Technical Report

INbreast:
Toward a Full-field Digital Mammographic Database

^s C. Moreira, MSc student, Igor Amaral, MSc, Ine

Ine ^s Domingues, MSc,
 nio Cardoso, MD, Maria Joa
Anto ~o Cardoso, PhD, Jaime S. Cardoso, PhD

Rationale and Objectives: Computer-aided detection and diagnosis (CAD) systems have been developed in the past two decades to
assist radiologists in the detection and diagnosis of lesions seen on breast imaging exams, thus providing a second opinion. Mammo-
graphic databases play an important role in the development of algorithms aiming at the detection and diagnosis of mammary lesions.
However, available databases often do not take into consideration all the requirements needed for research and study purposes. This
article aims to present and detail a new mammographic database.
~o [CHSJ],
Materials and Methods: Images were acquired at a breast center located in a university hospital (Centro Hospitalar de S. Joa
Breast Centre, Porto) with the permission of the Portuguese National Committee of Data Protection and Hospital’s Ethics Committee.
MammoNovation Siemens full-field digital mammography, with a solid-state detector of amorphous selenium was used.
Results: The new database—INbreast—has a total of 115 cases (410 images) from which 90 cases are from women with both breasts
affected (four images per case) and 25 cases are from mastectomy patients (two images per case). Several types of lesions (masses, calci-
fications, asymmetries, and distortions) were included. Accurate contours made by specialists are also provided in XML format.
Conclusion: The strengths of the actually presented database—INbreast—relies on the fact that it was built with full-field digital mammo-
grams (in opposition to digitized mammograms), it presents a wide variability of cases, and is made publicly available together with precise
annotations. We believe that this database can be a reference for future works centered or related to breast cancer imaging.
Key Words: Mammographic database; CAD; computer-aided detection; computer-aided diagnosis.
ªAUR, 2012

A
ccording to the World Health Organization, breast mammographic examinations are performed annually on
cancer was responsible for approximately 519,000 asymptomatic women to detect early, clinically unsuspected
deaths in 2004: 16% of all cancer incidence among lesions. The age at which mass screening mammography is
women. In 2008, it was the most common form of cancer generally recommended in the United States is 40 (5). In
and cancer related death in women worldwide (1). In Portugal, Europe, screening at 40 to 50 years old is still not consensual
1500 women die every year from breast cancer, whereas in the (6). However, in women with genetic mutations or significant
European Union it is responsible for one in every six deaths family history of breast cancer, screening should start earlier,
from cancer in women (2). For this reason, early detection usually 10 years earlier than the age of diagnosis of the youngest
and diagnosis of breast cancer is essential to decrease its associ- relative (never before 25) (5).
ated mortality rate. Therefore, mass screening is recommended Mammography comprehends the recording of two views
by the medical community (2,3). for each breast: the craniocaudal (CC) view, which is a top
X-ray mammography is currently considered the best to bottom view, and a mediolateral oblique (MLO) view,
imaging method for breast cancer screening and the most which is a side view (Fig 1) (6). The images can be acquired
effective tool for early detection of this disease (4). Screening on x-ray film, such as a film-screen mammogram, or in digital
format, such as with digital mammography (full-field digital
mammography [FFDM] and computed radiography) (7).
Acad Radiol 2012; 19:236–248 When radiologists examine mammograms, they look for
From the Faculdade de Medicina, Alameda Prof. Herna ^ni Monteiro, specific abnormalities (8). The most common findings seen
Universidade do Porto, 4200-319, Porto, Portugal (I.C.M., M.J.C.); Hospital
de Sa ~o Joa~o, Porto, Portugal (I.C.M., A.C.); INESC Porto, Porto, Portugal on mammography are masses, calcifications, architectural
de
(I.C.M., I.A., I.D., M.J.C., J.S.C.); Escola Superior de Tecnologia da Sau distortion of breast tissue, and asymmetries when comparing
do Porto, Polite cnico do Porto, Portugal (I.C.M.); Faculdade de Engenharia,
Universidade do Porto, Portugal (I.D., J.S.C.); Champalimaud Cancer
the two breasts and the two views. To standardize the termi-
Center, Breast Unit, Lisbon, Portugal (M.J.C.). Received May 4, 2011; nology of the mammographic report, the assessment of find-
accepted September 26, 2011. Funded by the Portuguese Agency for ings and the recommendation of action to be taken, the
Innovation (ADI), through project QREN reference 3472 ‘‘Semantic PACS.’’
We acknowledge the participation of Emılio de Azevedo Campos on the American College of Radiology (ACR) has developed the
person of Pedro Cardoso. We also thank Doctor Francisco Pimentel for Breast Imaging Reporting and Data System (BI-RADS) scale
helpful scientific comments. Address correspondence to: I.C.M. e-mail:
icm@estsp.ipp.pt
(9). Based on level of suspicion, the previously mentioned
ªAUR, 2012
lesions can be placed into one of six BI-RADS categories:
doi:10.1016/j.acra.2011.09.014 category 0, exam is not conclusive; category 1, no findings;

