This Article Abstract Figures Only Full Text (PDF) Services Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via Google Scholar Google Scholar Articles by Meeson, S Articles by Ramsdale, M L Search for Related Content PubMed PubMed Citation Articles by Meeson, S Articles by Ramsdale, M L

Image features of true positive and false negative cancers in screening mammograms

¹ National Co-ordinating Centre for the Physics of Mammography, Department of Medical Physics, St. Luke's Wing, Royal Surrey County Hospital, Guildford GU2 7XX, ² Warwickshire, Solihull & Coventry Breast Screening Centre, Coventry and Warwickshire Hospital, Stoney Stanton Road, Coventry CV1 4FH, and ³ Jarvis Breast Screening Centre, 60 Stoughton Road, Guildford GU1 1LJ, UK

	Abstract

Top Abstract Introduction Methods Results Discussion Conclusions References

 
The location, tissue background and imaging characteristics of true positive and false negative screens of breast cancers have been studied. This data can aid decisions in optimizing the display of mammographic information with the objective of minimizing false negative screens. Screening mammograms for four groups of women were digitized; those with screen detected cancers, those with false negative interval cancers, and matched normals for both groups. The optical density (OD) distribution in the main breast region of each mammogram was determined. The OD in three regions of interest around the cancers was also measured. Cancer locations were mapped and warped onto a typical image to show their spatial distribution. Where a cancer was detectable by calcifications alone it had a relatively low probability of being a false negative interval cancer. The mean OD differences between the cancer and the cancer background region (excluding calcifications) were approximately a factor of two lower in dense breasts compared with other breast types. Poorly defined masses that became interval cancers had mean OD differences that were approximately a factor of 0.1 OD lower than those that were detectable by screening. 22% of false negative cancers were located near the chest wall edge of the mammograms compared with 10% of the true positives. The results indicate the importance of effectively displaying information in the lighter areas of the mammogram, corresponding to glandular tissues, with sufficient contrast for suspicious mammographic details to be detected. Where the mean OD differences between the cancer and its background region are low, as measured for some poorly defined masses, there is an increased risk of a false negative interval cancer. Particular attention should be given to the chest wall area of the film, especially in the lower retroglandular region, during routine screening.

	Introduction

Top Abstract Introduction Methods Results Discussion Conclusions References

 
Mammography is generally accepted to be the most efficient way of screening for breast cancers in women. Studies have shown that mammography screening reduces breast cancer mortality rates in women aged 50–69 years [1, 2], and in the UK screening has led to a reduction in breast cancer mortality [3]. However, it is well documented that breast screening with mammography as the primary modality for imaging is not a flawless test [4, 5]. The incidence of interval cancers for the early years of the UK screening programme have been reported by some regions as being somewhat higher than expected [6, 7]. Reviews of screening mammograms suggest that a considerable proportion of these interval cancers could have been detected at screening. Since the prognoses of interval cancers are likely to be worse than screen detected cancers, it is important to minimize the number of false negative screens. As a step towards this, the location, tissue background and imaging characteristics, including optical density (OD), of true positive and false negative screens of breast cancers were studied. This data can aid in optimizing display of mammographic information with the objective of minimizing false negative screens. The basic quantitative tools necessary for quantifying some important aspects of image quality in clinical mammograms have been developed and described in earlier work [8–10]. These make it possible to evaluate quality measures of individual mammograms using a quantitative rather than a subjective approach. These tools, and those developed specifically for this work, were used in conjunction with the visual gradings of radiologists to determine how interval cancers compared with screen detected cancers.

