Tải bản đầy đủ - 0 (trang)
BREAST IMAGING, COMPUTER-AIDED DETECTION, AND COMPUTER ASSISTED CLASSIFICATION

BREAST IMAGING, COMPUTER-AIDED DETECTION, AND COMPUTER ASSISTED CLASSIFICATION

Tải bản đầy đủ - 0trang

434



PACS: A Guide to the Digital Revolution



computer-associated artificial intelligence to diagnosis. However, it was also

the last major modality to have a commercially marketed digital system. Furthermore, in many departments mammography is the last division converted

to digital format.

The delay in the development and deployment of full field digital mammography (FFDM) primarily relates to the exacting spatial resolution

required in order to detect and correctly analyze microcalcifications. Many

malignant microcalcifications may be as small as 0.3 mm in size, 1 order of

magnitude smaller than most radiographically identified abnormalities. Controversy still exists regarding the optimum, cost-effective acquired pixel size

for the appropriate visualization and analysis of microcalcifications. For

example, a 50 or 100 mm pixel pitch might enable the identification of small

clusters of microcalcifications. However, a 25 mm pixel pitch may enable a

more rigorous analysis of the morphology of the individual microcalcifications, thus improving specificity by allowing better discrimination between

benign and malignant microcalcifications. However, with each halving of

acquired pixel pitch, there is a quadrupling of the number of pixels, which

substantially increases the cost of storage and time required for processing

and network transmission of the data.

A second major challenge faced by digital mammography is the need

for superb contrast over a wide dynamic range. The breast contains benign,

dense fibroglandular tissue and lucent fatty tissue as well as masses and/or

microcalcifications that may be of benign or malignant origin. Uniformly

dense mammograms and even mammograms with a combination of different densities can make discriminating subtle areas of malignancy most challenging. Improving visualization requires improved contrast resolution,

which entails optimizing the bit depth of each pixel since contrast resolution

is a function of the bit depth. Improved contrast resolution must also be optimized from a cost-benefit perspective similar to that for spatial resolution.

The requirements for a wide dynamic range, with optimal spatial and contrast resolution, have been met through extensive research and confirmed

with the approval process’s mandatory equivalency testing, allowing for the

ultimate introduction of digital mammography into the marketplace.

Because of these technological advances, new applications are being developed for FFDM. For example, tomosynthesis shows promise as a technique

to improve visualization of breast cancer.

In addition to the acquisition of the digital mammogram, downstream

requirements for processing (as well as computer-aided detection [CAD] and

potentially computer-assisted classification [CAC]) had to be developed. The

rigorous requirements related to the display for softcopy interpretation,

which is the preferred manner for digital interpretation, and production of



BREAST IMAGING, CAD, AND CAC



435



high-resolution hard copies (as needed) required extensive evaluation and

approval. The data storage requirements also had to be met in a costeffective manner. FFDM storage requirements, by virtue of the optimum

contrast and spatial resolution needed, are particularly large and hence costly.

Facilities with robust picture archiving and communication systems (PACS)

usually have a cost-effective archive. However, smaller radiology practices or

freestanding mammography boutiques may need to employ a specialized

mammography PACS or utilize offsite digital archiving, which adds significantly to the cost of FFDM.

Due to these requirements, the overall cost for complete FFDM

systems is substantially higher than for conventional mammography. This

cost has been partially mitigated by increased patient throughput available

with FFDM, which is in the range of 150% to 200% of standard film-screen

examinations. In addition, increased reimbursement for each case is allowed

by most health insurers, which is frequently 150% of the standard rate.

Now that FFDM has become a reality, the major benefits of the digital

image—CAD, CAC, and PACS—can be exploited. The digital image may

allow improved accuracy with the utilization of the CAD and CAC software.

Integration into a PACS allows optimized distribution and storage. However,

as these benefits are fundamentally related to the digital image, similar

benefits may potentially be obtained from computed radiography (CR)

technology once its ability to provide diagnostic accuracy is proven to

approximate conventional film-screen images and FFDM. Computed radiography mammography, when introduced, potentially has the benefit of

reduced costs, which in the competitive mammography marketplace may

exert downward pressure on the cost of FFDM.



