
each site.

5.3. Results
In the following tables and figures, ‘with CD’ indicates that there was a valid mAs and kVp
for that image and the other calibration data were either available or defaulted; whilst ‘without
CD’ indicates that no calibration data at all were used other than knowing the digitizer to
pixel value conversion. When used in this mode, GenerateSMF (version 2.2β) defaults all the
values except mAs which is estimated as outlined in section 5.1.
GenerateSMF categorizes the SMF results based upon a number of simple metrics derived
from a close, automated, inspection of the images. For example, ‘Poor’ indicates that over
5% of the pixels in the internal region of the breast (away from the breast edge) had estimated
h values that were physically impossible, that is either greater than H or less than 0. In like
manner, ‘Excellent’ indicates that both of these totals were under 5%. It is important to note
that even when GenerateSMF judges the result to be ‘Poor’, ‘Usable’ SMF% and SMF volume
values might well be output where ‘Usable’ indicates that SMF% is between 0 and 100%, and
SMF volume is positive.
Table 4 presents a summary of the results of running GenerateSMF retrospectively both
with and without using recorded calibration data. As expected from the earlier theoretical
work in this paper, without CD the percentage of ‘excellent’ values is just 97.7% whilst with
CD it is 99.3%.
However, since the thrust of this work is towards retrospective use of GenerateSMF when
little or no calibration data is likely to be present, of more significance is that without CD we
achieve usable results in some 99.3% of the images compared to just 95.4% of the images
with CD. Obviously, if we ran SMF and allowed it to use calibration data (along with the
compensation technique) when available the results would be significantly better.
Breast composition using retrospective SMF 2707

100

80

minus SMF Volume estimated without CD

SMF Volume estimated with CD
60

40

20

-20

-40

-60

-80
0 50 100 150 200 250 300
Average of SMF Volume estimated with and without calibration data

Figure 6. Bland–Altman plot showing agreement between SMF volume estimated with and
without calibration data. The additional horizontal lines represent the 95% limits of agreement.

Table 4. General overview of running SMF with and without calibration data (CD).

With CD Without CD
Total number of images 3867 4028
‘Excellent’ results 3841 (99.3%) 3935 (97.7%)
‘Poor’ results 25 (0.7%) 92 (2.3%)
‘Crash’ results 1 1
‘Usable’ results 3844 (95.4% of 4028) 4001 (99.3% of 4028)

Ideally, we expect that SMF% and SMF volume estimated with and without CD should
be equal, and that the difference between the two measures should therefore be zero. Such
agreement is best measured using Bland–Altman plots, which plot the difference between
two measures of the same value against their average. Such plots illustrate good agreement
between the measures when the points are close to the zero horizontal line. Figures 6 and 7
show good agreement for SMF volume and SMF%, respectively, for small values of SMF%
respectively SMF volume, but poorer agreement at higher values.
Next, we compared the breast thickness H estimated with and without calibration data.
We also compared the estimated and recorded breast thickness (available in 3515 cases)
against the tube voltage, expecting that the bigger the breast, the higher the kVp, as automated
systems and radiographers both aim to reduce radiation dose to a minimum. Figure 8 shows
that the results are as expected. However, note that they also show that (a) the breast thickness
estimated using calibration data agrees closely with the value recorded; but (b) breast thickness
estimated without using any calibration data consistently underestimates the recorded value.
The underestimates are given numerically in the lower two rows of table 5.
Table 5 shows the consistency of the breast composition measures with and without
calibration data. The consistency of these results is estimated as the median difference
between mammograms taken for the same woman on the same day, of the same view for left–
right comparisons, and the same side for CC/MLO comparisons; see table 6. Importantly,
overall, there are no significant differences in left/right or CC/MLO consistency, depending
on whether SMF% or SMF volume are estimated using calibration data or without.
2708 R Highnam et al

40
35
30

minus SMF% estimated without CD

25

SMF% estimated with CD

20
15
10
5
0
-5
-10
-15
-20
-25
-30

10 20 30 40 50 60
Average of SMF% estimated with and without calibration data

Figure 7. Bland–Altman plot showing agreement between SMF% estimated with and without
calibration data. The additional horizontal lines represent the 95% limits of agreement.