	Methods

Top Abstract Introduction Methods Results Discussion Conclusions References

 
Case selection
Screening mammograms for two groups of women were selected and digitized. The groups represented women with true positive screen detected cancers (TPSC), and women who presented with interval cancers following a false negative screen (FNIC). Matched normals were selected for both the screening mammograms of the TPSC and the FNIC. The case studies involved incident screening mammograms from breast screening centres where the standard practice was to use single view mammograms (mediolateral oblique (MLO) view). 90 cases in each group were identified and analysed. The TPSC were cancers positively identified following the assessment of a screening mammogram and confirmed by pathology. The breast screening centres were asked to identify women with screen detected cancer who had at least one previous screen where they were judged to be true negative. The FNIC were cancers diagnosed in the period between the last screen and the next scheduled screen, following an earlier negative or benign assessment of a mammogram. The matched normals were women whose mammograms showed no signs of cancer and who were asymptomatic at the time of the scheduled screen. Normals also had a subsequent screen in which they were judged to be true negatives. Cases were selected between early 1992 and 1998. The case selection process was randomized to prevent bias. As the smallest group, the FNIC were selected first. All the FNIC between the study dates were used from one of the centres. The cases from the second, larger centre were evenly selected from the cases occurring between the specified dates. Similarly, the TPSC were selected evenly from the cases occurring between the study dates. Matched normals were independently selected for the TPSC and FNIC groups. The matched normal was the next woman with a true negative mammogram that was screened on the same day as the woman in the cancer group, and who also had a true negative mammogram 3 years later at a subsequent screen. The method of selecting matched normals was chosen to ensure that target film density, imaging system, location and the process of mammogram review were all as closely matched as possible. A total of 360 case files were assessed.

Imaging systems used by breast screening centres
160 screening mammograms using Sterling Microvision C (Sterling UK, Stevenage, UK) film with Sterling Microvision Detail screens from 160 women attending screening on mobile and static units were collected from the first screening centre. Mammograms were taken using either a GE Medical Systems 600 TS Senix (GE Medical Systems Europe, Paris, France) or a Siemens Mammomat 2 (Siemens plc Medical Engineering, Bracknell, UK) operated under automatic exposure control (AEC). The films were processed using a Sterling T6 processor operating at 34°C on extended cycle using Sterling chemicals. The films were batch processed at the end of each day of screening.

206 mammograms using the Fuji UM-MA (HC) film (Fuji Photo Film (U.K.), London, UK) with Fuji UM fine screens from 200 women (includes three women with bilateral breast cancer) attending screening on mobile units were collected from the second screening centre. Mammograms were taken using a Siemens Mammomat 300 or Mammomat 2, operated under AEC. The films were processed using a Fuji FPM 3000 processor operating at a developer temperature of 34°C on extended cycle and using Photosol chemicals (Photosol Ltd., Basildon, UK). The films were batch processed at the end of each day of screening.

At both centres a tube potential of 28 kV was used for the majority of mammograms. However, 30 kV was selected for a small number of women with "large" breasts, at the discretion of the radiographers. A molybdenum/molybdenum target/filter combination was used together with the bucky grid. A target OD of 1.6 for a 4 cm thickness of polymethylmethacrylate was used and the correct operation of the AEC checked regularly. The general practice was to have the AEC chamber positioned at the chest wall for the quality control films and to move the chamber for the mammograms, at the discretion of the radiographers.

Mammogram digitization and radiological review
Screening mammograms were digitized using a Lumisys Lumiscan 150 HR laser scanner (Lumisys, Sunnyvale, CA) to enable them to be analysed electronically. Each mammogram was digitized with a pixel size of 210 µm. The scanner's detected signal was digitized into image pixel values. Grey scale step wedge films, which were produced with a sensitometer, were digitized with the mammograms to enable conversion from image pixel value to OD units [10].

Two expert film readers with considerable experience in mammography independently reviewed the mammograms. The mammograms were viewed in low ambient light conditions on a masked light box normally used for reviewing mammograms. The mammogram assessments included using the screening mammograms to grade the breast composition as either fatty, mixed density or dense. The cancer locations and outlines on the screening mammograms were recorded on transparent overlays for each of the TPSC and FNIC. Cancer details were also recorded and cross-referenced with the case reports to confirm cancer type and whether calcifications were involved. Radiological and assessment information from screening reports, interval cancer reports and pathology reports were recorded for each case. This included the cancer diameter, which was defined as the maximum dimension of the cancer measured by pathology [11]. The report data was combined with the results of mammogram image analysis to form a record in a database for subsequent analysis.

