Computers in Biology and Medicine 40 (2010) 853–859

Contents lists available at ScienceDirect

Computers in Biology and Medicine

0010-48

doi:10.1

n Corr

Hippokr

Avenue

E-m

journal homepage: www.elsevier.com/locate/cbm

Downgrading BIRADS 3 to BIRADS 2 category using a computer-aidedmicrocalcification analysis and risk assessment systemfor early breast cancer

Georgia Giannakopoulou a,b,n, George M. Spyrou c,d, Argyro Antaraki c, Ioannis Andreadis c,f,Dimitra Koulocheri a,b, Flora Zagouri a, Afroditi Nonni e, George M. Filippakis a, Konstantina S. Nikita f,Panos A. Ligomenides c, George C. Zografos a

a Breast Unit, 1st Department of Propaedeutic Surgery, Hippokratio Hospital, School of Medicine, University of Athens, 114, Vas. Sofias Avenue, 11527 Athens, Greeceb Department of Radiology, Hippokratio Hospital, 114, Vas. Sofias Avenue, 11527 Athens, Greecec Informatics Laboratory, Academy of Athens, 4, Soranou Ephessiou St., 115 27, Greeced Biomedical Informatics Unit, Biomedical Research Foundation, Academy of Athens, 4, Soranou Ephessiou St., 115 27, Greecee Department of Pathology, School of Medicine, University of Athens, 75 Mikras Asias St., 115 27Athens, Greecef Department of Electrical and Computer Engineering, National Technical University of Athens, 9 Iroon Polytexneiou St., 157 80 Athens, Greece

a r t i c l e i n f o

Article history:

Received 5 August 2009

Accepted 22 September 2010

Keywords:

Early breast cancer

Microcalcifications

BIRADS 3

CAD system

SVABB

25/$ - see front matter & 2010 Elsevier Ltd. A

016/j.compbiomed.2010.09.005

esponding author at: Breast Unit, 1st Departm

atio Hospital, School of Medicine, University

, 11527 Athens, Greece. Tel./fax: +30 210722

ail address: ggiannakopoulou@gmail.com (G.

a b s t r a c t

This paper explores the potential of a computer-aided diagnosis system to discriminate the real benign

microcalcifications among a specific subset of 109 patients with BIRADS 3 mammograms who had

undergone biopsy, thus making it possible to downgrade them to BIRADS 2 category. The system

detected and quantified critical features of microcalcifications and classified them on a risk percentage

scale for malignancy. The system successfully detected all cancers. Nevertheless, it suggested biopsy for

11/15 atypical lesions. Finally, the system characterized as definitely benign (BIRADS 2) 29/88 benign

lesions, previously assigned to BIRADS 3, and thus achieved a reduction of 33% in unnecessary biopsies.

& 2010 Elsevier Ltd. All rights reserved.

1. Introduction

Breast cancer is the most frequently diagnosed cancer and isthe second leading cause of death in women. Mammographyremains the most effective imaging technique for early detectionof breast cancer [1]. Non-palpable breast lesions that appear onmammograms usually involve masses, calcifications, architecturaldistortion and asymmetric density. Microcalcifications occur in30–50% of breast cancer cases and constitute one of the mostimportant diagnostic markers in both benign and malignantlesions of the breast [2,3].

Unfortunately the precision of mammography, expressed as itspositive predictive value TP/(TP+FP), where TP is the number of truepositives and FP is the number of false positives, varies from 15–30%[4,5] and 10–30% of breast cancers may be missed. Therefore, doublereading is recommended [6], in order to achieve increasedsensitivity of mammography [7–9]. However, double reading is

ll rights reserved.

ent of Propaedeutic Surgery,

of Athens, 114, Vas. Sofias

1910.

Giannakopoulou).

not feasible for all mammographic centers and therefore computeraided diagnosis (CAD) systems have been introduced. CAD can bedefined as the diagnosis made by a radiologist who uses the outputfrom a computerized analysis of a medical image as a ‘‘secondopinion’’ in detecting and diagnosing lesions, with the final diagnosisbeing made by the radiologist (Fig. 1) [10–13]. Nevertheless, a CADsystem cannot replace radiologists as second-reader since its effectis inferior to double reading by radiologists [7–9].

A number of researchers have attempted to develop featureextraction and classification techniques for the detection andcharacterization of microcalcifications [7,14–34]. Many investi-gators have developed computerized methods to distinguishbetween malignant and benign clusters based on morphologi-cal-spatial features of individual calcifications or cluster as awhole [14,17,18,27–31] while others have introduced texturefeatures in their studies [15,16,23,32,33].