236
Academic Radiology, Vol 19, No 2, February 2012 INBREAST: TOWARD AN FFDM DATABASE

Figure 1. Mammogram examples:

(a) craniocaudal (CC) view of the right
breast; (b) CC view of the left breast;
(c) mediolateral oblique (MLO) view of the
right breast; (d) MLO view of the left breast.

category 2, benign findings; category 3, probably benign find- lesions. To decrease this rate, computer-aided detection and
ings; category 4, suspicious findings; category 5, a high prob- diagnosis (CAD) systems have been developed in the past
ability of malignancy; and category 6, proved cancer (Table 1). two decades to assist the radiologists in the interpretation of
In case of categories 4 and 5, a biopsy is needed to exclude or the medical images (14,15). To design, test, and tune such
confirm malignancy (10). Other important characteristic computational systems, researchers demand a large number
referred by the ACR is the breast composition tissue, related of mammograms (16). These datasets need to be digital, so
to the breast density shown in x-rays. There are four cate- if the images are acquired on x-ray film, they have to be digi-
gories ranging from 1, for low density (fatty tissue), to 4, for tized (15). Therefore, mammographic databases play an
very high density (dense tissue) (11). important role in the development of algorithms aiming at
Several studies (8,12,13) concluded that detection of detecting and diagnosing lesions. They are also important to
suspicious findings by radiologists is a repetitive and allow comparison of results from different studies (17–19).
fatiguing task, leading to a 10%–30% rate of undetected A different application is the use of database images

237
MOREIRA ET AL Academic Radiology, Vol 19, No 2, February 2012

TABLE 1. Breast Imaging Reporting and Data System unless it is a case from a patient with one breast only from
Assessment Categories previous mastectomy. Image acquisition should be adequate
in terms of patient positioning, x-ray exposure, and with an
Category Description
absence of image blur due to patient motion.
0 Needs additional imaging evaluation and/or prior
mammograms for comparison
Ground Truth
1 Negative
2 Benign finding(s) Biopsy proof for all cases should be available. Annotations
3 Probably benign finding(s). Short-interval follow-up should include the ‘‘ground truth’’ (GT) concerning
is suggested. cytology/histology for all cases, the location and boundaries
4 Suspicious anomaly. Biopsy should be considered.
of the lesion with the outline marking performed by an
5 Highly suggestive of malignancy. Appropriate
imaging specialist.
action should be taken.
6 Biopsy proven malignancy
Associated Information

for teaching and training students in this specific medical Clinical history such as age, family history, and previous biop-
field. sies can be useful for studying subpopulations of women
The common practice in the development of CAD algo- (eg, women <50 years of age), and it may improve the perfor-
rithms has been the use of private sets of mammograms to mance of a CAD scheme by incorporating nonradiographic
design and evaluate the performance of the algorithms. This information. Additional information such as breast density
impairs the fair judgment of the quality of the individual (preferably given by a standard like ACR) and BI-RADS
work and the comparison of the accuracy of different methods classification are also mandatory.
because performance is database-dependent. Good results can
have been obtained in databases with ‘‘easy’’ cases, whereas Organization of Database
bad accuracies may have been achieved by using ‘‘difficult’’
databases (16,18,20). Public available databases could A specific file format for digital mammograms does not exist.
provide a common ground for researchers to develop, test, Medical images are usually saved in the DICOM (Digital
and compare their methods. However, to be effective, Imaging and Communications in Medicine) format that
certain criteria should be met by the database. gathers not only the image but also some related metadata
This article aims to present a new mammographic research (21). A division of the images on training and test sets should
database originated at Centro Hospitalar de S. Jo~ao (CHSJ) at also be suggested. By doing so, different methods can be
Porto, Portugal, the INbreast database with the purpose of compared.
developing CAD methods and to overcome some limitations
of existent databases. Details of the design of this database will Distribution of Database
be presented in this article as follows: requirements of digital
The database should be available, preferentially over the
mammographic databases, existent available databases,
Internet. Continuous user support is also indispensable.
description of INbreast database, description of the findings
in the database, proposal for a methodology for performance
evaluation, discussion, and conclusion. AVAILABLE DATABASES
There are several image databases, some public and some
REQUIREMENTS FOR A DIGITAL restricted to individual groups, which are used by researchers
MAMMOGRAPHIC DATABASE in the breast cancer area. However, these often do not meet all
the requirements needed for a study (16–18,20,22–24).
According to previous studies (16,18), mammographic
databases should take into consideration the following
The Mammographic Image Analysis Society Digital
requirements. Mammogram Database

Case Selection The Mammographic Image Analysis Society Digital

Mammogram Database (MIAS) (25), despite being the oldest
The database should include various cases with images with available database, is still widely used in literature. This data-
normal breasts and all types of findings, and also all types of base has been reduced in resolution and is reachable at
breast density. Normal images with structures that may be http://peipa.essex.ac.uk/info/mias.html. Although images
misleading (eg, superimposed tissue that looks like a mass) are still available, it is no longer supported.
are important in order to make the classifiers more robust. MIAS consists of 161 cases, 322 digitized MLO images,
The cases should be collected by a specialist experienced in with all types of findings, including benign and malign lesions,
mammography. Each case should contain four standard views, and also normal images. It has a high percentage of spiculated