Image analysis
The image processing software package Aphelion (Amerinex Applied Imaging, Inc, Northampton, MA) was used to semi-automatically create regions of interest (ROIs) representing the pectoral muscle, main breast and skin edge. Images were filtered to remove calcifications, film emulsion "pick-off" and random noise [9, 10]. Three additional ROIs were manually drawn for each cancer around the cancer outline, in the central portion of the cancer avoiding complex edges, and a ring of local background breast tissue surrounding the cancer. These ROIs are shown schematically in Figure 1. All ROIs were subject to image analysis, including measuring the area, maximum, minimum and mean OD. Bilateral and multifocal cancers were treated separately.

View larger version (88K): [in this window] [in a new window]   Figure 1. Oblique view mammogram showing the three regions of interest selected for the analysis of cancer features.

View larger version (49K): [in this window] [in a new window]   Figure 2. Schematic showing the regions of interest used to assess distribution of cancer locations throughout the breast. The breast has been divided into five mutually exclusive anatomical regions of interest and one rectangular region of interest at the chest wall edge of the film.

	Results

Top Abstract Introduction Methods Results Discussion Conclusions References

View larger version (35K): [in this window] [in a new window]   Figure 3. Histogram of the mean optical density (OD) in the main breast regions of all the mammograms used in this study. (Labels on the mean OD axis represent the mid-points of the ranges used to sort the data.)

View larger version (22K): [in this window] [in a new window]   Figure 4. The relationship between the mean optical density (OD) in the main breast region and the mean OD in the whole cancer region of interest (ROI).

View larger version (34K): [in this window] [in a new window]   Figure 5. Histogram showing the distribution of mean optical density (OD) in the cancer background region of interest (ROI) of all the cancer group mammograms used in this study. (Labels on the Mean OD axis represent the mid-points of the ranges used to sort the data.)

Cancer diameter
The mean cancer diameters, as measured by pathology, are shown in Table 4 and Table 5. Table 6 shows the percentage of cancers divided into bands of 5 mm. The FNIC cancers at the time of diagnosis are generally larger than the TPSC cancers (group difference had a p-value of 0.010, ² test). Overall, the mean diameter±standard error in the mean for FNIC is 21.2±1.1 compared to 17.0±1.0 for TPSC.

View larger version (50K): [in this window] [in a new window]   Figure 6. Chart showing time lag between false negative screening mammogram and breast cancer being confirmed by mammography. (Labels on the lag time axis represent the upper limit of the ranges used to sort the data.)

View larger version (96K): [in this window] [in a new window]   Figure 7. Cancer centre locations shown for both the interval cancers following false negative screen (a) and true positive screen detected cancers (b) cases.

	Discussion

Top Abstract Introduction Methods Results Discussion Conclusions References

Both cancer groups had similar numbers of cancer features that were classified as masses without calcifications. More FNIC than TPSC were classified as masses with calcifications, but where a cancer could be detected by calcifications alone it had a relatively low probability of being a FNIC (p=0.010). Similar results were reported in previous studies [1, 13].

Figure 3 shows that the mean OD in the main breast region of the MLO view mammograms varied widely from mammogram to mammogram, with an overall mean of 1.6 OD. These variations in the mean OD can be explained by the way the AEC functions. The AEC system varies exposure to achieve the desired OD for only a relatively small area of the mammogram, i.e. that immediately above the detector. The mean OD in the main breast ROI will therefore vary depending upon the distribution of breast tissue. This contention is supported by similar results reported in earlier work [8, 9]. The mean OD in the cancer background ROIs also varied over a wide range of optical densities, with the majority of cancers being located against mean background ODs between 1.1 and 2.0.