The American College of Radiology (ACR) [35,36] has developedthe Breast Imaging Reporting and Data System (BI-RADS) in anattempt to create a unique and common lexicon for use in reportingfindings on mammograms and a common language for unambigu-ously describing the level of suspicion and recommended follow-upfor a mammographic lesion. The BIRADS lexicon describes sevencategories: Category 0 recommends additional imaging (spot

Fig. 1. Mammogram diagnosis with and without the use of a CAD system. This figure (designed by our group) points out the improvement in diagnosis by using

information technologies.

G. Giannakopoulou et al. / Computers in Biology and Medicine 40 (2010) 853–859854

compression, special mammographic views or ultrasound), Category1 corresponds to a negative mammogram, Category 2 featuresbenign findings, Category 3 characterizes probably benign findingand a 6-month follow-up is suggested, Category 4 suggests biopsyfor patients with lesions classified as suspicious, Category 5 refers tolesions highly suggestive of malignancy that must be excised andfinally Category 6 represents histologically proven malignancy andthe patient must follow appropriate therapy.

Despite ACR reports that BIRADS category 3 lesions have lowprobability of malignancy (less than 2%) and suggests short-interval follow-ups [35–38], biopsies are performed occasionallyfor BIRADS 3 lesions, and this topic has been a matter ofconsiderable debate [39,40]. Biopsy is usually performed in asmall subset of patients with BIRADS 3 diagnosis when there isstrong demand from concerned patients or the physician insists,for reasons such as a future pregnancy or augmentation orreduction surgery, and the presence of malignancy in the same orthe opposite breast [41,42].

Frequently, in cases where the image findings are clearenough, the support of a CAD system is not so important.Nevertheless, in obscure cases such as BIRADS 3, where theevidence for a clear decision is inadequate, the contribution of aCAD system is expected to be important. On the other hand, CADsystems have their own classification difficulties apart from anytechnical or methodological limitations that they may have. Thekey question is to investigate whether a CAD has the potentialto support sufficiently a diagnosis of obscure cases, shiftingthe diagnostic index towards a correct and more stable value(e.g. BIRADS 2 or 4 categories).

Since, to our knowledge, there is no such a study, we presentour work that investigates the potentiality of a CAD system todiscriminate the real benign microcalcifications among a subset ofpatients with BIRADS 3 mammograms who underwent biopsy,thus giving radiologists the opportunity to downgrade fromBIRADS 3 to BIRADS 2 category in certain cases and therebyeliminate unnecessary biopsies in probably benign lesions. It is

obvious that any reduction in the unnecessary biopsies is animportant goal that, if achieved, could improve patient care byavoiding unnecessary interventions, reducing demands on thehealthcare system, and reducing the overall cost of caring forpatients. In order to investigate the possibility of unnecessarybiopsy reduction, we used a computer aided microcalcificationanalysis and risk assessment system that has been developed inour laboratory [18–21]. It quantifies critical features of micro-calcifications and classifies them according to their probability ofmalignancy. This system has been evaluated on cases covering awide range of BIRADS categories by verification with histologicalbiopsy findings and the evaluation results are very promising[18–21]. This CAD system can be made available to other researchgroups in a collaborative program regulated by established rules.

2. Methods and materials

The data-set for this study were obtained from 109 consecu-tive women (median age 51 years, range 36–90 years ) all with109 BIRADS 3 mammograms with microcalcifications clustersthat had undergone stereotactic vacuum-assisted breast biopsy(SVABB) in our Breast Unit because their family history wasstrongly positive or because the patient and the referringphysician expressed particular concern. A post-SVABB mammo-gram and an X-ray of the cores after the procedure confirmed theexcision of microcalcifications.

Tissue samples were formalin-fixed, paraffin-embedded andhaematoxylin–eosin stained. Histopathological examination ofthe samples revealed 6 malignancies, 88 benign cases and 15atypias. The malignancies comprised invasive ductal and in situductal or lobular carcinoma. The benign cases encompassedfibrocystic changes, apocrine metaplasia, sclerosing adenosis,epithelial hyperplasia, adenosis, fibrosis, mammary duct ectasia,papilloma, fibroadenoma, radial scar and diabetic mastopathy and

G. Giannakopoulou et al. / Computers in Biology and Medicine 40 (2010) 853–859 855

mucocele. The cases of atypia were either ductal or lobular (ADHor ALH).

Permission was obtained from the local institutional reviewboard for publication of the findings summarized in this study.