9.0

8.0

7.0
Breast Thickness (H) cm

6.0
H recorded
5.0
H estimate with CD
4.0
H estimate without CD
3.0

2.0

1.0

0.0
25.0 26.0 27.0 28.0 29.0 30.0 31.0 32
kVp

Figure 8. Median estimated and recorded breast thickness versus tube voltage for 3515 images
with a recorded breast thickness, 3844 images with CD and known kVp and for 3922 images
without any CD used but with known kVp.

Table 5. Consistency with and without calibration data, median values.

With CD Without CD

SMF volume 64.5 cm3 57.1 cm3

SMF% 23.9% 26.7%
Hestimated 4.9 cm 3.6 cm
Hrecorded – Hestimated 0.0 cm signed 1.1 cm signed
0.5 cm Abs 1.2 cm Abs

We have previously reported that the correlation coefficients between paired mammograms
with calibration data, taken on the same day between left and right, and between CC and MLO
is high (Jeffreys et al 2006) with, for example, r = 0.92 (P < 0.001) for ln(SMF volume) and
r = 0.85 (P < 0.001) for ln(SMF%) between left and right breasts.
Breast composition using retrospective SMF 2709

120.0

100.0

SMF Volume cm3

80.0

60.0

40.0

20.0

0.0
0 1 2 3 4 5 6 7
SCC

Figure 9. SCC versus SMF volume generated using no calibration data showing median and
inter-quartile range values.

120.0

100.0

80.0
SMF Volume cm 3

60.0

40.0

20.0

0.0
0 1 2 3 4
Wolfe

Figure 10. Wolfe versus SMF volume generated using no calibration data and showing median
and inter-quartile ranges.

Table 6. Median signed differences (inter-quartile range) in SMF results, comparing data estimated
with and without calibration data.
With CD Without CD
Left − right difference
SMF volume 1 cm3 (−8 to 10) 1 cm3 (−9 to 10)
SMF% −0.1% (−2.6 to 2.3) −0.1% (−3.4 to 2.7)

MLO − CC difference
SMF volume 5 cm3 (−5 to 17) 4 cm3 (−8 to 15)
SMF% −0.6% (−3.7 to 2.3) −1.2% (−4.9 to 2.1)

Finally, figures 9 and 10 show the association between SCC and Wolfe grades and SMF
volume estimated without calibration data. As expected, there is no association, since both
Wolfe and SCC measure the percentage area of dense breast tissue, which is not obviously
related to the absolute volume. This is consistent with the results with calibration data reported
in Jeffreys et al (2006). On the other hand, the relationship between SCC and Wolfe grades
with SMF% is shown in figures 11 and 12, revealing a strong positive relationship.
2710 R Highnam et al

50.0

40.0

SMF%
30.0

20.0

10.0

0.0
0 1 2 3 4 5 6
SCC

Figure 11. SCC versus SMF% when generated using no calibration data; the median values and
inter-quartile ranges are shown.

50.0

40.0
SMF%

30.0

20.0

10.0

0.0
0 1 2 3 4
Wolfe

Figure 12. Wolfe versus SMF% when generated using no calibration data; the median values and
inter-quartile ranges are shown.

6. Discussion

We have introduced the concept of calibration parameter compensation (CPC), and

demonstrated that it enables GenerateSMF to be applied retrospectively to large datasets of
mammograms, for which calibration data are often not available. This potentially facilitates
epidemiological analysis of mammography datasets using the SMF representation, previously
thought to be not possible. Knowing approximately the shape of the correct transfer function,
and then estimating H, CPC fixes the transfer function so that it crosses the h-axis at a point
determined by information gleaned from the image, information which effectively serves as
‘ground truth’. We have demonstrated that this implicit constraint propagation procedure
provides a better basis for analysis of GenerateSMF than our previous Taylor’s series analysis,
in particular of the way in which the process minimizes the effects of errors.
The transfer function corrected using such ‘ground truth’ relates h and pixel values. This
suggests that SMF volume, being the sum of h values over the image, should be robust to errors
in the calibration data. Equally, the theory suggests that if errors are being compensated in
breast thickness, then SMF%, which uses breast thickness to estimate breast volume, should
be very susceptible to calibration data errors. Indeed, our previous work in this area (Highnam
and Brady 1999) and that of others (Blot and Zwiggelaar 2005) has focused on SMF% being
seriously compromised by errors in breast thickness. The theoretical sensitivity work found
in Highnam and Brady (1999) explores the effects of errors in each of the input parameters
and concludes the following:
Breast composition using retrospective SMF 2711