COMPUTER-AIDED DETECTION AND

COMPUTER-ASSISTED CLASSIFICATION

Early and accurate detection of breast cancer is important for optimizing

therapy and affording the patient a better prognosis. Mammography, in conjunction with physical examination, is the most effective tool available today

for mass screening for the early detection of breast cancer. It can potentially

reduce mortality by as much as 40% to 63%. Various studies have shown

that a significant number of biopsy-proven malignancies are not detected by

mammography. Mammography has particular difficulty in detecting breast

cancers in women with dense breast tissue that may obscure the lesion or

make it more difficult to identify even if visible. Subtle areas of architectural

distortion can also be challenging. Other false negatives may relate to visible



436



PACS: A Guide to the Digital Revolution



cancers being misinterpreted or overlooked because of inexperience, distraction, or fatigue caused by the large number of examinations presented

to the limited number of mammographers or by suboptimal film quality.

Therefore, methods that can overcome these difficulties in interpreting

mammograms are of major importance. It has been shown that breast cancer

detection rates can be substantially improved by using second readers.

However, in many instances it is not feasible to perform a second read by

human mammographers. Potentially, even after a second reader’s interpretation, significant numbers of missed cancers might still exist and may be

detected by the addition of computer software.

For this reason CAD appears to be one of the most important or, as

some suggest, the most important collateral benefit from the availability of

digital mammography, since it allows cost-effective and efficient acquisition

of the digital image to computer software that can then serve as the second

reviewer (or even the third reviewer after the double reading by the human

mammographers) to optimize the detection of abnormalities. CAD can overcome much of the difficulty and expense inherent in a reading by a second

mammographer and may conceptually be even more accurate than some

mammographers in detecting abnormalities.

These detection systems are popularly known by the acronym CAD,

which has been used to represent varied abbreviations including computeraided detection, computer-aided diagnosis, computer-assisted detection, and

computer-assisted diagnosis. In this chapter, CAD will be used to represent

computer software that detects abnormalities. In addition, the data can also

be analyzed by specialized software that can help classify the likelihood of

malignancy of any identified abnormalities. These systems will be called

CAC in this chapter, representing computer-assisted classification.

This idea of using computers in facilitating the mammography examination is not a new one. In August of 1967, Winsberg, Elkin, et al., from

the Albert Einstein College of Medicine of Yeshiva University, authored an

article in Radiology titled “Detection of Radiographic Abnormalities in Mammograms by Means of Optical Scanning and Computer Analysis,” in which

they noted, “Because of the problems inherent in the routine viewing of large

numbers of examinations of presumably asymptomatic patients we have

proposed the automation of reading of the radiographs by means of optical

scanning and computer interpretation.” It has taken over 35 years to even

approximate this idealized vision of mammography.

After the digital image is acquired, software algorithms are employed

to meticulously review the data and search for areas that appear to be different from the surrounding tissue. These algorithms can identify microcalcifications and masses. However, the relatively rare but visually challenging



BREAST IMAGING, CAD, AND CAC



437



areas of architectural distortion are not as well identified by CAD, requiring more research. Integration of data from both breasts, which would allow

identification of potential asymmetry, is also presently under development.

Integration of prior mammograms may also be useful in identifying developing densities or other interval changes.

The sensitivity of detection of abnormalities, the prime role of CAD,

is related to a great extent to the specificity of the system. For example, using

just the number of microcalcifications as a discretionary feature, a system set

for a minimum threshold of 5 microcalcifications per area may discount early

malignancies that manifest themselves with only 4 microcalcifications.

However a threshold level set at 3 microcalcifications would identify those

malignant clusters as well as any other clusters of 4 microcalcifications. Thus,

to maximize true positive identifications (reduce false negatives) it is necessary to allow for more false positives, although too many false positives may

be disturbing and potentially discredit the CAD system for some users.