2.1. Computer-aided microcalcification analysis and risk assessment

system

Two independent radiologists using a computer-aided diag-nosis system analyzed retrospectively the 109 consecutivemammograms with microcalcifications clusters of these specialBIRADS 3 patients who had already been biopsied. The radiolo-gists were blinded to the results of the pathologic analysis. It isimportant to note that the radiologists in this study werescreening radiologists of similar diagnostic experience and thatthe mammograms were examined by the one or the otherdepending on the Breast Unit work load.

All mammograms were digitized at high resolution using aMicrotek Scanmaker 9800XL scanner with a resolution of 300 dotsper inch (dpi), 16-bit grayscale depth and .tiff uncompressed asthe image format. The image size for each image was 2116�2791while the size in megabytes was approximately 11 MB for eachimage.

The radiologists entered the digitized mammograms into the‘Hippocrates-mst’ computer-aided microcalcification analysis andrisk assessment system [18–21]. A region of interest (ROI)containing a microcalcification cluster was selected correspond-ing to the biopsy region. For this purpose, the radiologists wereprovided with the post-SVABB mammograms.

Detection of the microcalcifications in the ROI followed andimage analysis was performed for the microcalcification featureextraction and quantification (Fig. 2). The microcalcificationfeatures which the cad system considered were: (i) size,(ii) circularity, (iii) existence of dark center, (iv) level of bright-ness, (v) irregularity of shape, (vi) level of branching and

Fig. 2. The CAD form for the detection of microcalcifications. Microcalcifications are d

(vii) circumvolution. These features were calculated for everymicrocalcification of the cluster inside a sub region of the initialROI which was selected by the user. Subsequently, the systemuses a two-stage procedure for the risk assessment of the selectedROI. In the first stage, an empirical model comprising rules fromthe relevant literature and simulating a medical expert’s way ofclassifying a microcalcification is used in order to classify eachmicrocalcification and categorize it according to its risk using theseven characteristic feature space. In the second stage, the finalrisk assessment for the selected cluster is obtained from a threecharacteristic feature space including (i) the risk distribution ofmicrocalcifications, (ii) the number of microcalcifications at highrisk and (iii) the cluster polymorphism. The results from the‘Hippocrates-mst’ system are presented on a risk percentage scalefrom 0% to 100% (Fig. 3). According to its developers, a percentage,below 35% declares definitely benign findings, suggests follow-upand discourages biopsy [18]. A percentage from 35% to 55%declares a benign state but a biopsy test is suggested, while apercentage from 55% to 75% indicates malignancy with doubtsand a biopsy test is recommended. Finally, percentages higherthan 75% declare a definite malignancy and biopsy cannot beavoided.

It must be emphasized that CAD systems only make sugges-tions to the physician and it is the physician that makes the finaldecision whether a case has to be sent for a biopsy test or not.

2.2. Statistical analysis

In order to evaluate the performance of the computer-aidedmicrocalcification analysis and risk assessment system in the taskof predicting the probability of malignancy, we compared itsresults to the pathologic reports. In our analysis, the positiveresults included all malignant and atypical cases (ADH and ALH)and concerned lesions that forced surgical excision whereas thenegative results included benign cases only and thus concerned

etected in the ROI, classified and colored according to the risk estimation model.

Fig. 3. The CAD risk estimation form.

G. Giannakopoulou et al. / Computers in Biology and Medicine 40 (2010) 853–859856

lesions that did not require excision An atypical lesionwas considered as a positive result since the patient is at higherrisk of having or of developing breast carcinoma than the generalpopulation and recommendations suggest surgical excisionalbiopsy [43,44]. We produced the receiver operating characteristic(ROC) curve and calculated the area under the ROC curve(AUC) [45,46]. We also calculated the system’s sensitivity andspecificity. Statistical analysis was carried out with the statisticalsoftware R 2.10.1.

Fig. 4. Distribution of the estimated risk for malignant, atypical and benign cases.

3. Results

The ‘Hippocrates-mst’ system classified all 109 cases andsuggested biopsy for 76 women with estimated risk greater than35% and acquitted of pathology 33 women with values below 35%and discouraged biopsy [18]. The distribution of the estimatedrisk for the malignancies, for the benign cases and for the cases ofatypias is shown in Fig. 4.

The analysis included all 109 cases and analyzed lesions thatneeded surgical excision (malignant and atypical ones: positivebiopsy result) as well as lesions that did not require excision(benign: negative biopsy result) (Table 1).The group of malignant-atypias had a mean estimated risk 57.6% (SD: 24.5) with 95%confidence interval (46.5–68.7) whereas the benign lesionsshowed a mean estimated risk 50.4% (SD: 26) with 95% confidenceinterval (44.9–55.8).