• Errors in many of the input parameters such as mAs and film speed lead (to a first
approximation) to translation shifts in the h values across the image, i.e. ∀ (x, y) ∈ Image,
hint (x, y) = hint (x, y) + k where k is a constant.
• Errors in input parameters that relate to image contrast such as film gradient and tube
voltage tend to generate higher h values in some places and lower h values in other so the
overall effect on SMF% and SMF volume is less.
• If all the other parameters are fixed then it is a reliable breast thickness estimate which
is critical with small errors giving rise to large swings in SMF%. For example, a 0.1 cm
change in breast thickness is capable of changing SMF% by 3–4%.

It is easy to see how the current SMF implementation compensates for the translation
errors, and is not entirely able to deal with non-linear changes in the transfer function. Yet,
the results shown in this paper strongly suggest that SMF volume and SMF% (even more so)
are robust to errors. We have argued in this paper that previous analyses have not matched
the implementation, and that, in particular, SMF% is robust because for most breasts the
Breast Volume (the denominator) is much greater than SMF volume (the numerator, and is
not changed by the CPC process) so that even quite large changes in the Breast Volume do not
change SMF% significantly; small dense breasts might not be so robust.
Despite the encouraging results using GenerateSMF without calibration data, we also
note that having calibration data generally provides more accurate results. Figure 8 shows this
clearly.
It is important to note that the estimated SMF volume increases when the calibration
data are known, whereas SMF% decreases. This cautions against using absolute levels of
the SMF breast composition measures between databases. For example, if for one database
we declare SMF% over 37% to be class 4 on the Wolfe scale, we may find that the next database
has overall increased levels of SMF% due to lack of calibration data, and many women have
(apparently) class 4 breasts. We will explore this issue in later papers.
On a related note, the database used in this work is from the UK where most mammograms
were routinely performed (following national screening standards) at 28 kVp with a Mo/Mo
filter/target combination. Consequently, the defaults used in this work (namely 28 kVp,
Mo/Mo) are likely to be good estimates. Anecdotal evidence suggests that mammograms in
the US and elsewhere have a higher degree of variability in technique and thus it may be that
the current SMF implementation would be less good at compensating for errors over a large,
US database for which almost nothing is known about the imaging of the mammograms. This
is, of course, a very pessimistic assumption. Note also that newer systems have a wider variety
of kVp and target/filter combinations, only some of which are set automatically. If these are
not known, then again the defaults might not give satisfactory results. Of course, the defaults
can easily be changed.
This work, and almost all other work to date, deals with ‘raw’, ‘unprocessed’ images,
whereas many of the new digital machines output images ready for display or ‘presentation’.
The post-processing that occurs to make those images ready for presentation is likely to hinder
the automatic generation of quantitative volumetric-based values.
The fundamental SMF assumption groups fibrous, glandular, cancerous and other tissues
into just the one class—‘interesting’. Since we aim to use SMF to estimate breast cancer risk
which might be, for example, related only to the amount of glandular tissue, it is of some
concern that SMF has just the one big class. Nonetheless, the correlations between SMF%
and SCC and Wolfe are very strong and SCC and Wolfe have both been shown to correlate
with breast cancer risk indicating that the big grouping does not remove the important aspects
of the data.
2712 R Highnam et al

GenerateSMF estimates breast thickness by assuming the existence of a fatty breast edge
and, on that basis, that a smooth curve exists slightly inside the breast edge for which all points
have h(x, y) = 0 and h(x, y) = H. An apparently easier way would involve simply selecting the
minimum pixel value (highest film density) ‘well within the breast’ and declaring that value to
have h(x, y) = 0 and h(x, y) = H. However, we contend that such a method will not be robust
over large databases due to the need to segment out the breast edge, will suffer from noise
issues, and might give very misleading results on the densest breasts where there simply is no
‘entirely fatty’ region inside the breast.
Finally, as suggested earlier, there will be outliers which have such extreme error values
that the default and estimated values are way-off and CPC is unable to correct for those
extremes. Further work should investigate the ‘poor’ results and will likely need to investigate
ways of estimating kVp and some of the other non-translation error parameters from the
images rather than accepting defaults.