Potentially, practices whose referral base has little tolerance for a low sensitivity and will readily accept lower specificity can have their needs met by

increasing the false positives allowed. A “dial up or down” feature to allow

for practices or mammographers to set their own sensitivity/specificity level

would be a desirable feature.

The utilization of a CAD system requires a digital image. This digital

image can be derived from 3 different mechanisms: (1) manual digitization

of a conventional film-screen mammogram, which is labor intensive, (2) a

FFDM machine, or (3) CR (potentially). Conceptually, the end result is a

digital image representing the underlying breast tissue. However, there are

inherent problems associated with digitization of conventional film-screen

mammograms. This is because of the lack of reproducibility of the digitized

image. The use of calibration strips can allow for an approximation, but each

time an analog image is digitized, the results are slightly different, and this

could affect the results of the CAD analysis. A slightly different input due

to fluctuations in the digitization process could yield a very different output,

which could create medicolegal difficulties. Thus, it is simpler to deal with

images that are inherently digital instead of trying to digitize analog ones.

Inherently digital images can also be sent more efficiently through a CAD

system because they require no digitization step (although no definitive

studies have yet been published on this). For this reason, the CAD-marked

images should not have to be archived, as long as the version of the CAD

system used is documented and can be reproduced as needed in the future

in conjunction with the original digital mammography data.

Because there are 2 different primary signs of breast cancer, microcalcifications and masses, it is necessary to have 2 different detection algorithms.



438



PACS: A Guide to the Digital Revolution



The microcalcification algorithm identifies and evaluates pixels that may represent calcium. The individual microcalcifications can be outlined, their size

quantified, and the features of each analyzed. The cluster as a whole is also

analyzed. The minimum number of adjacent calcifications within a prescribed area may be calculated and compared to the software’s minimum

threshold level deemed appropriate for analysis. Mathematical models are

then used to evaluate the combined data and determine which cluster of

microcalcifications should be presented to the mammographer for further

evaluation. With respect to the evaluation of masses, the software can search

for an area with different characteristics from the background. It can refine

its analysis with evaluation of the mass’s density and its border characteristics, such as spiculation.

CAD has been successful in detecting approximately 80% to 90% of

all malignancies. However, the systems at present appear to be more sensitive in detecting clusters of microcalcifications than masses, with microcalcifications detected in the upper 90% range and masses detected in the 80%

to 90% range. The less frequently occurring architectural distortion is relatively poorly identified. As can be expected, CAD will benefit some radiologists more than others. Experience appears to be inversely proportional to

the amount of improvement seen with CAD.

Computer-aided detection is primarily a sensitivity-enhancing tool,

with potentially deleterious effects not only to specificity, but also to recall

rates and positive predictable value. Recall rates appear to have a modest

deterioration that appears because of the sensitivity improvement. Similarly,

there have been only minimal effects on positive predictive value (PPV). This

is because the additional false positive biopsies recommended are offset by

the additional true positives. When the two are combined, the PPV

approaches the original PPV, but with more cancers detected. Therefore, it

is better if abnormalities detected by either the mammographer or CAD can

be directed to a CAC system. Such a classification system can then evaluate

and exclude many questionable areas, thereby making the correct diagnosis

as accurate, sensitive, and specific as possible while maintaining reasonable

recall rates. Similar to CAD, these classification systems also need to have

different algorithms to classify microcalcifications and masses, but they do

not use the same algorithms as those used for detection. Once detected, each

microcalcification is characterized according to several properties, including

brightness, area, length, and shape. After these quantitative parameters are

obtained, the microcalcification cluster can be classified into each of 3 major

subdivisions: first is a group that characterizes variables in morphology of

the individual calcifications; second is a group that reflects the heterogeneity of the morphologic features related to the individual calcifications; third