Table 2 shows the contingency table of CAD system resultsand biopsy results (positive and negative). The system indicatedbiopsy for 17 of the 21 lesions that should be excised(sensitivity¼81% with 95% CI: 58–94). All cancers wereidentified but 4 atypical lesions were misread and follow-upwas suggested instead. The system did not refer for biopsy 29 out

of the 88 benign cases but incorrectly suggested biopsy for 59benign lesions giving a specificity of 33% (95% CI: 23–43).

Table 3 presents the distribution of ADH/ALH cases on the CADrisk percentage scale as well as their summary statistics. The CADsystem referred 11 cases of atypia for biopsy (73.33%). It isnoteworthy that almost half of the atypias had an estimated riskof over 55%.

Using the results from Table 1, the receiver operatingcharacteristic (ROC) curve of Fig. 5 was produced. The area under

Table 1Frequencies of computer-aided microcalcification analysis and risk assessment system results in lesions that should be surgically excised (malignant+atypical) or not

(benign lesions).

Biopsy results System results (%)

N Min Max Mean SD 95% CI

Benign cases(no surgical excision

required)

N1¼88 0 100 50.4 26.0 (44.9–55.8)

Malignant and atypical cases (lesions

that required surgical excision)

N2¼21 7.5 95.4 57.6 24.5 (46.5–68.7)

Overall summary N¼109 0 100 51.8 25.8 (47.0–56.7)

N: number, SD: standard deviation, 95% C.I: 95% confidence interval.

Table 2Contingency matrix of biopsy and CAD results (lesions that should be surgically

excised (malignant+atypical) or not (benign lesions)).

CAD results Biopsy results Total

Malignant and atypical cases Benign cases

CAD+ 17 59 76

CAD� 4 29 33

Total 21 88 109

CAD+: positive CAD results (estimated risk Z35%-biopsy suggestion).

CAD�: negative CAD results (estimated risk o35%-follow-up).

G. Giannakopoulou et al. / Computers in Biology and Medicine 40 (2010) 853–859 857

the ROC curve (AUC) was 0.58 with 95% confidence interval of(0.45–0.71).

4. Discussion

This is the first study of the capability of a CAD system todiscriminate the real benign microcalcification clusters amongBIRADS 3 mammograms and thus downgrade them to BIRADS 2category and discourage biopsy. It is important to emphasize thatthese patients represent a small subset of BIRADS 3 patients thatconsidered biopsy.

A number of studies have focused on the usage of a CADsystem in an effort to assist radiologists to characterize variousbreast lesions on mammograms and make recommendations.Buchbinder et al. [47] performed a computer aided classificationfor masses assigned to BIRADS 3 category and upgraded 38/42malignant lesions initially assigned to BIRADS 3 to BIRADS 4category (sensitivity 90%) and on the other hand assigned 12/64benign lesions as BIRADS 2. Marx et al. [48], using a CAD system innon-palpable masses and calcifications, detected 32/36 cancercases, and discouraged biopsy for 11.3% of normal cases and 5.9%of benign case, but overall did not change sensitivity and slightlyincreased specificity . Finally, Burnside et al. [28] achieved an AUCvalue of 0.919 with a 10% threshold for biopsy; according toBayesian network probability estimates 34 patients with mam-mographic microcalcifications would not have undergone biopsy(39%), without missing any breast cancer.

It is important to note that single reading with CAD has beenproven to be less effective than double reading by radiologists, asthere is no difference in cancer detection rates but double readingwith arbitration shows a significantly better recall rate. Addition-ally, CAD increases the biopsy rate of healthy women [8,9].

In our study, the ‘Hippocrates-mst’ computer-aided micro-calcification analysis and risk assessment system correctlyidentified all malignancies and proposed biopsy. Furthermore,the analysis that we performed concerned lesions, which requiredsurgical excision (malignant and atypical ones) and lesions, which

did not require excision and should be followed-up (benign). Thisanalysis was performed based on recommendations that suggestsurgical excisional biopsy of atypical lesions found in minimallyinvasive biopsies because these patients are often at higher risk ofhaving or of developing breast carcinoma [43,44]. We can see thatthe system did not demonstrate stable behavior for the cases ofADH or ALH as it suggested biopsy for 73.33% but 4 of them weremissed. We must note that almost half of the atypias wereestimated as having over 55% risk, indicating malignancy withdoubts. On the other hand, all cancers appeared in the region over35%, with the majority in the 55–75% range.

