British Journal of Radiology 74 (2001),1048-1051 © 2001 The British Institute of Radiology
A mammographic dilemma: calcification or haemosiderin as a cause of opacities? Validation of a new digital diagnostic tool
M Yam, DPhil1,
J Tchou, MD, PhD2,
R English, MBChB, MRCP, FRCR3,
R Highnam, DPhil4,
R Highnam, DPhil4,
D Roskell, MA, BMBCh, MRCPath5,
M Greenall, MBChB, ChM, FRCR6 and
M Brady, FRS, FEng1
1Medical Vision Laboratory, Robotics Research, Engineering Science, University of Oxford, Parks Road, Oxford OX1 3PJ, UK, 2Department of General Surgery, Johns Hopkins University School of Medicine, Baltimore, MD, USA, Departments of 3Radiology, 5Histopathology and 6Surgery, Oxford Radcliffe Hospitals NHS Trust, Headley Way, Headington, Oxford OX3 9DU, UK and 4Mirada Solutions Ltd., Oxford Centre for Innovation, Mill Street, Oxford OX2 0JX, UK
 |
Abstract
|
|---|
Core biopsies of an area of microcalcification demonstrated large collections of macrophages containing haemosiderin, with evidence of minimal microcalcification on H&E staining. Algorithms were developed that were capable of differentiating with high accuracy those signs due to calcification, using quantitative measurements such as the apparent volume composition of calcium. Using the linear attenuation coefficients of calcification and assuming an ellipsoid model for the 3-dimensional shape of calcification, we computed the relative calcification volume for each region of interest. The difference in the linear attenuation coefficients of iron and calcification allowed the two to be differentiated on a mammogram based on this measure of relative calcification volume.
|
Introduction
|
|---|
The purpose of this article is to describe a new method of determining the constitution of mammographic opacities, in particular those appearing as tiny flecks of radiodense material such as calcification.
The discussion arose in relation to a core biopsy of the breast that was performed for the evaluation of microcalcification. The pathologist reported that the sample contained haemosiderin and asked whether this, rather than calcium deposition, was the aetiology of the mammographic opacities. To attempt to resolve this, the Medical Vision Laboratory applied newly developed techniques of image analysis to see whether calcification or haemosiderin was detectable.
The problem arose in the case of a 75-year-old woman who had undergone a mammographic localization biopsy of screen-detected suspicious microcalcification of the upper outer quadrant of her left breast. Pathological examination demonstrated ductal carcinoma in situ of at least 1.4 cm in diameter with positive margins. Wide excision and axillary node sampling showed no residual in situ carcinoma or lymph node involvement. Her post-operative course was complicated by formation of a haematoma in the breast. The drain placed at the wide local excision site was removed a few days after surgery. Mammography 1 year later showed no evidence of recurrence. New microcalcification was seen in the left upper outer quadrant on the second year follow-up mammogram. Five core biopsies were taken from this area. Two flecks of calcification were present in the specimen as confirmed by radiography. Histology of all five core biopsies showed breast tissues containing fibrosis with widespread granulomatous inflammation and fat necrosis. Large collections of macrophages were present containing haemosiderin pigment forming some linear aggregates within the fibrous tissue, but no microcalcification was seen (Figure 1
) in the histological sample.
Histological analysis of the biopsy samples suggested that the mammographic opacities may have resulted from iron, rather than calcium deposition. Conventional magnification radiography was applied to the block of tissue known to contain haemosiderin in the core biopsy samples in order to determine whether radio-opaque material was present.
|
Materials for image analysis
|
|---|
The two mammograms, taken under craniocaudal (CC) and mediolateral oblique (MLO) compressions, were digitized to a resolution of 50 µm per pixel using a Lumisys LS85 laser film digitizer (Lumisys Inc., Sunnyvale, CA). Two image samples, each of 250 pixels x 200 pixels (corresponding to an actual film area of 1.25 cm x 1 cm), were extracted from each mammogram. The image samples taken from the CC and MLO views were denoted by I1CC, I2CC and I1MLO, I2MLO respectively. I1CC corresponded to the breast region where a core biopsy was carried out during CC compression and where iron pigment was detected. I1MLO contained a bright region, which was considered to correspond to the iron pigment detected in I1CC. I2CC and I2MLO both contained an isolated microcalcification, which correspond to two views of the same object. The MLO–CC correspondence of regions was determined by the radiologist.