BREAST IMAGING, CAD, AND CAC



439



is a group that characterizes variables in the distribution of the calcifications

within a cluster. The spatial orientation appears to be most helpful in distinguishing benign from malignant clusters. Artificial intelligence then can

analyze all the information and compare it to known statistics representing

frequency of malignancy. The system then can give a final assessment as to

the likelihood of an abnormality being malignant (see Figure 21.1). This

analysis cannot be performed by the mammographer because many of the

underlying features analyzed are too small (malignant microcalcifications are

in the 0.3 mm range) and the calculations too complex, making the analysis

beyond human capability. Algorithms for evaluating masses include analysis

of features that characterize shape, definition of the margins, and various

aspects of spiculation (see Figure 21.2). The information derived can be used

to approximate the likelihood of malignancy (high, intermediate, or low) of

any abnormalities identified.

Prototype classification schemes that give gross estimates of malignancy are currently being evaluated and show promise in initially being used



FIGURE 21.1



Computer analysis of a cluster of microcalcifications. (Courtesy of

Professor Isaac Leichter, Siemens Computer Aided Diagnosis Ltd.)



440



PACS: A Guide to the Digital Revolution



FIGURE 21.2



Computer analysis of a breast mass. (Courtesy of Professor Isaac

Leichter, Siemens Computer Aided Diagnosis Ltd.)



by physicians to potentially (and at their discretion) modify Breast Imaging

Reporting and Data System (BIRADS) assessments. Systems of this nature

may be particularly helpful in evaluating abnormalities that have a preliminary diagnosis of probably benign. With more refinements, the CAC systems

may ultimately be able to give a probability range of an abnormality’s potential to be malignant, for example, 20% to 25% chance. The referring physician and patient can then incorporate these data to more accurately assess

risk-benefit relationships of the various options for management of the identified abnormality. In addition, combination systems with CAD and CAC,

in tandem, are also being evaluated and appear to be on the horizon to aid

the mammographer’s interpretation.

Ultimately, robust algorithms combining CAD and CAC systems may

be employed to prescreen the large number of screening mammograms from

the patients recommended to have this examination. Only selected cases that



BREAST IMAGING, CAD, AND CAC



441



fall outside predetermined threshold levels for probability of disease will then

need to be evaluated by the mammographer for final assessment. Similar

computer methods have been applied medically for cervical cancer screening and also in nonmedical commercial endeavors. Despite the significant

research that has already been undertaken, particularly over the past decade,

additional research is required to fulfill the promise envisioned by Winsberg,

Elkin, et al., in 1967 to make mammography as sensitive, specific, readily

available, and cost effective as possible.



PACS AND BREAST IMAGING

The integration of breast imaging into PACS is associated with several

unique challenges with respect to softcopy image interpretation, communication for consultation, and archival storage. When an organization commits

to FFDM, it must decide how the images will be interpreted. The preferred

method of interpretation is via softcopy display, although some practices may

still prefer printed film displayed on a view box. Softcopy reading requires

high-resolution monitors to enable the physician to visualize the minute

details for accurate diagnosis. Printed film interpretation of images derived

from FFDM requires a specialized high-resolution printing capability.

If the decision is to use the digital image for only softcopy interpretation, no hardcopy production should be necessary unless specifically required

for distribution outside the institution. Using the PAC system’s existing

archive appears intuitively to be the most appropriate method. However,

during a transition phase (which in mammography usually takes up to 3

years), previous images are needed for comparison in order to optimize diagnosis. The comparison with the previous examinations enables appropriate

diagnoses of any new masses or microcalcifications or allows for the identification of any developing densities. Therefore, in the interim, adjacent to

softcopy monitors used for FFDM acquired images, a view box for the conventional historical image review is necessary. An alternative is the digitization of the historic films with subsequent archiving into the PACS system

and visualization as necessary on a softcopy monitor for direct comparison

to the FFDM image. This process, in addition to being costly and tedious,

has the inherent limitation of potential loss of information related to the digitization process. Therefore, most institutions have chosen to have a dualmethod display: softcopy for the current FFDM images and films on a view

box for the historical image comparison.