Furthermore, the CAD system downgraded 29 out of 88 benigncalcification clusters, previously assigned as BIRADS 3, to BIRADS2 category, and so achieved a reduction of 33% in unnecessarybiopsies. However, the system recommended biopsy for theremaining 59 benign microcalcification clusters and consequentlyhad low specificity. The large number of false positive results ofthe CAD system and no false negatives (concerning cancer),indicated that a positive test is in itself poor for confirming cancerand further research is called for this problem. However, anegative result was very good at suggesting that a woman did nothave cancer, and at this primary level of prevention, it correctlyidentified 33% of those who did not have cancer (i.e. specificity).

Nevertheless, the system’s specificity would be more appro-priately compared to the physician’s decision for biopsy referrals,since all patients in the present study had been referred forbiopsy. The CAD system behavior demonstrated satisfactoryperformance regarding its sensitivity, even though the analysiswas applied in a target group with higher probability ofmalignancy for BIRADS 3 category (6/109¼5.5%) than theexpected (ACR reports probability of malignancy o2%) [35,36].The distribution of benign cases for which the estimated risk wasbelow the CAD boundary of malignancy (55%) presented decreas-ing tendency as risk percentage increased. Additionally, thedistribution of the 43 benign cases estimated as malignancies(CAD result 455%), remained higher in the area of malignancywith doubts (55–75%). It would be interesting to note that thebenign cases with an increased cancer estimated risk, concernedmostly pleomorphic or amorphous clusters of microcalcifications,which according to pathology reports corresponded to sclerosingadenosis.

In Fig. 5 the area under the empirical ROC curves formammography was 0.58. This implies that, if we select randomlytwo patients-one with a lesion that needs surgical excision(malignancy or atypia) and one without—the probability is 0.58that the patient with the lesion that requires surgical excision,will have the more suspicious result [45]. We can also be 95% surethat the 95% CI in Fig. 5 includes the true values of the AUC. Thelower bound of the 95% CI of the AUC is 0.45, therefore the test ismarginally better than making the diagnostic decision based onpure chance. The special case of BIRADS 3 lesions and the limitednumber of women undergoing biopsy result in a small proportion

Table 3Distribution of ADH/ALH cases in the CAD risk percentage scale.

CAD risk percentage scale of atypias

0–35% 35–55% 55–75% 75–100% Total

26.7 (N¼4) 26.7 (N¼4) 20.0 (N¼3) 26.7 (N¼4) 100.0 (N¼15)

Summary statistics

Min Max Mean SD

7.5 95.4 55.2 28.1

The CAD system suggested biopsy (risk estimation: Z35%) to 11 cases of atypia (73.33%) but misread 4 of them and proposed follow-up (risk estimation: o35%).

We must note that 7/15 (46.67%) of ADH/ALH cases had an estimated risk of over 55%.

Fig. 5. ROC curve for CAD performance in selected BIRADS 3 cases (lesions that

require surgical excisional biopsy/lesions that require follow-up). Analysis

included lesions that require surgical excisional biopsy (malignant andatypical

lesions) and lesions that require follow-up (benign) N=109, sensitivity: 81%

(95% CI: 58-94), specificity: 33% (95% CI: 23-43), AUC=0.58 (95% CI: 0.45-0.71),

N: number, 95% C.I: 95% confidence interval.

G. Giannakopoulou et al. / Computers in Biology and Medicine 40 (2010) 853–859858

of malignancies as well as conservative AUC and 95% confidenceinterval estimations. We plan to carry out further investigationsusing larger samples which may lead to better performance andfirmer conclusions.

In our previously published work, we performed imageanalysis using our CAD system on mammograms with micro-calcification clusters in all BIRADS categories and an achieved AUCvalue of 0.856 [18]. Consequently, we assume that the compli-cated nature of BIRADS 3 category along with the selected specialsubset led to the decreased AUC value. Nevertheless, the AUC thatwe obtained in this study, although it is not close to 1(the idealcase), gives us good evidence for the usefulness of CAD,considering the difficulty of dealing with the BIRADS 3 category.

5. Conclusions

The ‘Hippocrates-mst’ computer-aided microcalcification ana-lysis and risk assessment system has been tested in clinicalpractice to determine whether it can be used as a second opinion

by radiologists in managing BIRADS category 3 mammogramswith calcification clusters. A limitation of the study was that thesample of patients consisted only of the small subset of BIRADS 3cases who were referred for biopsy. The system detectedsuccessfully all cancer cases and recommended surgical excisionalbiopsy and, additionally, confirmed that a woman with CAD riskpercentage less than the given threshold had no cancer at thepresent time. Unfortunately the system did not predict success-fully the presence of atypia that could lead to cancer as itsuggested biopsy for 73.33% but missed 4 of them, thus leading toa lower sensitivity of 81% with 95% CI (58–94). However, itacquitted of biopsy 29 out of 88 benign calcification clusters, andachieved a reduction of 33% in the unnecessary biopsies. Inconclusion, the CAD system gave promising results in thedirection of downgrading the assignment from BIRADS 3 toBIRADS 2, thus accomplishing elimination of unnecessary biopsiesand furthermore eradication of patients’ anxiety. Nevertheless,examining the CAD performance on a larger data-set of BIRADS 3cases as well as the fine-tuning of CAD parameters concerningoptimization of risk estimation are among our future plans.

