Key words
mammography - breast cancer - breast density - volumetric measurement
Introduction
Mammographic breast density has been shown to be one of the strongest known markers
of breast cancer risk and has been proposed as a variable for individual risk assessment
[1]
[2]
[3]
[4]. Some investigators have used breast density as an intermediate end point for interventional
studies [5]
[6]. An assessment of radiographic breast density is required in every mammography report
and is an important variable in research studies. Breast density may in the future
become a factor for individualizing breast cancer screening regimens according to
each woman’s risk profile and the expected sensitivity of mammography given her individual
breast density [7]
[8]. A number of different reporting schemes have been developed, with the American
College of Radiologists Breast Imaging Reporting and Data System (BI-RADS) perhaps
being the most widely used system. However, visual assessment of breast density has
limited intra- and interobserver reproducibility [9]
[10]
[11]. A variety of approaches have been tested to objectify breast density assessment
[12]
[13]
[14]
[15]. A drawback of most approaches is the demand on reader time, which limits their
use in the clinical setting and population studies alike. Given the importance of
breast density for risk stratification, an accurate, fast, and reproducible method
for assessing breast density is needed [4]
[16]. Volumetric breast density measurement provides an estimate of breast percent density
(PD) without reader interaction. The method uses a model of the imaging chain to estimate
total breast volume and the amount of glandular tissue present. An important advantage
of this approach is that it avoids subjectivity, which is introduced whenever different
readers rate breast density on the same study, as this software-based method always
produces the same result when presented with identical image input. However, while
the algorithm has been calibrated to volume measurements of sample breasts, it is
not clear to what extent the produced result deviates from reality. Also, there is
no study data on the reproducibility of measurements when there is variation in data
input, for example in repeated examinations of the same patient, when differences
in projection angle, breast compression and image acquisition parameters may affect
the apparent breast density. Given that a possible application of this algorithm is
its use in longitudinal studies of breast density, this application requires testing
in a sample of consecutive mammograms. This information is necessary for estimating
the magnitude of error resulting from variations in the imaging chain and provides
a measure of the reproducibility of the process as a whole. The aim of this study
was to assess the reproducibility of breast density assessment using the R2 Quantra
software in serial mammography examinations and to compare its performance with that
of human readers.
Materials and Methods
Patients
We searched our records from June 2002 to December 2006 for patients satisfying the
following inclusion criteria: two consecutive examinations performed on the same mammography
unit no more than 24 months apart, raw image data stored in the picture archiving
and communication system (PACS), unremarkable mammography reports for at least one
breast, and minimum of 18 months of normal follow-up of the eligible breast(s). The
exclusion criteria were: previous surgery on the eligible breast(s), change in hormone
status such as starting or stopping hormone-replacement therapy or menopause, and
technical deficits of the mammogram such as inadequate positioning or presence of
large skin folds.
A total of 170 patients were identified. Raw image data of the two consecutive mammography
examinations were sent to an R2 CenovaTM server for analysis by the R2 QuantraTM breast density assessment algorithm. In 29 patients, the algorithm failed to produce
results for one or both examinations. These patients were excluded from the analysis.
Therefore, 141 patients were included in the study. In 21 patients, the algorithm
produced results but marked the results as potentially inaccurate. This occurs when
there is a discrepancy between the measurements in the CC and MLO projections. This
was recorded for subgroup analysis of the reproducibility.
Only one breast per patient was chosen for analysis to avoid linkage of data points.
If both breasts were eligible for analysis, one side was chosen at random. Institutional
review board was obtained.
Image acquisition
All patients underwent digital mammography using the same full-field digital mammography
system with a flat-panel detector and a cesium iodide absorber, field size 19 × 23 cm,
pixel size 100 μm, image matrix size 1914 × 2294 (Senographe 2000 D, General Electric
Healthcare, Chalfont St. Giles). All mammograms were acquired in standard craniocaudal
and mediolateral oblique projections using automatic optimization of acquisition parameters
and standard supplier presets.
Image analysis
For all patients included in the study, breast density was assessed both visually
and with the automatic software tool.
Visually, breast density was assessed by three independent, board-certified radiologists
of our hospital using the BI-RADS lexicon. Reading was performed on a diagnostic mammography
workstation (syngo MammoReport, Siemens Medical, Erlangen, Germany) in a blinded manner
without knowledge of the woman’s age, the original mammography interpretation, and
risk profile for breast cancer. The three observers independently assessed the mammograms
for breast density, assigning one of the BI-RADS breast density categories on a standardized
form. The first mammogram of each patient was read first, followed by another reading
session for the second mammogram after an interval of 4 weeks or more. In a third
reading session, again after an interval of at least 4 weeks, the first set of mammograms
was read a second time to estimate the intra-rater reproducibility.