238
Academic Radiology, Vol 19, No 2, February 2012 INBREAST: TOWARD AN FFDM DATABASE

TABLE 2. Most Used Databases in Literature

MIAS (24) DDSM (29) BancoWeb (35)

Origin UK USA Brazil

Year 1994 1999 2010
Number of cases 161 2620 320
Views MLO MLO and CC MLO, CC, and other
Number of images 322 10,480 1400
Mode of image acquisition Screen film Screen film Screen film
Image type file PGM LJPEG TIFF
Resolution 8 bits/pixel 8 or 16 bits/pixel 12 bits/pixel
Lesion type All kinds (with special concentration All kinds All kinds
of spiculated masses)
Ground truth Center and radius of a circle Pixel level boundary ROI is available in
around the interest area of the findings a few images only
BI-RADS No Yes Yes
Breast density Yes (not ACR) Yes (in ACR standard) Yes (not ACR)
Clinical history No Age Yes
Search system No Yes, but not functional Yes
Access Yes Yes Yes
Support No No Yes

ACR, American College of Radiology; CC, craniocaudal; DDSM, Digital Database for Screening Mammography; LJPEG, lossless JPEG (Joint
Photographic Experts Group); MIAS, Mammographic Image Analysis Society Digital Mammogram Database; MLO, mediolateral oblique; PGM,
portable gray map; ROI, region of interest; TIFF, tagged image file format; UK, United Kingdom; USA, United States of America.

masses but, in Rangayyan’s article (24), the author noticed that information included is the age, but it has breast density
there was an unexpected elevated number of benign findings annotations (ACR) and BI-RADS annotations.
in relation to the malign ones. It contains breast density infor- Image annotations include pixel level boundary of the find-
mation, but not classified according to the ACR standards. ings. There are several articles whose authors got satisfactory
However, because of the increasing usage of the ACR classi- results using this type of annotation (32–35). However, as
fication, it was decided to classify the set of mammograms noted in other studies (8,22,23), they are not adequate for
according to that standard (26). the validation of segmentation algorithms because the
MIAS annotations consist in the center and radius of a circle precision is not good enough.
around the area of interest. These types of annotations are not
considered sufficient for some studies, as the one done by The BancoWeb LAPIMO Database
Oliver et al (20), where all circumscribed and spiculated
lesions had to be manually segmented. Another drawback is A more recent database is the BancoWeb LAPIMO Database
the resolution to which the images have been digitized, which (36). After registration, users can gain access and contribute to
makes MIAS unsuitable for experiments on detection of the database at http://lapimo.sel.eesc.usp.br/bancoweb/.
microcalcifications (MCCs) (27). However, in previous It has 320 cases, 1473 images with MLO, CC, and magni-
work (28), the authors achieved a 100% detection rate of fication views, with normal images, and images with benign
MCCs by applying two different detection methods. Llobet and malign findings. Background patient information along
(29) considered that, in the case of calcifications, the GT with BI-RADS annotations is available. Annotations exist in
region contains more healthy tissue than affected tissue. For only some of the images, in the form of a region of interest
this reason, calcifications were not included in his study. (ROI), but all have textual description of the findings. We
did not find any published work related to this database,
The Digital Database for Screening Mammography probably because it is a recent project. A summary of these
databases can be found in Table 2.
The most used database is the Digital Database for Screening Antoniou (19) refers to a web-accessible mammographic
Mammography (DDSM) (30), and it is accessible at http:// database, called MIRAcle DB, which is still in an experi-
marathon.csee.usf.edu/Mammography/Database.html, but mental stage and became available online in the summer of
is also no longer supported. It is the largest public database, 2011. This database, as well as others, is cited in the literature.
with 2620 cases including two images from each breast However, as can be seen in Table 3, most of them are not
(MLO and CC), for a total of 10,480 images, with all types available and consequently details could not be found.
of findings from normal images to images with benign and There are also two grids that are a combination of images
malign lesions. Some of the cases in this database were from multiple resources, which are available to the user as
collected from the Nijmegen Database (31). The only patient a single database, but can be stored in several servers.