7. Conclusion

Retrospective use of SMF, and use of SMF without calibration data, is possible and yields
quantitative results which strongly correlate with SCC and Wolfe. This suggests that SMF
can be applied to any number of retrospectively collected images, with or without calibration
data, and reliable breast composition measures generated automatically.
Future work will include showing that the breast composition measures generated by SMF
(with and without calibration data) correlate directly with breast cancer risk.

Acknowledgments

The work of Pat Forrest in collecting data is greatly appreciated. The Centre for Public
Health Research is supported by a Programme Grant from the Health Research Council of
New Zealand. Dr Jeffreys (née Okasha) was previously employed at the University of Bristol,
and she is grateful for financial support from Breast Cancer Campaign, Breast Cancer Research
Trust and World Cancer Research Fund International. Part of this work was funded by the EU
MammoGrid project. Finally, many thanks to Chris Tromans for useful discussions around
SMF and for showing the benefits of the transfer function approach. We also thank the referees
for their numerous suggestions as to better presentation insightful comments.

References

ACR (American College of Radiology) 1998 Breast Imaging and Reporting and Data System (BI-RADS) 3rd edn
(Reston, VA: ACR)
Blot L and Zwiggelaar R 2005 A volumetric approach to glandularity estimation in mammography: a feasibility
study Phys. Med. Biol. 50 695–708
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
Burch A and Law J 1995 A method for estimating compressed breast thickness during mammography Br. J. Radiol.
68 394–9
Byng J, Boyd N, Fishell E, Jong R and Yaffe M 1994 The quantitative analysis of mammographic densities
Phys. Med. Biol. 39 1629–38
Heine J and Malhotra P 2002 Mammographic tissue, breast cancer risk, serial image analysis, and digital
mammography: 1. Tissue and related risk factors Acad. Radiol. 9 298–316
Highnam R and Brady M 1999 Mammographic Image Analysis (Dordrecht: Kluwer )
Highnam R, Brady M and Shepstone B 1994 Computing the scatter component of mammographic images IEEE
Trans. Med. Imaging 13 301–13
Breast composition using retrospective SMF 2713

Highnam R, Brady M and Shepstone B 1996 A representation for mammographic image processing Med. Image
Anal. 1 1–18
Highnam R, Brady M and Shepstone B 1998 Estimation of breast thickness in mammography Br. J. Radiol. 646–53
Hufton A, Astley S, Marchant T and Patel H 2006 A method for the quantification of dense breast tissue from digitised
mammograms Proc. IWDM 2004 vol 2 (at press)
Jeffreys M, Warren R, Davey Smith G and Gunnell D 2003b Breast density: agreement of measures from film and
digital image Br. J. Radiol. 76 561–3
Jeffreys M, Warren R, Gunnell D, McCarron P, Highnam R and Davey Smith G 2003a Body mass index in young
adulthood and breast cancer risk (abstract) Australas. Epidemiol. 10 17
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. (at press)
Johns P and Yaffe M 1987 X-ray characterization of normal and neoplastic breast tissue Phys. Med. Biol. 32 675–95
Linguraru M, Brady M and Yam M 2002 Detection of microcalcifications using SMF Int. Workshop on Digital
Mammography (Lecture Notes in Computer Science) (Berlin:: Springer)
Marias K, Behrenbruch C, Highnam R, Parbhoo S, Seifalian A and Brady M 2004 A mammographic image analysis
method to detect and measure changes in breast density Eur. J. Radiol. 52 276–82
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
Pawluczyk O, Augustine B J, Yaffe M J, Rico D, Yang J and Mawdsley G E et al 2003 A volumetric method for
estimation of breast density on digitized screen-film mammograms Med. Phys. 30 352–64
Tromans C 2006 Measuring breast density from x-ray mammograms PhD Thesis University of Oxford
Wolfe J N 1976 Breast patterns as an index of risk for developing breast cancer Am. J. Roentgenol. 126 1130–37
Yam M, Highnam R, Brady M and English R 1999 Detecting calcifications using the hint representation
Proc. Computer Assisted Radiology and Surgery (CARS) pp 373–77