The BI-RADS scheme of breast densities, developed by the American College of Radiology
(ACR) is intended to provide a standardized classification system for mammographic
studies. The ACR classification identifies four categories of breast composition:
(1) the breast is almost entirely fat (< 25 % glandular); (2) there are scattered
fibroglandular densities (25 – 50 % glandular); (3) the breast tissue is heterogeneously
dense (approximately 51 – 75 % glandular); and (4) the breast tissue is extremely
dense (> 75 % glandular).
For the software-based analysis, raw image data were sent to a dedicated server running
the R2 Quantra software. Briefly, R2 Quantra™ is a software tool for automatically
calculating volumetric breast density from the ratio of fibroglandular tissue to the
estimated total breast volume. The algorithm uses a physical model of the imaging
process to deduce the density and composition of breast tissue from the degree of
X-ray attenuation on mammograms. To achieve this, the algorithm estimates the amount
of fibroglandular tissue an X-ray beam must have passed to deposit the amount of energy
measured at the detector. Images are processed within minutes. The output of the R2
Quantra software includes the estimated total breast volume and fibroglandular tissue
volume in ml (cm3) and the calculated breast PD ([Fig. 1]).
Fig. 1 Representative mammogram of the right breast in craniocaudal and mediolateral oblique
projection and the corresponding datasheet provided by the QuantraR2 software.
Abb. 1 Repräsentative Mammografie-Aufnahmen der rechten Brust in craniocaudaler und mediolateral-obliquer
Ausrichtung sowie das korrespondierende Datenblatt der QuantraR2-Software.
Statistical analysis
Data analysis was performed using statistical software packages (SPSS, version 18.0;
SPSS Chicago, Illinois; MedCalc 12.3.0). The intra- and inter-rater reproducibility
as well as the inter-examination reproducibility of the visual and software-based
analysis were assessed by calculating the intraclass correlation coefficient (ICC).
For comparison with other studies of visual density assessment, quadratic-weighted
kappa values were also calculated for the intra- and inter-rater reproducibility.
For the correlation of categorical BI-RADS density levels of examinations 1 and 2
versus ordinal volumetric breast density values, BI-RADS classes 1 – 4 were replaced
with the mean PD value of the respective category (1 = 12.5 %; 2 = 37.5 %; 3 = 62.5 %;
4 = 87.5 %), and the ICC was calculated.
To investigate the effects of different compression forces on breast density estimates
by volumetric assessment, we assigned the patients to one of four subgroups based
on the magnitude of the difference in compression force applied for the first and
the second mammogram in each patient. For each subgroup, the inter-examination agreement
of the measured breast density was determined. Differences in correlation coefficients
were tested for statistical significance using the Fisher r-to-z transformation.
Results
The patients had a mean age of 62 years (range, 45 – 78 years). 61 patients underwent
mammography in the setting of surveillance after breast surgery and had one unaffected
breast. The remaining 80 patients had workup of a palpable lump or unclear ultrasound
findings. The median interval between the first and the second examination was 13.2
months with a range of 9 – 24 months. 29 patients were premenopausal, 112 patients
were postmenopausal. Of the premenopausal patients, 6 patients took oral contraceptive
agents. Of the postmenopausal patients, 12 received hormone replacement therapy and
17 received antihormonal therapy.
The results for inter-rater agreement in visual breast density assessment between
pairs of observers for both examinations, 1 and 2, are summarized in [Table 1]. The inter-rater agreement ranged from 0.71 – 0.77 (ICC).
Table 1
Inter-rater variability. Intraclass correlation coefficients (ICC) and quadratic-weighted
kappa values were calculated. Numbers in parentheses represent 95 % confidence intervals.