6. Summary

Mammography screening widely offers the opportunity ofdetecting early breast cancer. However, 10–30% of breast cancersmay be missed and a second reading is recommended. Doublereading succeeds an increased sensitivity of mammography.Nevertheless, double reading is not feasible for all mammographiccenters. Therefore, computer aided diagnosis (CAD) systems wereintroduced and aimed at being a second-look aid for radiologists.Our study investigates the potential of a computer-aided micro-calcification analysis and risk assessment system to discriminatethe real benign microcalcifications among a special group ofpatients with BIRADS 3 mammograms who were referred forbiopsy. We analyzed 109 consecutive BIRADS 3 mammogramswith SVABB-proven results. The computer aided diagnosis (CAD)system detected and quantified critical features of microcalcifica-tions and classified them according to their probability ofmalignancy on a risk percentage scale. The system identifiedand proposed biopsy for all cancer cases. Unfortunately it missed4 atypical lesions which should be referred for surgical excisionbecause of their high probability of having or developingmalignancy, causing the system’s sensitivity to decrease to 81%with 95% CI (58–94). However, the system downgraded 29 out of88 benign calcification clusters, previously assigned to BIRADS 3,to BIRADS 2 category, and so achieved a reduction of 33% inunnecessary biopsies. The results justified that the CAD systemhas the potential to improve radiologist’s performance in thedetection and evaluation of microcalcifications, as it revealed themalignant clusters among our small subset of BIRADS 3

G. Giannakopoulou et al. / Computers in Biology and Medicine 40 (2010) 853–859 859

mammograms and unveiled 33% of the real benign cases,consequently downgrading them from BIRADS 3 to BIRADS 2category. Therefore, with further development and investigationusing larger samples it could be used to detect cancerous lesionsand eliminate unnecessary biopsies of benign lesions.

Conflict of interest statement

None declared.

Acknowledgements

We would like to thank the President of the Hellenic Antic-ancer Institute, E. Frangoulis, the Head of the Radiology Depart-ment of Hippokratio Hospital of Athens, Spyridon X. Kavvadiasand the Professor of Patholgy of Medical School of University ofAthens, Efstratios Patsouris for their contribution to the organiza-tion of the breast unit.

References

[1] A. Jemal, R. Siegel, E. Ward, Y. Hao, J. Xu, T. Murray, et al., Cancer statistics, CACancer J. Clin. 58 (2) (2008) 71–96.

[2] K. Kerlikowske, D. Grady, J. Barclay, E.A. Sickles, A. Eaton, V. Ernster, Positivepredictive value of screening mammography by age and family history ofbreast cancer, JAMA 270 (1993) 2444–2450.

[3] A.J. Evans, A.R. Wilson, H.C. Burrell, I.O. Ellis, S.E. Pinder, Mammographicfeatures of ductal carcinoma in situ (DCIS) present on previous mammo-graphy, Clin. Radiol. 54 (1999) 644–646.

[4] D.D. Adler, M.A. Helvie, Mammographic biopsy recommendations, Curr.Opinion Radiol. 4 (1992) 123–129.

[5] D.B. Kopans, The positive predictive value of mammography, AJR 158 (1991)521–526.

[6] A.S. Majid, E.S. de Paredes, R.D. Doherty, N.R. Sharma, X. Salvador, Missedbreast carcinoma: pitfalls and pearls, Radiographics 23 (4) (2003) 881–895.

[7] A. Malich, D.R. Fischer, J.C.A.D. Bottcher, For mammography: the technique,results, current role and further developments, Eur. Radiol. 16 (2006)1449–1460.

[8] P. Taylor, H.W. Potts, Computer aids and human second reading asinterventions in screening mammography: two systematic reviews tocompare effects on cancer detection and recall rate, Eur. J. Cancer 44 (6)(2008) 798–807 Epub 2008 March 18.

[9] M. Noble, W. Bruening, S. Uhl, K. Schoelles, Computer-aided detectionmammography for breast cancer screening: systematic review and meta-analysis, Arch. Gynecol. Obstet. 279 (6) (2009) 881–890 Epub 2008 Nov 21.