Each image sample was first converted to a normalized mammographic image representation, the 
representation [1, 2], which is developed by modelling the physics of the X-ray imaging process and discarding various undesirable image degrading factors, such as scatter and extrafocal radiation. In the hint representation, each pixel value represents the thickness of "interesting" (or non-adipose) tissue in the cone of breast tissue compressed between the compression plate and the imaging surface above that pixel.
Using isocontours and imposing an area constraint as described by Yam et al [3], seven candidate regions were extracted from the four image samples: a1, a2 and a3 from I1CC (Figure 2a); b from I2CC; c1 and c2 from I1MLO (Figure 2b); and d from I2MLO. Owing to the projective nature of overlapping structures, the number of regions extracted from the CC and MLO views was different. The relative calcification volume 
was then computed for each of these candidate regions, as described below.
View larger version (87K):
[in this window]
[in a new window]
|
Figure 2. (a) Regions a1, a2 and a3 extracted from image sample I1CC of the craniocaudal view. (b) Regions c1 and c2 extracted from image sample I1MLO of the mediolateral oblique view. These regions correspond to the breast region where biopsy was performed.
|
|
Analysis of relative calcification volume : method and findings
Our method is based upon the quantity , which is the ratio of the volume of calcification 
to the estimated 3-dimensional (3D) volume 
of the region. Essentially, gives the volume composition of calcification within that region of interest.
An outline of the method is given here; it is described in greater detail in Yam et al [3]. Using a normalized mammographic image representation , we computed, for each pixel within the candidate regions, the calcification thickness 
, which refers to the amount of calcification present in the compressed breast tissue above that pixel. Using the values of the linear attenuation coefficients at 18 keV (µint=1.028, µfat=0.558 [4] and µcalc=26.1) for a typical calcification composed of calcium salts such as calcium hydroxyapatite, it was derived by Highnam and Brady [1] that:
where is the thickness of interesting (or non-adipose) tissue within the column of compressed breast tissue above a certain pixel, and 
is the average thickness of interesting tissue in the surrounding background area. Summing values of all pixels within the region gives us the total volume of calcification in that volume of breast tissue. The estimated 3D volume was approximated from the spatial dimensions of the projected 2-dimensional (2D) region using an ellipsoid model. The volume ratio , was then computed by:
Assuming the entire volume of interest is composed of calcification and has a 3D shape that approximates an ellipsoid, this quantity should have a theoretical value near to 1. However, experimental values are expected to deviate slightly from 1. This will be further discussed below.
Since iron has a linear attenuation coefficient approximately 10 times higher than that of calcium [5], attenuation of an X-ray beam through a small volume of iron would be comparable with that through a much thicker layer of calcification. Hence, if a substantial amount of iron were present within a region of interest, the computed calcification volume would be large and the resulting would be well above 1.
Table 1 presents the and values of the extracted regions. It is interesting to note that of region a3 was very close to that of region c2, and likewise between region b and d. Also, the sum of of regions a1 and a2 is also comparable with that of region c1. This supports, quantitatively, the hypothesis that the respective regions are projections of the same anatomical structure in different views.
The values shown in Table 1
approximate the volume composition of calcium present in the extracted regions of interest. Apart from region a1, which has a relatively high of 1.53, all the other regions have a less than 1. However, there is a significant discrepancy between of a1 and that of its corresponding region in the MLO view, c1, the of which is only 0.49. If the high of a1 resulted from the presence of iron, we would expect the of c1 also to be high. Moreover, given the fact that iron has a much higher X-ray attenuation ratio than calcium, the should be higher still should a substantial amount of iron be present within the region.
Integrating information from both views and noting the linear attenuation coefficients of calcium and iron, we concluded that the high 
of region a1 is more likely to be a result of projection of more than one overlapping structure or significant deviation of its 3D shape from the ellipsoid model, rather than being owing to the presence of a significant amount of iron.
|
Discussion
|
|---|
Detection of microcalcification is an integral part of the diagnosis of breast cancer by mammography [6, 7]. However, some heavy metals have been recorded as mimicking calcification on mammography [8–10], which could lead to an erroneous diagnosis. The method described provides a tool for making this distinction.