239
240

MOREIRA ET AL
TABLE 3. Other Databases Referred in Literature

Nijmegen (37) Trueta (38) IRMA (39) MIRAcle (19) LLNL (39)  laga (38)
Ma NDMA (40)

Origin The Netherlands Spain Germany Greece USA Spain USA

Year 1998 2008 2008 2009 Unknown Unknown Unknown
Number of cases 21 89 Unknown 196 50 35 Unknown
Views MLO and CC MLO and CC MLO and CC Unknown MLO and CC MLO and CC Unknown
Number of images 40 320 10,509 204 198 Unknown 1,000,000
Mode of image acquisition Screen film FFDM Screen film Unknown Unknown Unknown Unknown
Image type file Unknown DICOM Several Unknown ICS Raw Unknown
Resolution 12 bits/pixel 12 bits/pixel Several Unknown 12 bits/pixel 12 bits/pixel Unknown
Lesion type MCCs All kind All kind Unknown Calcifications Masses Unknown
Ground truth Center and radius of Center and radius of Several Region of Outline of Pixel level Unknown
a circle around the a circle around the Interest calcifications annotations
interest area interest area
BI-RADS Unknown Yes Yes Yes Unknown Unknown Unknown

Academic Radiology, Vol 19, No 2, February 2012

Breast density Unknown ACR ACR No Unknown Unknown Yes
Clinical history Unknown Unknown Yes No Yes Unknown Unknown
Search system No Unknown Unknown YES Unknown Unknown Unknown
Access No No Yes Summer 2011 Paid Unknown No
Support No Yes Yes Yes Unknown Unknown Unknown

ACR, American College of Radiology; BI-RADS, Breast Imaging Reporting and Data System; CC, craniocaudal; DICOM, Digital Imaging and Communications in Medicine; FFDM, full-field
digital mammography; ICS, Image Cytometry Standard; IRMA, Image Retrieval in Medical Applications; LLNL, Lawrence Livermore National Laboratory; MCC, microcalcification; MLO, medio-
lateral oblique; NDMA, National Digital Medical Archive.
Academic Radiology, Vol 19, No 2, February 2012 INBREAST: TOWARD AN FFDM DATABASE

collected, from which 90 have two images (MLO and CC)

of each breast and the remaining 25 cases are from women
who had a mastectomy and two views of only one breast
were included. This sums to a total of 410 images. Eight of
the 91 cases with 2 images per breast also have images acquired
in different timings (follow-up).
The database includes examples of normal mammograms,
mammograms with masses, mammograms with calcifications,
architectural distortions, asymmetries, and images with
multiple findings (Fig 2). According to BI-RADS, a mass is
defined as a three-dimensional structure demonstrating
convex outward borders, usually evident on two orthogonal
views. Benign calcifications are usually larger than calcifica-
Figure 2. Chart describing the findings in the INbreast database. tions associated with malignancy, are usually coarser, and are
often round with smooth margins and are much more easily
Magic-5 (41) (previously known as GPCalma) is an Italian seen. Calcifications associated with malignancy are usually
database built in 2002 and containing 967 cases with images very small. An architectural distortion is defined as a focal
in MLO, CC, and lateral views, making a total of 3369 images. interruption of the normal mammographic pattern of lines
The screen films were digitized with a resolution of 12 bits and (converging at the nipple), usually presenting as a star-
saved in the DICOM format. Both masses and MCCs are shaped distortion, with no definite mass visible. An asymme-
present and the GT consists in the centre and radius of a circle try lacks convex outward borders of a mass and it and can be
around the interest area. Patient age is available but it has no represented in three ways: size asymmetry (difference in
BI-RADS categorization and the density classification is not volume between the right and left breast), focal asymmetry
on the ACR standard. Magic-5 limitations are related to the (unilateral, localized area of parenchyma), and global asymme-
different environments were images were acquired, making try (difference in the amount of parenchyma between the
them very heterogeneous. MammoGrid (42) is a collaboration right and left breast) (43). Concerning this distinction
between the United Kingdom, Italy, and Switzerland, with between asymmetries, this work does not take that into
images being standardized using the Standard Mammogram consideration.
Form representation and saved in the DICOM format. This The graphic in Figure 2 shows that there is a big promi-
grid has both screen films and FFDM images and annotation nence of calcifications on our database. This reflects the real
workstations are available in the participating hospitals. The population, where calcifications are the most common
main limitation of MammoGrid is that it is only available finding in mammography (44).
for associated institutions. Images contain findings of six types: asymmetries,
calcifications, distortion, masses, multiple findings (Fig 3),
and normal (Fig 1).
INBREAST DATABASE DESCRIPTION
The main characteristic of this work is the carefully associ-
The database was acquired at the Breast Centre in CHSJ, ated GTannotation. Most of the databases, such as MIAS, only
Porto, under permission of both the Hospital’s Ethics provide a circle around the area of interest. DDSM does have
Committee and the National Committee of Data Protection. pixel-level contours but, as noticed in a previous study (22),
The images were acquired between April 2008 and July 2010; they are not exact, which can impact accuracy measures by
the acquisition equipment was the MammoNovation Siemens incorrectly assigning some target pixels to background and
FFDM, with a solid-state detector of amorphous selenium, vice versa.
pixel size of 70 mm (microns), and 14-bit contrast resolution. The annotations were made by a specialist in the field, and
The image matrix was 3328
4084 or 2560
3328 pixels, validated by a second specialist, between April 2010 and
depending on the compression plate used in the acquisition December 2010. When there was a disagreement between
(according to the breast size of the patient). Images were saved the experts, the case was discussed until a consensus was
in the DICOM format. All confidential medical information obtained. Annotations were made on OsiriX, an open-
was removed from the DICOM file, according to Supplement source picture archiving and communication system
55 of the DICOM standard; the correspondence between (PACS) workstation, running on a Macintosh platform.
images of the same patient is kept with a randomly generated Each finding has a label that identifies the type of lesion.
patient identification. There are seven types of annotations: asymmetry (Fig 4),
INbreast has FFDM images from screening, diagnostic, and calcification (Fig 5a), cluster (of MCCs), mass (Fig 5b),
follow-up cases. Screening is made according to national and distortion (Fig 6a), spiculated region (Fig 6b) and pectoral
regional standards (5). Diagnostic is made when screening muscle (only in the MLO view; Fig 7). For the types, asym-
shows signs of anomaly. In follow-up images, cancer was metry (Fig 4), calcification (Fig 5a), mass (Fig 5b), distortion
previously detected and treated. A total of 115 cases were (Fig 6a) and pectoral muscle (Fig 7), a detailed contour of