[10] R.F. Brem, J. Baum, M. Lechner, S. Kaplan, S. Souders, L.G. Naul, et al.,Improvement in sensitivity of screening mammography with computer-aided detection: a multiinstitutional trial, AJR 181 (2003) 687–693.

[11] S.M. Astley, F.J. Gilbert, Computer-aided detection in mammography, Clin.Radiol. 59 (2004) 390–399.

[12] M.L. Giger, Computerized analysis of images in the detection and diagnosis ofbreast cancer, Semin. Ultrasound CT MRI 25 (2004) 411–418.

[13] J.M. Ko, M.J. Nicholas, J.B. Mendel, P.J. Slanetz, Prospective assessment of thecomputer-aided detection in interpretation of screening mammography, AJR187 (2006) 1483–1491.

[14] Y. Jiang, R.M. Nishikawa, D.E. Wolverton, C.E. Metz, M.L. Giger, R. Schmidt,et al., Malignant and benign clustered microcalcifications: automated featureanalysis and classification, Radiology 198 (1996) 671–678.

[15] H.P. Chan, B. Sahiner, N. Petric, M.A. Heavie, K.L. Lam, D.D. Adler, et al.,Computerized classification of malignant and benign microcalcifications onmammograms: texture analysis using an artificial neural network, Phys. Med.Biol. 42 (1997) 549–567.

[16] H.P. Chan, B. Sahiner, K.L. Lam, N. Petric, M.A. Helvie, M.M. Goodsitt, et al.,Computerized analysis of mammographic microcalcifications in morpholo-gical and texture feature spaces, Med. Phys. 25 (1998) 2007–2019.

[17] R. Nakayama, Y. Uchiyama, R. Watanabe, S.H. Katsuragawa, K. Namba, K. Doi,CAD scheme for histological classification of clustered microcalcifications onmagnification mammograms, Med. Phys. 31 (4) (2004) 789–799.

[18] G. Spyrou, S. Kapsimalakou, A. Frigas, K. Koufopoulos, S. Vassilaros,P. Ligomenidis, Hippocrates-mst: a prototype for computer-aided micro-calcification analysis and risk assessment for breast cancer, Med. Bio. Eng.Comput. 44 (2006) 1007–1015.

[19] G. Spyrou, M. Nikolaou, M. Koussaris, A. Tsibanis, S. Vassilaros,P. Ligomenides, A system for Computer Aided Early Diagnosis of breastcancer based on microcalcifications analysis, in: Proceedings of the FifthEuropean Systems Science Congress, October 2002, Crete, Res-Systemica 2,2002.

[20] G. Spyrou, K. Koufopoulos, S. Vassilaros, P. Ligomenides, Computer aidedimage analysis and classification schemes for the early diagnosis of breastcancer, Hermis Int. J. Comput. Math. Appl. 4 (2003) 175–181.

[21] A. Frigas, S. Kapsimalakou, G. Spyrou, K. Koufopoulos, S. Vassilaros,A. Chatzimichael, et al., Evaluation of a breast cancer computer aideddiagnosis system, Stud. Health Technol. Inform. 124 (2006) 631–636.

[22] D.L. Thiele, C. Kimme-Smith, T.D. Johnson, M. McCombs, L.W. Bassett, Usingtissue texture surrounding calcification clusters to predict benign vsmalignant outcomes, Med. Phys. 23 (1996) 549–555.

[23] A.P. Dhawan, Y. Chitre, C. Kaiser-Bonasso, M. Moskowitz, Analysisof mammographic microcalcifications using gray-level image structurefeatures, IEEE. Trans. Med. Imaging 15 (1996) 246–259.

[24] L. Shen, R.M. Rangayyan, J.E.L. Desautels, Application of shape analysisto mammographic calcifications, IEEE. Trans. Med. Imaging 13 (1994)263–274.

[25] M.A. Gavrielides, J.Y. Lo, Floyd CE Jr, Parameter optimization of a computer-aided diagnosis scheme for the segmentation of microcalcification clusters inmammograms, Med. Phys. 29 (4) (2002) 475–483.

[26] S. Lee, C. Lo, C. Wang, P. Chung, C. Chang, C. Yang, et al., A computer-aideddesign mammography screening system for detection and classification ofmicrocalcifications, Int. J. Med. Inf. 60 (1) (2000) 29–57.

[27] D. Betal, N. Roberts, G.H. Whitehouse, Segmentation and numerical analysisof microcalcifications on mammograms using mathematical morphology,British J. Radiol. 70 (1997) 903–917.

[28] E.S. Burnside, D.L. Rubin, J.P. Fine, R.D. Shachter, G.A. Sisney, W.K. Leung,Bayesian network to predict breast cancer risk of mammographic micro-calcifications and reduce number of benign biopsy results: initial experience1, Radiology 240 (3) (2006) 666–673.