With respect to distribution of images, 3 separate situations must be

considered. Many facilities run screening programs. Approximately 5% to



442



PACS: A Guide to the Digital Revolution



15% of patients who are seen at a screening facility will require immediate

additional imaging, usually performed within 2 weeks at a diagnostic facility. This will require the screening images to be available in order to appropriately work up the abnormality. Other patients who have had FFDM

performed will need to have images distributed to referring physicians, such

as breast surgeons or other consultants (and potentially to the operating

room during a surgical procedure). There are also images that the patient

may wish to distribute for a second opinion to physicians not affiliated with

the original practice. With respect to patients who have had screening

images, the site where the screening images were obtained may either be

FFDM or conventional film-screen, and the site where the patient will

undergo the diagnostic study may likewise have either of the 2 types of

imaging equipment. As with the routine comparison with historic films, 1

set of images may be on softcopy display and the second set of images may

be on a film view box.

Distribution to referring physicians and/or the operating room within

the institution for gross visualization (not primary diagnosis) may be performed over the PACS enterprise distribution system (typically a Web

browser–based interface) even though the display systems on most of the

clients is not suitable for primary diagnosis. For most exams, this would

appear to be adequate particularly for abnormalities such as large masses or

clusters. For off-site distribution, facilities must use high-resolution copiers,

which need to be specifically approved by the FDA for use with mammography images obtained with FFDM. These images can be shipped as any

other images and/or given to the patients for personal distribution. Direct

electronic image distribution to off-site locations would be desirable provided that necessary issues, such as network security, bandwidth, and Health

Insurance Portability and Accountability Act (HIPAA) regulations, are

worked out.

Mammography images, by virtue of their high spatial resolution and

need for contrast over a large dynamic range, require relatively large

amounts of storage. Currently, mammography images are stored with lossless compression. No studies have been performed yet with respect to storing

mammography in a lossy form. This burden on a PACS system is substantial and clearly has to be calculated into the size of storage requirements

when an organization with a PACS is contemplating FFDM integration or

when an FFDM system is bought for a freestanding organization without a

PACS already in place.

It is unclear whether it is really necessary to store the outputs of CAD

and CAC. So long as the version of the algorithms used at the time of diagnosis is clearly documented, the results should be completely reproducible.



BREAST IMAGING, CAD, AND CAC



443



This is analogous to the use of window and leveling in computed tomography (CT) exams. Even though the radiologist often views CT images under

3 different window/level settings, there is no need to store a separate copy

of the images for each window/level setting used. Rather, applying the standard settings to the original pixel data will consistently yield the same results.

So, too, knowing which versions of CAD and CAC were used should permit

presentation of the marking data at any future time exactly as they were presented to the mammographer at the time of initial interpretation.

In summary, mammography has had a long history with the digital revolution, but it is only now with the advent of FFDM that the benefits of

CAD, CAC, and PACS can be realized.



◗ REFERENCES

Australian Health Technology Assessment Commission. Review of automated and

semi-automated cervical screening devices—a summary. Canberra, Australia:

Commonwealth Department of Health and Family Services; 1998.

Baker JA, Rosen EL, Lo JY, et al. Computer-aided detection (CAD) in screening

mammography: sensitivity of commercial CAD systems for detecting architectural distortion. AJR Am J Roentgenol. October 2003;181:1083–1088.

Bamberger P, Novak B, Fields, S. Optimizing parameters for computer aided diagnosis of microcalcifications in mammography. Acad Radiol. 2000;7:406–412.

Beam CA, Sullivan DC, Layde PM. Effect of human variability on independent

double reading in screening mammography. Acad Radiol. 1996;3:891–897.

Bird RE, Wallace TW, Yankaskas BC. Analysis of cancers missed at screening mammography. Radiology. 1992;184:613–617.

Birdwell RL, Ikeda DM, O’Shaughnessy KF, Sickles EA. Mammographic characteristics of 115 missed cancers later detected with screening mammography

and the potential utility of computer-aided detection. Radiology. 2001;

219:192–202.