241
MOREIRA ET AL Academic Radiology, Vol 19, No 2, February 2012

Figure 3. Database examples: multiple

findings. (a) Craniocaudal view of the right
breast; (b) mediolateral oblique view of
the right breast.

Figure 4. Annotation examples: asymme-

try. (a) Mediolateral oblique (MLO) view of
the right breast; (b) MLO view of the left
breast.

the finding was made. An ellipse enclosing the entire cluster Center</key> followed by a the coordinates of the point
was adopted to annotate the clusters of MCCs (Fig 5a). in the centre of the ROI, the tag <key>Name</key> fol-
When the mass is spiculated, besides a contour of the denser lowed by the type of finding (mass, calcification, distortion,
region, we added an ellipse enclosing all the spicules spiculated region) and some other general information’s
(Fig 6b). about the ROI;
The annotations were saved in XML format with the  After the general information, for each ROI, a list of
following structure. contour points is presented between the tags <array> and
</array>.
 A standard header with the XML version and type of
encoding information; Information regarding patient’s age at the time of image
 The tag <key>NumberOfROIs</key> followed by an acquisition, family history, ACR breast density annotation
integer that indicates the number of annotations present and BI-RADS classification is also provided (see Fig 8 for
in the image; the distribution of BI-RADS classification on the database).
 For each ROI, there is a tag <key>Area</key> followed A biopsy result for BI RADS 3, 4, 5, and 6 cases is also dis-
by the value of the area of the current ROI, the tag <key> played whenever performed. The remaining cases were

242
Academic Radiology, Vol 19, No 2, February 2012 INBREAST: TOWARD AN FFDM DATABASE

Figure 5. Annotation examples: (a) cluster;

(b) masses.

Figure 6. Annotation examples: (a) distor-

tion; (b) spiculated region.

considered benign and therefore a biopsy was therefore not With the precise annotations in INbreast, future studies can
performed. Consequently, a biopsy was performed on be developed that cannot be performed with the currently
56 cases, of which 11 were found to be benign and the available databases. Shape information is highly indicative of
remaining 45 were malignant. The overall distribution of the malignancy of a mass (45) and therefore automatic shape
benign/malignant cases is shown in Figure 8. assessment in the mammogram in often pursued. However,
The database is available at http://medicalresearch. the coarse-grained annotation of current databases does not
inescporto.pt/breastresearch/GetINbreastDatabase.html. A allow a proper validation of the discoveries. Also, MCC
division of the database into train and test sets is also suggested. grouping and distribution is the mammogram is important

243
MOREIRA ET AL Academic Radiology, Vol 19, No 2, February 2012

Figure 7. Annotation example: pectoral muscle.

Figure 8. Charts of (a) the Breast Imaging Reporting and Data

System images distribution (b) benign/malignant cases distribution.
to the correct diagnosis. Again, the usual annotation of the
MCCs with a single region enclosing all MCCs is insufficient
for the development of automatic methods. Finally, the distribution of the patient age, also included in
the database, is portrayed in Figure 13.
FINDINGS CHARACTERISTICS
PERFORMANCE EVALUATION
One of the most important breast characteristic is density.
Dense breasts are harder to analyze through mammography In CAD research, the quality of the detection algorithm is
than nondense ones. For each image in our database, its usually reported with the miss detection rate, false-positive
density in ACR standard scale is available. A distribution of rate, or similar metrics. With masses the computation of
density for each BI-RADS class is presented in Figure 9. such quantities is usually a straightforward process; because
There are a total of 116 masses among 107 images the number of objects is usually small, the correspondence
(z1.1 masses per image). The number of masses (normalized between detected and manually annotated masses is usually
by the total number of each class images) for each one of the clear. After adopting a distance notion between a detected
BI-RADS classes is shown in Figure 10; the average mass size and manually annotated mass (eg, a measure of overlap
is 479 mm2 (with a standard deviation of 619 mm2); the small- between the masses), a mass in considered correctly detected
est mass has 15 mm2 and the biggest has an area of 3689 mm2. is its distance to an automatically detected mass is below
Localization distribution of Masses is depicted in Figure 11. a certain threshold. From there, the miss detection rate is
Concerning calcifications, they are present in 301 of the just the number of undetected reference mass and the false-
410 images. The tag ‘‘cluster’’ was only used in 27 sets of calci- positive rate is just the number of automatically detected
fications, in 21 images (z1.3 clusters per image). Of these masses minus correctly classified masses.
21 images, only 2 had no single calcifications annotation. A This procedure is far from being easily extended to MCCs,
total of 6880 calcifications were thus individually identified with tens of objects per image. By computing the perfor-
in 299 images (z23.0 calcifications per image). BI-RADS mance as the result of accumulating local errors, we will likely
distribution is depicted in Figure 12. incur in many-to-one or one-to-many correspondences