For the TPSC and the two matched normal groups, the mean OD in the main breast ROI were related to the breast tissue type. The mean OD was lowest for the dense breasts and highest for the fatty breasts. This suggests that the AECs on the X-ray sets used operated imperfectly, leading to systematic underexposure for dense breasts. This may be a factor that reduces the effectiveness of mammograms in dense breasts. The mean OD in dense breasts was similar for both cancer groups. FNIC were more common in dense breasts than TPSC, in agreement with work relating breast density to interval cancer occurrence [14]. The local cancer contrast was lower in the FNIC fatty breasts than in the TPSC fatty breasts, but equivalent to the FNIC mixed density breasts. It is likely that identifying cancers in FNIC fatty breasts was made more difficult by this lower density contrast, accounting for why FNIC also occur in some types of fatty breasts. In both cancer groups the local cancer contrast was approximately a factor of two lower in the dense breasts than in other types of breast, making visualizing suspicious breast features more difficult. This lower contrast in dense breasts is likely to be one factor explaining the difficulty radiologists have in detecting lesions in this type of breast.

Spiculate masses were the most common feature in both cancer groups, with comparable local cancer contrast. A greater number of spiculate masses that became interval cancers were histological grade III rather than grade I or II compared with those detected by screening (p=0.021). A poorly defined mass is at increased risk of being a FNIC, which may be related to these features being more difficult to locate in screening mammograms. The local cancer contrast of poorly defined masses that became FNIC were approximately 0.1 OD lower than for TPSC, indicating that lower local contrast may have been the reason they were not detected by screening. This suggests that a number of cancers appear in screening mammograms with features that are at the borderline between being detectable and not detectable, i.e. poorly defined masses with a low local contrast. Thus variations in the presentation of image information that affect local contrast can be expected to affect cancer detection rates.

The recall rate for the FNIC normal group is based on a small number of cases, which has a large associated error. However, the recall rate expressed as a percentage of the number of women is comparable to the national rate of 3.1% for approximately the mid-point of the study period [15]. (NB. The national rate has since risen slightly to 3.9% for the period 1998/1999 [16]). Therefore in this study, five times more women were recalled for further assessment following their routine screening mammogram from the false negative group than normal. This may be related to the false negative group including women who were considered to have had either benign or minimal signs of abnormalities. On a retrospective review, many of the interval cancers occurred in approximately the same region of the breast that the initial referral investigated. Recently reported work suggests a number of procedures that may help when interpreting indeterminate microcalcifications and areas of architectural distortion [17].

In order to investigate cancer locations on false negative and true positive mammograms, a warping technique was used to map each screening mammogram to the selected typical breast image. The use of an affine warping transformation with manually obtained control points is sufficient where the aim is to pool the locations of cancer centres for a group of women. More sophisticated algorithms [18] have been proposed and developed, which are more accurate. However, these are designed to accurately register two mammograms of the same woman that have been exposed at different times to look for changes that may suggest abnormalities and changes in the breast. They will therefore need to be capable of dealing with variations due to breast compression and positioning.

The distributions of cancer centres for the two cancer groups are different. While there are similar numbers of cancer centres in the central and pectoral muscle ROIs this is not the case elsewhere in the breast. Small numbers of cancers were found in the skin edge regions of both cancer location images. Since cancers do appear in this region there are associated implications for film design and use, since it is important to effectively display the full extent of the breast on a mammogram, including the darker parts of the mammogram at the skin edge [19]. While many cancer centres are located in the upper retroglandular region, it is the lower retroglandular region where the greatest difference was identified. 10% of FNIC were located in the lower retroglandular region, whereas none of the TPSC were centred there. Comparing the chest wall band of the images with the rest of the breast image, 22% of the FNIC, compared with 10% of the TPSC, were located in the chest wall band (p=0.020). A figure of 10% for TPSC is in line with the 12% quoted following a study of cancers located at the posterior edge of the film [20]. Another study has also suggested that "missed lesions" are frequently found in the retroglandular region of the breast [13]. Both craniocaudal and mediolateral oblique views of the breast were used in that study. The definitions of the breast locations in that paper were not explicit and therefore may be different to those used in this study. However, notably more cancers were located towards the chest wall edge of the mammogram and behind the central area of dense glandular tissues in false negative than true positive cancer groups. In other work, cancer locations have been recorded for screen detected cancers in the UK [21, 22], but no comparison was made with false negative interval cancer locations. In those studies a mean breast was used as a reference and a sites-of-occurrence map was constructed using a new coordinate system based on the locations of the pectoral muscle and nipple. The warped lesion centroids were found to most likely be in the upper part of the breast. Further direct comparison between the cancer centroids in the two studies is not a simple task owing to the different methods of presenting the data. A novel technique has also been used to record the geographical distribution of breast cancers [4]. Cancer locations were recorded using a stylized diagram and a simple "point and click" technique. Whilst it did find that the distribution of false negative and screen detected cancers differed in the MLO view, and some clustering observed in this study was also seen in the published maps, a larger central distribution of cancers was recorded for both the false negative and true positive cases reported in our study than in the previous work. Overall, further investigation is required to determine if the results of this study are more widespread, but they do suggest that there is an imbalance in cancer detection in mammograms and that more attention should be given close to the chest wall edge of the film, and particularly the lower retroglandular region, during routine screening.