Brem RF, Baum J, Lechner M, et al. Improvement in sensitivity of screening mammography with computer-aided detection: a multiinstitutional trial. AJR Am J

Roentgenol. September 2003;181(3):687–693.

Buchbinder S, Novak B. Quantitative characterization of mass lesions on digitized

mammograms for computer assisted diagnosis. Invest Radiol. 2000;35:366–372.

Buchbinder S, Leichter I, Bamberger P, et al. Analysis of clustered microcalcification using a single numeric classifier extracted from mammographic digital

images. Acad Radiol. November 1998;5:779–784.

Buchbinder S, Leichter I, Lederman R, et al. Can the size of microcalcifications

predict malignancy of clusters in mammography? Acad Radiol. 2002;9:18–25.



444



PACS: A Guide to the Digital Revolution



Buchbinder S, Leichter I, Lederman R, et al. Computer-aided classification of

BI-RADS category 3 breast lesions. Radiology. 2004;230:820–823.

Burhenne LW, Wood S, D’Orsi C, et al. Potential contribution of computer-aided

detection to the sensitivity of screening mammography. Radiology. 2000;215:

554–562.

Cady B, Michaelson JS. The life sparing potential of mammographic screening.

Cancer. 2001;91:1699–1703.

Ciatto S, Del Turco MR, Morrone D, et al. Independent double reading of screening mammograms. J Med Screen. 1995;2:99–101.

Destounis SV, Dinitto P, Young-Logan W, et al. Can computer-aided detection with

double reading of screening mammograms help decrease the false-negative rate?

Initial experience. RSNA. 2004;578:584–585.

Duffy SW, Tabar L, Chen HH, et al. The impact of organized mammography service

screening on breast carcinoma mortality in seven Swedish counties. Cancer.

2002;95:458–469.

Fields S, Lederman R, Buchbinder S, et al. Improved mammographic accuracy with

CAD assisted ranking of lesions. In: Proceedings of the 6th International Workshop on Digital Mammography, June 2002, Bremen, Germany.

Fields S, Leichter I, Lederman R, et al. Improving the Performance of Mammographic Assessment by the use of an Advanced Two-tiered CAD/CAC System.

In: Proceedings of the 7th International Workshop on Digital Mammography,

June 2004, Charolotte, NC, USA.

Freer T, Ulissey M. Screening mammography with computer-aided detection:

prospective study of 12,860 patients in a community breast center. Radiology.

2001;220:781–786.

Helvie MA, Hadijiiski L, Makariou E, et al. Sensitivity of Noncommercial Computer-aided Detection System for Mammographic Breast Cancer Detection:

Pilot Clinical Trial RSNA. 2004;208:214.

Improving the performance of mammographic assessment by the use of an advanced

two-tiered CAD/CAC system. Digital Mammography. 2002.

Kopans D. Breast Imaging. 2nd ed. Philadelphia: Lippincott-Raven; 1998.

Leichter I, Fields S, Nirel R, et al. Improved mammographic interpretations of

masses using computer-aided diagnosis: Eur Radiol. 2000;10:377–383.

Leichter I, Lederman R, Buchbinder S. Mammographic lesions: what diagnostic role

does the individual mc’s shape play compared to the cluster geometry? AJR Am

J Roentgenol. 2004;182:705–712.

Leichter I, Lederman R, Novak B, et al. The use of interactive software program

for quantitative characterization of microcalcifications of digitized film-screen

mammograms. Invest Radiol. June 1999;34:394–400.

Lederman R, Leichter I, Buchbinder S, et al. Stratification of mammographic

computerized analysis by BIRADS categories, Eur Radiol. 2003;13:347–353.



Tài liệu bạn tìm kiếm đã sẵn sàng tải về

BREAST IMAGING, COMPUTER-AIDED DETECTION, AND COMPUTER ASSISTED CLASSIFICATION

Tải bản đầy đủ ngay(0 tr)

×