244
Academic Radiology, Vol 19, No 2, February 2012 INBREAST: TOWARD AN FFDM DATABASE

Figure 9. Distribution of density across the Breast Imaging Reporting and Data System scale.

Figure 10. Normalized distribution of masses across the Breast Figure 12. Normalized distribution of calcifications across the
Imaging Reporting and Data System scale. Breast Imaging Reporting and Data System scale.

Figure 13. Distribution of age across the Breast Imaging Reporting

and Data System scale.

between reference and automatically detected MCCs. Gener-

ally, we argue, as others before (46), that the most interesting
Figure 11. Percentage of masses on each quadrant. UIQ, upper measures arise when one defines the (dis)similarity as the result
internal quadrant; LIQ, lower internal quadrant; LOQ, lower outer of optimizing a global function defined over all reference and
quadrant; UOQ, upper outer quadrant. detected objects simultaneously.

245
MOREIRA ET AL Academic Radiology, Vol 19, No 2, February 2012

Figure 14. Toy example with three ground

truth microcalcifications (MCCs) and two
automatically MCCs. Using directly the
Euclidean distance between the centroids,
the MCCs results in the assignment
problem results in the matrix at the right.
The output would match ground truth (GT)
1 to automatic detection (AD)2 and GT2 to
AD1. Saturating the distances to T2 = 5
would correctly match GT1 to AD1.

To evaluate the miss detection rate and false-positive rate, We do acknowledge that not all images in the database
we propose to start by computing the distance between each respect all quality assessment criteria, because the examination
reference MCC and each actually detected MCC; then we technique and patient related factors have some limitations
solve the matching problem on the resulting bipartite graph (eg, previous surgery). Therefore, the database reflects
by minimizing the assignment cost (= distance). Only pairs a wide variability of cases and conveys the reality of the routine
with average error-distance below a certain threshold (T1) work of a radiographer.
are assumed correctly matched (the other pairs are assumed Despite the fact that our database has a limited number of
to originate from a false-positive staff line being matched to images, we strongly believe that it is more important to have
an undetected true staff line and are therefore unmatched). imaging diversity, than a large number of similar images.
Now the two metrics result as the number of unmatched Zheng et al (47) claim that, in the development of CAD
MCC (false positive) and unmatched reference MCC (missed systems, including difficult cases leads to better results than
to detect). It should be noted that these metrics only measure simply increasing the size of the database with easy masses.
whether MCC are found, not how good the match is. Nevertheless, increasing the size of the database will be a future
The detection threshold T1 should reflect our imprecision phase of this research.
acceptance in the detection process. A final remark is related Annotation is a subjective, tedious, and extremely time-
to the notion of distance between two MCCs. Because consuming task. Specialists are needed to perform the annota-
MCCs are very small structures, it is a valid assumption to treat tion, which can turn into an extremely difficult and costly task.
them as singular points, their centroids. The obvious choice of That is probably the main reason why the currently available
embracing the Euclidean distance between the centroids for databases do not have accurate contours. In the present
the previously described global optimization can lead to work, there was a big concern in making precise annotations.
unnatural results (see Fig 14). However, only two specialists were involved in the process. For
The insight is that because high errors penalize a lot the that reason, the project will therefore integrate in the database
global optimization, the final correspondence result tries to additional specialists contributions to continuously improve
avoid such solutions. A workaround is to saturate the the database and annotations quality.
Euclidean distance:
CONCLUSIONS AND FUTURE WORK
d ðreference MCC; detected MCCÞ
¼ minðT2; d ½reference MCC; detected MCCÞ We consider that this project has the potential to be a unique
work in the field of mammographic databases. Notwithstanding
where T2 is a saturation value (an alternative approach would the importance of the digitized databases, technological
be to use a sigmoid function to compress the Euclidean advances in image acquisition devices for radiology, together
distance). The motivation for this saturation process is that with the ubiquity of the computer, led to the development of
erring by T2 (eg, 100) is the same as erring by any value above the FFDM, where the digitalization-related loss of information
T2 (eg, 700). is absent. Thus, the development of new databases that cover
such technological advancements is a crucial step to develop
future CADs.
DISCUSSION
With this database, we aim at increasing available
Having in attention the actual state-of-the-art on breast cancer resources in the breast imaging diagnostic field. This
research, FFDM databases are the natural step in the evolution updated set of images can be used not only for research
of mammographic databases. As noted by Oliver and purposes, but also in medical practice, for instance, in
colleagues (20), there is no public available database made a teaching environment. Within our team, we are interested
with digital mammograms. In this work, we address this gap in the development of a CAD system. Some MCCs and
by proposing a FFDM database with a wide variety of findings. mass detection methods are currently being implemented