	Conclusions

Top Abstract Introduction Methods Results Discussion Conclusions References

 
Results indicate the importance of effectively displaying information in the lighter areas of mammograms, corresponding to glandular tissues. Mean OD in the main breast was generally related to breast tissue type, with mean OD lowest for dense breasts. Local cancer contrast was also related to breast tissue type. For both cancer groups local cancer contrast was approximately a factor of two lower in dense breasts than in other types of breast. When combined, these characteristics understandably make it more difficult to detect lesions in this type of breast.

Whilst a cancer detected by calcifications alone had a relatively low probability of being false negative, spiculate masses were the most common feature in both cancer groups. Poorly defined masses are at increased risk of being false negative interval cancers, which may be related to the features being more difficult to locate since poorly defined masses that became false negatives had local cancer contrasts approximately 0.1 OD lower than for true positives.

However, not all false negatives are necessarily related to OD. More false negatives were located in the lower retroglandular region of the mammograms than true positives, a typically more adipose region of the mammogram than the main breast. This suggests more attention should be given to this region of the mammogram during routine screening.

There appears to be further scope to optimize the presentation of mammographic images. Better contrast display, either using the new generation of very high contrast film/screen systems or through further technological developments such as digital mammography, may further reduce the incidence of false negative interval cancers in the UK NHS Breast Screening Programme.

	Acknowledgments

 
The authors would like to thank the clinical and support staff at the Warwickshire, Solihull & Coventry Breast Screening Centre and the Jarvis Breast Screening Centre in Guildford, who helped to identify and locate the cases used in this study.

	Footnotes

 
Current address for Dr S Meeson: Clinical Physics, Barts and The London NHS Trust, 56-76 Ashfield Street, The Royal London Hospital, London E1 1BB, UK.

Received for publication April 23, 2002. Revision received August 21, 2002. Accepted for publication September 24, 2002.

	References

Top Abstract Introduction Methods Results Discussion Conclusions References

 