[29] I. Leichter, R. Lederman, S.S. Buchbinder, P. Bamberger, B. Novak, S. Fields,Computerized evaluation of mammographic lesions: what diagnostic roledoes the shape of the individual microcalcifications play compared with thegeometry of the cluster? AJR 182 (2004) 705–712.

[30] I. Leichter, R. Lederman, P. Bamberger, B. Novak, S. Fields, S.S. Buchbinder, Theuse of an interactive software program for quantitative characterization ofmicrocalcifications on digitized film-screen mammograms, Invest. Radiol.34 (6) (1999) 394–400.

[31] M.S. Soo, E.L. Rosen, J.Q. Xia, S. Ghate, J.A. Baker, Computer-aided detection ofamorphous calcifications, AJR 184 (2005) 887.

[32] A. Papadopoulos, D.I. Fotiadis, A. Likas, Characterization of clusteredmicrocalcifications in digitized mammograms using neural networks andsupport vector machines, Artif. Intell. Med. 34 (2005) 141–150.

[33] H.S. Sheshadri, A. Kandaswamy, Computer aided decision system for earlydetection of breast cancer, Indian J. Med. Res. 124 (2006) 149–154.

[34] J.J. Fenton, S.H. Taplin, P.A. Carney, L. Abraham, E.A. Sickles, C. D’Orsi, et al.,Influence of computer-aided detection on performance of screening mam-mography, N. Engl. J. Med. 356 (14) (2007) 1399–1409.

[35] American College of Radiology (ACR), in: Illustrated Breast Imaging Reportingand Data System (BI-RADS), 4th edn., American College of Radiology, Reston,VA, 2003.

[36] S. Obenauer, K.P. Hermann, E. Grabbe, Applications and literature review ofthe BI-RADS classification, Eur. Radiol. 15 (2005) 1027–1036.

[37] X. Varas, F. Leborgne, J. Leborgne, Nonpalpable probably benign lesions: roleof follow up mammography, Radiology 184 (1992) 409–414.

[38] E. Sickles, Periodic mammographic follow up of probably benign lesions:results of 3184 consecutive cases, Radiology 79 (1991) 463–468.

[39] E.A. Sickles, S.H. Parker, Appropriate role of core breast biopsy in themanagement of probably benign lesions, Radiology 188 (1993) 315.

[40] W.W. Logan-Young, J.A. Janus, S.V. Destounis, N.Y. Hoffman, Appropriate roleof core breast biopsy in the management of probably benign lesions,Radiology 190 (1994) 313–314.

[41] K. Kerlikowske, R. Smith-Bindman, B.M. Ljung, D. Grady, Evaluation ofabnormal mammography results and palpable breast abnormalities, Ann.Intern. Med. 139 (2003) 274–285.

[42] American College of Surgeons and American College of Radiology, Physicianqualifications for stereotactic breast biopsy: a revised statement, Bull. Am.Coll. Surg. 83 (1998) 30–33.

[43] C. Balu-Maestro, F. Ettore, C. Chapellier, I. Peyrottes, P. Leblanc-Talent, Whenshould caution be used with regards to histopathologic findings of imaging-guided breast micro- and macro-biopsies? J. Radiol. 87 (3) (2006) 265–273.

[44] R.F. Brem, M.C. Lechner, R.J. Jackman, J.A. Rapelyea, W.P. Evans, L.E. Philpotts,J. Hargreaves, S. Wasden, Lobular neoplasia at percutaneous breast biopsy:variables associated with carcinoma at surgical excision, AJR Am.J. Roentgenol. 190 (3) (2008) 637–641.

[45] A.N. Obuchowski, Receiver operating curves and their use in radiology,Radiology 229 (2003) 3–8.

[46] Ho S Park, Mo J Good, C.H. Jo, Receiver operating characteristic (ROC) curve:practical review for radiologists, Korean J. Radiol. 5 (2004) 11–18.

[47] S.S. Buchbinder, I.S. Leichter, R.B. Lederman, B. Novak, P.N. Bamberger,M. Sklair-Levy, et al., Computer-aided classification of BI-RADS category 3breast lesions, Radiology 230 (2004) 820–823.

[48] C. Marx, A. Malich, M. Facius, U. Grebenstein, D. Sauner, S.O.R. Pfleiderer,et al., Are unnecessary follow-up procedures induced by computer-aideddiagnosis (CAD) in mammography? Comparison of mammographic diagnosiswith and without use of CAD, EJR 51 (2004) 66–72.