246
Academic Radiology, Vol 19, No 2, February 2012 INBREAST: TOWARD AN FFDM DATABASE

and tested, and classification methodologies are also under 18. Nishikawa RM. Development of a common database for digital mammog-
raphy research. Chicago, IL: University of Chicago, 1996.
development. 19. Antoniou AC. A web-accessible mammographic image database dedi-
Careful annotation is considered as an advantage over the cated to combined training and evaluation of radiologists and machines.
currently available databases. This can motivate computer 9th International Conference on Information Technology and Applications
in Biomedicine, Cyprus, 2009.
vision researchers to develop methodologies that take advan- 20. Oliver A, Freixenet J, Martı J, et al. A review of automatic mass detection
tage of a better precision of the shape of the lesions to improve and segmentation in mammographic images. Medical Image Anal 2010;
detection and/or malignancy classification algorithms. The 14:87–110.
21. Bernarding J, Thiel A, Decker I, et al. Implementation of a dynamic
high image diversity of this database will provide a challenge platform-independent DICOM-server. Comput Methods Programs Bi-
believed to be difficult to overcome but extremely useful to omed 2001; 65:71–78.
design more robust CADs. For these reasons, we believe 22. Singh S, Bovis K. An evaluation of contrast enhancement techniques for
mammographic breast masses. IEEE Trans Inform Technol Biomed
that this database can be a reference for future works in the 2005; 9:109–119.
breast cancer imaging area. 23. Song E, Xu S, Xu X, et al. Hybrid Segmentation of Mass in Mammograms
The development of this database in an ongoing work; we Using Template Matching and Dynamic Programming. Academic Radi-
ology 2010; vol. 17:1414–1424.
plan to extend the number and variability of cases and at 24. Rangayyan RM, Mudigonda N, Desautels J. Boundary modelling and
the same time, improving the quality of annotations, or shape analysis methods for classification of mammographic masses.
even adding more case-related information, depending on Med Biol Eng Comp 2000; 38:487–496.
25. Suckling J. The Mammographic Image Analysis Society Digital Mammo-
the feedback provided by potential users. We are also starting gram Database. In Exerpta Medica. International Congress Series 1069,
to research the development of CAD systems benefiting from York, England. 1994; 1069:375–378.
the unique characteristics of INbreast database. 26. Oliver A, Llado X, Pe
rez E, et al. A statistical approach for breast density
segmentation. J Digital Imaging 2010; 23:527–537.
27. Domınguez AR, Nandi AK. Detection of masses in mammograms using
REFERENCES enhanced multilevel-thresholding segmentation and region selection
based on rank. In: IASTED International Conference on Biomedical Engi-
1. World cancer report. Lyon, France: International Agency for Research and neering. Austria: Innsbruck, 2007; 1:370–375.
Cancer, 2008. 28. Reljin B, Milosevic Z, Stojic T, et al. Computer aided system for segmen-
2. Health statistics: atlas on mortality in the European Union. Luxembourg: tation and visualization of microcalcifications in digital mammograms.
Eurostat, 2009. Folia Histochem Cytobiol 2009; 47:525–532.
3. Fact sheet No. 297: Cancer. France: World Health Organization, 2009. 29. Llobet R, Paredes R, Perez Cortes JC. Comparison of feature extraction
4. Misra S, Solomon NL, Moffat FL, et al. Screening criteria for breast cancer. methods for breast cancer detection. In Iberian Conference on Pattern
Adv Surg 2010; 44:87–100. Recognition and Image Analysis (IbPRIA). 2005; II:495.
5. Lee CH, Dershaw DD, Kopans D, et al. Breast cancer screening with 30. Heath M, Bowyer K, Kopans D, et al. Current status of the Digital Database
imaging: recommendations from the Society of Breast Imaging and the for Screening Mammography. In Digital Mammography 1998; 457–460.
ACR on the use of mammography, breast MRI, breast ultrasound, and 31. Netsch T, Peitgen HO. Scale-space signatures for the detection of
other technologies for the detection of clinically occult breast cancer. J clustered microcalcifications in digital mammograms. IEEE Trans Med
Am Coll Radiol 2010; 7:18–27. Imaging (MedImg) 1999; 18:774–786.
6. European Cancer Observatory. Breast cancer screening programmes in 32. Joseph JYL, Gavrielides M, Markey M, et al. Computer-aided classifica-
the EU Member States in 2006. Available at: http://eu-cancer.iarc.fr/ tion of breast microcalcification clusters: merging of features from image
cancer-13-display-text-562-564.html,en. processing and radiologists. Medical Imaging Conference, San Diego,
7. Karellas A, Vedantham S, Suryanarayan S. Digital mammography: an CA. Proc SPIE 2003; 5032:882–889. Available at: http://adsabs.harvard.
emerging technology. Business Briefing: Future Directions in Imaging edu//abs/2003SPIE.5032..882L.
2006; 2:5. Available at: http://www.touchbriefings.com/pdf/1677/Future_ 33. Wang DF, Shi L, Heng PA. Automatic detection of breast cancers in
imaging_06_Con.pdf. mammograms using structured support vector machines. Neurocomput-
8. Sampat MP, Bovik AC, Whitman GJ, et al. A model-based framework for ing 2009; 72:3296–3302.
the detection of speculated masses on mammography. Med Phys 2008; 34. Sakellaropoulos P, Costaridou LI, Panayiotakis G. A wavelet-based
35:2110–2123. spatially adaptive method for mammographic contrast enhancement.
9. Orsi CJ. The American College of Radiology mammography lexicon: an Phys Med Biol 2003; 48:787–803.
initial attempt to standardize terminology. AJR Am J Roentgenol 1996; 35. Cheng HD, Cui M. Mass lesion detection with a FUZZY neural network.
166:779–780. Patt Recognit 2004; 37:1189–1200.
10. Orel SG, Kay N, Reynolds C, et al. BI-RADS categorization as a predictor of 36. Matheus BR, Schiabel H. Online mammographic images database for
malignancy. Radiology 1999; 211:845–850. development and comparison of CAD schemes. J Digital Imaging 2010;
11. Mammography Accreditation Program: Submitting Clinical Images. 24:500–506.
Preston White Drive, Reston: American College of Radiology, 2009; 37. Lo CS, Chung PC, Chang CI, et al. A computerized system for detection
20191–4397. and segmentation of clustered microcalcifications. Joint Conference of
12. Bird RE, Wallace TW, Yankaskas BE. Analysis of cancers missed at International Computer Symposium 1996; 247–253.
screening mammography. Radiology 1992; 184:612–617. 38. Oliver A. Automatic mass segmentation in mammographic images. PhD
13. Kolb TM, Lichy J, Newhouse JH. Performance of screening mammog- thesis, Universitat de Girona (Espan ~a) 2008.
raphy among women with and without a first-degree relative with breast 39. Oliveira JEE, Gueld MO, de Arau jo A, et al. Towards a standard reference
cancer. Ann Intern Med 2000; 133:855–863. database for computer-aided mammography. San Diego, CA: Med
14. Zheng B, Hardesty LA, Polle WR, et al. Mammography with computer- imaging Computer-aided diagnosis, 2008. 6915. Available at: http://spie.
aided detection: reproducibility assessment - initial experience. Radiology org/x648.html?product_id=770325.
2003; 228:58–62. 40. Gardner WD. Breast cancer database provides faster access to patient
15. Castellino RA. Computer aided detection (CAD): an overview. Cancer records. InformationWeek, 2005. Available online at: http://www.
Imaging 2005; 5:17–19. informationweek.com/story/showArticle.jhtml;jsessionid=NBXENMVY3E
16. Nishikawa RM. Mammographic databases. Breast Dis 1998; 10:137–150. P5AQSNDBGCKHSCJUMEKJVN?articleID=174400322. Accessed Jan.
17. Orellana CJ. Study of a mammographic CAD performance dependence on 18, 2011.
the considered mammogram set. Proceedings of the International Confer- 41. Tangaro S, Bellotti R, De Carlo F, et al. MAGIC-5: an Italian mammo-
ence of IEEE Engineering in Medicine and Biology Society, Vancouver, graphic database of digitised images for research. Radiol Med 2008;
British Columbia, Canada 2008; 4776–4779. 113:477–485.