1. Burrell HC, Sibbering DM, Wilson ARM, Pinder SE, Evans AJ, Yeoman LJ, et al. Screening interval breast cancers: mammographic features and prognostic factors. Radiology 1996;199:811–7.[Abstract/Free Full Text]
2. Nystrom L, Rutqvist LE, Wall S, Lindgren A, Lindqvist M, Ryden S, et al. Breast cancer screening with mammography: overview of Swedish randomised trials. Lancet 1993;341:973–8.[CrossRef][Medline]
3. Blanks RG, Moss SM, McGahan CE, Quinn MJ, Babb PJ. Effect of NHS breast screening programme on mortality from breast cancer in England and Wales, 1990–8: comparison of observed with predicted mortality. BMJ 2000;321:665–9.[Abstract/Free Full Text]
4. Brown M, Eccles C, Wallis MG. Geographical distribution of breast cancers on the mammogram: an interval cancer database. Br J Radiol 2001;74:317–22.[Abstract/Free Full Text]
5. Forrest P. Breast cancer screening: report to the Health Ministers of England, Wales, Scotland & Northern Ireland. London: HMSO, 1986.
6. Day N, McCann J, Camilleri-Ferrante C, Britton P, Hurst G, Cush S, Duffy S. Monitoring interval cancers in breast screening programmes: the East Anglian experience. Quality Assurance Management Group of the East Anglian Breast Screening Programme. J Med Screen 1995;2:180–5.[Medline]
7. Woodman CB, Threlfall AG, Boggis CR, Prior P. Is the three year breast screening interval too long? Occurrence of interval cancers in NHS breast screening programme's north western region. BMJ 1995;310:224–6.[Abstract/Free Full Text]
8. Meeson S, Young KC. Quantification of mammographic image quality in the NHSBSP. NHS Breast Screening Programme, 1999 (Occasional Report 99/2).
9. Meeson S, Young KC, Ramsdale ML, Wallis MG, Cooke J. Analysis of optical density and contrast in mammograms. Br J Radiol 1999;72:670–7.[Abstract]
10. Meeson S, Young KC, Rust A, Wallis MG, Cooke J, Ramsdale ML. Implications of using high contrast mammography X-ray film-screen combinations. Br J Radiol 2001;74:825–35.[Abstract/Free Full Text]
11. Vitak B, Olsen KE, Manson JC, Arnesson LG, Stal O. Tumour characteristics and survival in patients with invasive interval breast cancer classified according to mammographic findings at the latest screening: a comparison of true interval and missed interval cancers. Eur Radiol 1999;9:460–9.[CrossRef][Medline]
12. Russ JC. The image processing handbook (2nd edn). Boca Ration, FL: CRC Press, Inc., 1995.
13. Bird RE, Wallace TW, Yankaskas BC. Analysis of cancers missed at screening mammography. Radiology 1992;184:613–7.[Abstract/Free Full Text]
14. Mandelson MT, Oestreicher N, Porter PL, White D, Finder CA, Taplin SH, et al. Breast density as a predictor of mammographic detection: comparison of interval- and screen-detected cancers. J Natl Cancer Inst 2000;92:1081–7.[Abstract/Free Full Text]
15. NHS Breast Screening Programme Review 1995. Sheffield: NHS Breast Screening Programme, 1995.
16. Reducing the risk: Breast Screening Programme 2000. Sheffield: NHS Breast Screening Programme, 2000.
17. Burrell HC, Evans AJ, Wilson ARM, Pinder SE. False negative breast screening assessment. What lessons can we learn? Clin Radiol 2001;56:385–8.[CrossRef][Medline]
18. Sanjay-Gopal S, Chan H-P, Wilson T, Helvie M, Petrick N, Sahiner B. A regional registration technique for automated interval change analysis of breast lesions on mammograms. Med Phys 1999;26:2669–79.[CrossRef][Medline]
19. Meeson S, Young KC, Cooke J. Skin edge perception in mammograms: a comparison of two film screen combinations. Br J Radiol 2000;73:370–5.[Abstract]
20. Schmitt EL, Threatt B. Tumor location and detectability in mammographic screening. AJR 1982;139:761–5.[Abstract/Free Full Text]
21. Asquith J, Boggis C. Mammographic location of breast cancer. RAD Magazine 2000;26:21–2.
22. Caulkin S, Astley S, Asquith J, Boggis C. Sites of occurrence of malignancies in mammograms. In Karssemeijer N, Thijssen M, Hedriks J, Erning L, editors. Digital Mammography Nijmegen. Edited by Dordrecht, The Netherlands: Kluwer Academic Publishers, 1998:279–82.

This Article Abstract Figures Only Full Text (PDF) Services Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via Google Scholar Google Scholar Articles by Meeson, S Articles by Ramsdale, M L Search for Related Content PubMed PubMed Citation Articles by Meeson, S Articles by Ramsdale, M L