Tab. 1 Interrater-Variabilität. Es wurden die Intraklassen-Korrelationskoeffizienten (ICC)
und quadratisch-gewichteten Kappa-Koeffizienten (κ) bestimmt. Zahlen in Klammern stellen
die jeweiligen 95 % Konfidenzintervalle dar.
|
|
examination 1
|
examination 2
|
|
rater A vs. B
|
|
ICC
|
0.71 (0.62 – 0.78)
|
0.74 (0.65 – 0.80)
|
|
κ
|
0.69 (0.61 – 0.77)
|
0.73 (0.66 – 0.81)
|
|
rater A vs. C
|
|
ICC
|
0.77 (0.71 – 0.86)
|
0.74 (0.66 – 0.81)
|
|
κ
|
0.76 (0.79 – 0.82)
|
0.75 (0.66 – 0.83)
|
|
rater B vs. C
|
|
ICC
|
0.77 (0.69 – 0.83)
|
0.76 (0.68 – 0.82)
|
|
κ
|
0.69 (0.58 – 0.78)
|
0.72 (0.62 – 0.82)
|
[Table 2] summarizes the results for intra-rater agreement for examination 1, the inter-examination
variability for raters and volumetric measurements, as well as the comparison between
visual breast density assessment and volumetric analysis. The intra-rater agreement
ranged from 0.81 – 0.84 (ICC). The inter-examination agreement of examinations 1 and
2 for individual readers varied from 0.75 – 0.81 versus 0.91 for volumetric analysis.
The difference in the strength of correlation between volumetric and visual assessment
was statistically significant for all readers and constellations (p≤ 0.01). In patients
where breast density was marked as potentially inaccurate by the R2 Quantra software,
the inter-examination agreement was 0.90 (95 % confidence intervals, 0.77 – 0.96).
Table 2
Intra-rater agreement, agreement of R2 Quantra and visual assessment, and inter-examination
agreement for visual and software-based breast density assessment. Numbers in parentheses
represent 95 % confidence intervals.
Tab 2 Intrarater-Übereinstimmung, Übereinstimmung von R2 Quantra zu visueller Bestimmung
sowie Interexamination-Übereinstimmung für visuelle und Software-basierte Brustdichtebestimmung.
Zahlen in Klammern repräsentieren die 95 % Konfidenzintervalle.
|
|
|
rater A
|
rater B
|
rater C
|
quantra PD
|
|
intra-rater agreement
|
|
|
ICC
|
0.83 (0.78 – 0.88)
|
0.81 (0.74 – 0.86)
|
0.84 (0.77 – 0.88)
|
/
|
|
|
Κ
|
0.81 (0.75 – 0.87)
|
0.80 (0.72 – 0.87)
|
0.82 (0.75 – 0.89)
|
/
|
|
agreement quantra vs. visual assessment
|
|
examination 1
|
ICC
|
0.68 (0.58 – 0.76)
|
0.68 (0.58 – 0.76)
|
0.65 (0.55 – 0.74)
|
/
|
|
examination 2
|
ICC
|
0.69 (0.59 – 0.77)
|
0.63 (0.51 – 0.83)
|
0.73 (0.64 – 0.80)
|
/
|
|
inter-examination agreement
|
|
|
ICC
|
0.75 (0.67 – 0.83)
|
0.81 (0.74 – 0.86)
|
0.76 (0.67 – 0.84)
|
0.91[*] (0.87 – 0.93)
|
* indicates statistical significance of the difference in ICC compared with all other
ICC values (p≤ 0.01).
* Zeigt einen statistisch signifikanten Unterschied des ICC-Wertes im Vergleich zu
allen anderen ICC-Werten an (p≤0.01).
[Table 3] shows the inter-examination correlation of the volumetric analysis of the whole
group and for the four subgroups based on magnitude of difference in compression forces.
The inter-examination correlation of the volumetric analysis was similar in all groups,
regardless of the differences in mean compression forces.
Table 3
Inter-examination reproducibility of software-based analysis by magnitude of difference
in compression forces between the two mammography examinations. Numbers in parentheses
represent 95 % confidence intervals.
Tab. 3 Interexamination-Reproduzierbarkeit der Software-basierten Analyse in Abhängigkeit
vom Ausmaß der Kompressionskraftschwankungen zwischen den zwei Mammografieuntersuchungen.
Zahlen in Klammern repräsentieren die 95 % Konfidenzintervalle.
|
difference in compression force
|
N (%)
|
inter-examination reproducibility (ICC)
|
|
0 – 39 N
|
49 (35 %)
|
0.89 (0.82 – 0.94)
|
|
40 – 79 N
|
51 (36 %)
|
0.92 (0.86 – 0.95)
|
|
80 – 119 N
|
27 (19 %)
|
0.92 (0.83 – 0.96)
|
|
≥ 120 N
|
14 (10 %)
|
0.91 (0.73 – 0.97)
|
|
total
|
141 (100 %)
|
0.90 (0.87 – 0.93)
|
Discussion
The aim of our study was to assess the reproducibility of breast density measurement
in consecutive examinations using volumetric breast density analysis software and
to compare the results with the performance of human readers.