It must be stressed that the method only approximates the volume composition of calcifications, under the assumption of an ellipsoid 3D shape model. In this particular case, the major source of error comes from the projective nature of more than one overlapping structure and from the 3D shape model. An accurate picture of 3D structures from a single 2D view is hard to determine. However, combining information from both CC and MLO views would provide additional clues in this respect.
We conclude that regions a1, a2 and a3 in the CC view, and their corresponding regions c1 and c2 in the MLO view, have relative volumes more consistent with that of calcification. In other words, there is little evidence from our experimental results to suggest that a substantial amount of iron is present within these regions of interest. However, we do not rule out the possibility that there might be iron in the breast over a dispersed area, which could explain the biopsy findings. That might lead to an overall greater X-ray attenuation and thus a rise in brightness over a region, which we cannot investigate using our present method. The fact remains that the specific signs, i.e. the microcalcifications, seem unlikely to be iron and are perfect for investigation by our processing techniques.
The pathology blocks of the core biopsy were subjected to radiography. The magnification view showed two flecks of radio-opaque material that did not correspond to the extensive area of haemosiderin as demonstrated by the H&E stain.
|
Acknowledgments
|
|---|
The authors would like to thank the staff of the Breast Care Unit at the Churchill Hospital, Oxford, for their continuing help and support and, in particular, Liz Robinson for her help in the preparation of the manuscript.
|
Footnotes
|
|---|
The doctoral research of M Yam was supported by the Croucher Foundation, Hong Kong, and the Overseas Research Student (ORS) Awards Scheme, UK. The work of M Brady was supported by the EPSRC, UK under Grant GR/M54995. 
Received for publication December 11, 2000.
Accepted for publication August 24, 2001.
|
References
|
|---|
-
Highnam R, Brady M. Mammographic image analysis. Dordrecht: Kluwer Academic, 1999.
-
Yam M, Highnam R, Brady M. De-noising hint surfaces: a physics-based approach. In: Taylor C, Colchester A, editors. Proceedings of the 2nd International Conference on Medical Image Computing and Computer-Assisted Intervention; 1999 September 19–22; Cambridge, UK. Berlin: Springer, 1999:227–34.
-
Yam M, Brady JM, Highnam RP, English RE. Detecting calcifications using the hint representation. In: Lemke HU, Vannier MW, Inamura K, Farman AG, editors. Proceedings of the 13th International Congress on Computer Assisted Radiology and Surgery; 1999 June 23–26; Paris. Amsterdam: Elsevier 1999:373–7.
-
Johns PC, Yaffe MJ. X-ray characterisation of normal and neoplastic breast tissue. Phys Med Biol 1987;32:675–95.[Medline]
-
Hubbell JH, Seltzer SM. Tables of X-ray mass attenuation coefficients and mass energy-absorption coefficients. Gaithersburg, MD: National Institute of Standards and Technology, 1996. http://physics.nist.gov/PhysRefData/XrayMassCoef/
-
Egan RL, McSweeney MB, Sewell CW. Intramammary calcificiations without an associated mass in benign and malignant diseases. Radiology 1980;137:1–7.[Abstract/Free Full Text]
-
Moskowitz M. The predictive value of certain mammographic signs in screening for breast cancer. Cancer 1983;51:1007–11.[Medline]
-
Bruwer A, Nelson GW, Spark RP. Punctate intranodal gold deposits simulating microcalcifications on mammograms. Radiology 1987;163:87–8.[Abstract/Free Full Text]
-
Carter TR. Intramammary lymph node gold deposits simulating microcalcifications on mammogram. Hum Pathol 1988;19:992–4.[Medline]
-
Bertrand AF, Dubois-Toussaint S, Lorimier G, Basle MF, Bertrand G. Breast microopacities of lead origin mimicking microcalcifications. J Radiol 1995;76:213–5. [In French.][Medline]
Top
Abstract
Introduction
and total volume of calcification