247
MOREIRA ET AL Academic Radiology, Vol 19, No 2, February 2012

42. Warren R, Solomonides T, del Frate C, et al. MammoGrid: a prototype 45. Sun L, Li L, Xu W, et al. A novel classification scheme for breast
distributed mammographic database for Europe. Clin Radiol 2007; 62: masses based on multi-view information fusion. In Bioinformatics and
1044–1051. Biomedical Engineering (iCBBE), 2010 4th International Conference,
43. D’Orsi CJ, Mendelson EB, Ikeda DM, et al. Breast Imaging Reporting and 2010; 1–4.
Data System: ACR BI-RADS - Breast Imaging Atlas. Reston, VA: American 46. Liu G, Haralick RM. Optimal matching problem in detection and recogni-
College of Radiology; 2003. Available at: http://www.acr.org/ tion performance evaluation. Patt Recog 2002; 35:2125–2139.
secondarymainmenucategories/quality_safety/biradsatlas/biradsfaqs.aspx. 47. Zheng B, Wang X, Lederman D, et al. Computer-aided detection: the effect
44. Danikas D, STheodorou SJV, Kokkalis G, et al. Mammographic findings of training databases on detection of subtle breast masses. Acad Radiol
following reduction mammoplasty. Aesthetic Plast Surg 2001; 25:283–285. 2010; 17:1401–1408.

248