We found substantial, but not excellent, intra- and interobserver reproducibility
of the visual density classification, comparable to the results reported by other
studies. The inter-examination reproducibility of visual assessment was equal to or
slightly less than the intra-examination reproducibility, depending on the reader.
In comparison, breast density measurement by volumetric analysis showed an excellent
inter-examination reproducibility, which was significantly higher than that of human
readers. There was good agreement of the readers’ results with the volumetric analysis.
We found no influence of differences in breast compression on the reproducibility
of volumetric breast density analysis. Results that were marked as discrepant in CC
and MLO views and therefore potentially inaccurate by the software were as reproducible
as results that were not marked as potentially inaccurate.
Breast density has been shown to be the strongest known risk factor for breast cancer
[1]
[2]
[3]
[4]
[17]. There is some evidence that breast density may reflect changes in breast cancer
risk associated with interventions such as tamoxifen treatment [18]. From a clinical perspective, breast density has a strong effect on mammographic
sensitivity [19]
[20]. Future breast cancer screening programs may employ individualized screening regimens
for women according to their personal breast cancer risk as well as their chance of
benefiting from additional procedures like breast ultrasound or digital breast tomosynthesis
[21]
[22]. Therefore, accurate and reproducible measurement of breast density is very desirable
both in the clinical and research setting. The results of our study show that volumetric
analysis provides highly reproducible measurements of breast density in consecutive
examinations and clearly exceeds the performance of human readers. The method appears
to be robust with respect to differences in breast compression as well as the small
differences in breast orientation and projection angle, which may occur in consecutive
examinations. Volumetric analysis is therefore preferable to visual assessment in
the setting of longitudinal studies of breast density.
Most studies investigating the reproducibility of breast density assessment have looked
at intra- and inter-rater reproducibility. Software-based volumetric analysis always
yields the same result when confronted with the same mammogram, thereby eliminating
intra- and interobserver variability. As immediate acquisition of a second mammogram
after a satisfactory mammogram has been obtained is not possible for ethical reasons,
we used serial mammograms for estimating the reproducibility of the method. The reproducibility
of visual breast density assessment has been shown to be substantial but not perfect
[9]
[10]
[11]. Interactive thresholding in one study of digitized film mammograms improved both
the inter- and intra-rater reproducibility, with an increase in the intraclass coefficients
to 0.84 – 0.94 and 0.93 – 0.99, respectively.[12] Another study showed better correlation of the Cumulus method with another automated
density assessment algorithm than with the four-category BI-RADS scale on digitized
mammograms [13]. However, ours is the first study to investigate the reproducibility of breast density
assessment in serial examinations.
Three-dimensional imaging techniques, such as MR volumetry and digital breast tomosynthesis,
may yield similar information and potentially provide more accurate volume measurements.
However, the strength of quantifying breast tissue density from digital mammograms
is that these are inexpensive and widely available. A current limitation of the software
is the failure rate of around 8.5 % observed in this study, which may be improved
with future developments.
The results of our study are relevant both to the use of this method in longitudinal
studies and to the comparison of results obtained in different imaging centers, where
variations in imaging technique cannot be fully avoided. The lack of reader interaction
and the avoidance of intra-rater variability represent notable advantages over alternative
breast density assessment approaches. It should be noted that the high reproducibility
(precision) of this method does not allow assumptions about its accuracy, i. e. the
closeness of the software result to the true breast composition. While a highly accurate
measurement would be highly reproducible, high reproducibility does not prove high
accuracy. However, the high reproducibility of this algorithm means that changes in
breast density over time will be detected with much higher precision by volumetric
assessment than by visual assessment.
The major limitation of our study is the long interval between consecutive mammography
examinations in the same patients. While 1 – 2 years is the minimum interval for performing
serial mammography after an initial unremarkable mammogram, this is long enough for
changes in weight to occur and changes in hormone levels to manifest. The reproducibility
found in this study, therefore, very likely represents an underestimate.
In conclusion, volumetric breast density measurement is highly reproducible in serial
mammograms in a routine clinical setting. The performance significantly exceeds the
reproducibility of visual assessment by human readers. The method appears robust with
respect to variations in breast compression. Given the lack of reader interaction
and the avoidance of intra- and inter-rater variability, this method is a useful tool
for longitudinal studies of breast density and for the quantification of breast density
for breast cancer risk stratification.
