DOI: 10.1148/radiol.2432060253
(Radiology 2007;243:360-367.)
© RSNA, 2007
Breast Cancer Detection Rate: Designing Imaging Trials to Demonstrate Improvements1
Yulei Jiang, PhD,
Diana L. Miglioretti, PhD,
Charles E. Metz, PhD, and
Robert A. Schmidt, MD
1 From the Department of Radiology, University of Chicago, 5841 S Maryland Ave, MC2026, Chicago, IL 60637 (Y.J., C.E.M., R.A.S.); and Group Health Cooperative, Seattle, Wash (D.L.M.). Received February 9, 2006; revision requested April 7; revision received June 7; accepted July 7; final version accepted September 20. Supported in part by the National Cancer Institute through grants R01 CA92361 and U01CA86076.
Address correspondence to Y.J. (e-mail: y-jiang{at}uchicago.edu).
 |
ABSTRACT
|
|---|
Purpose: To estimate the extent of variability in screening mammography cancer detection rates and its effect on a hypothetical clinical trial of a new screening modality used to measure changes in cancer detection rate.
Materials and Methods: Each registry and the statistical coordinating center received institutional review board approval along with approval for consenting processes or a waiver of consent to enroll participants, link data, and perform analytic studies. This study was HIPAA compliant. The authors estimated the distribution of individual radiologists' breast cancer detection rates for 2 289 132 screening mammograms (9030 cancers) read by 510 radiologists in the United States who participated in the Breast Cancer Surveillance Consortium from 1996 through 2002. They then computed the distributions of breast cancer detection rates expected from a trial of screening mammography and multiple radiologists, as well as similar distributions for a hypothetical new modality that depicts one additional cancer per reader per 1000 screening examinations. Statistical power was calculated.
Results: The mean screening mammography cancer detection rate for individual radiologists was 3.91 cancers (standard deviation, 1.93; range, 0.25–13.75) per 1000 examinations. To achieve 80% power to detect a hypothetical increase of one additional cancer detected per reader per 1000 screening examinations, a trial in which a new modality was compared with standard mammography would require at least 25 radiologists each reading the images of at least 8000 screening examinations or 91 radiologists each reading the images of 1000–2000 examinations.
Conclusion: The low breast cancer prevalence in an average-risk screening population and the large interradiologist variability in the observed cancer detection rate suggest that for new technologies to demonstrate significant improvement in cancer detection rate in a clinical trial, very large samples of both radiologists and patients will be required.
© RSNA, 2007
|
INTRODUCTION
|
|---|
An important goal in the use of new breast imaging technologies or modalities is to detect more breast cancers in screening populations compared with the numbers of cancers detected with standard screen-film mammography. Clinical trials in which full-field digital mammography and computer-aided detection have been compared with standard screen-film mammography have not consistently shown an increase in cancer detection rate (1–8). Post hoc analysis results suggest that many trials lack statistical power owing to the small numbers of cancers included (4,9), which reflect the low prevalence of preclinical breast cancer in an average-risk screening population (10). Interradiologist variability in the interpretation of mammograms (11–20) might also have contributed to the inconsistent results in these trials, but the extent of that effect is unknown (8). The purpose of our study was to estimate the extent of variability in screening mammography cancer detection rates and its effect on a hypothetical clinical trial of a new screening modality used to measure changes in cancer detection rate.
|
MATERIALS AND METHODS
|
|---|
Data Sources
Cancer detection rates were estimated from 2 289 132 screening mammograms (9030 cancers) read by 510 radiologists who participated in the Breast Cancer Surveillance Consortium (21,22). The Consortium includes seven regional registries: the Carolina Mammography Registry, Chapel Hill, NC; the Colorado Mammography Project, Denver, Colo; the New Hampshire Mammography Network, Lebanon, NH; the New Mexico Mammography Project, Albuquerque, NM; the San Francisco Mammography Registry, San Francisco, Calif; the Vermont Breast Cancer Surveillance System, Burlington, Vt; and Group Health Cooperative, Seattle, Wash. Each registry collects information on mammography examinations performed at participating facilities in its catchment area and links these examinations to a state cancer registry or regional Surveillance, Epidemiology, and End Results program and to pathology databases to determine cancer status and tumor characteristics. Cancer ascertainment from these combined sources is estimated to be more than 94.3% complete (23). Each registry and the statistical coordinating center received institutional review board approval along with approval for consenting processes or a waiver of consent to enroll participants, link data, and perform analytic studies. All procedures and our study were compliant with the Health Insurance Portability and Accountability Act, and all registries and the statistical coordinating center received a federal Certificate of Confidentiality that protects the identities of research subjects.
We included only screening mammograms obtained between January 1, 1996, and December 31, 2002, in women without a personal history of breast cancer. A mammogram was considered a screening examination if the radiologist indicated that it was obtained for routine screening. Mammograms that included only unilateral views and those that followed a mammogram or other related radiologic examination within the preceding 9 months were excluded because they were likely diagnostic examinations. Mammograms were also excluded if computer-aided detection was used or if the radiologist read fewer than 500 mammograms during the study period.
A mammogram was considered positive if it was given a BI-RADS (Breast Imaging Reporting and Data System) assessment score of 0 (need additional imaging evaluation, 176 922 mammograms [7.7% of total cases]), 4 (suspicious abnormality, 9540 mammograms [0.4%]), or 5 (highly suggestive of malignancy, 1244 mammograms [0.1%]) (24). A mammogram was considered negative if it was given an assessment score of 1 (negative, 1 552 941 mammograms [67.8% of total cases]), 2 (benign finding, 488 543 mammograms [21.3%]), or 3 (probably benign finding) with a recommendation for short-interval or routine follow-up (42 346 mammograms [1.8%]). Mammograms given an assessment score of 3 with a recommendation for immediate follow-up (17 596 mammograms [0.8% of total cases]) were recorded as having a score of 0 (positive), because the 0 assessment more appropriately matches the recommendation. The cancer detection rate was calculated (by D.L.M.) separately for each radiologist as the number of breast cancers detected per 1000 screening mammograms (invasive carcinoma or ductal carcinoma in situ diagnosed within 1 year after an examination with a positive result).
Statistical Analyses
We estimated the population probability distribution of observed single-reader cancer detection rates by scaling the histogram of the calculated single-reader cancer detection rates. We assumed that this probability distribution accurately reflects single-reader cancer detection rates observed in trials. At least three sources of variability affect the observed single-reader cancer detection rate: (a) the single-reader screening mammogram volume, which is associated with intrareader variability in the observed single-reader cancer detection rate (smaller volume produces greater variability); (b) interreader variability—the between-reader variability in cancer detection rates that would persist even when measured from infinitely large mammogram volumes; and (c) possible variation in patient population demographics and breast cancer incidence. We did not separate these sources of variability because the observed single-reader cancer detection rates in future trials would be influenced by all these sources of variability combined. Probability distributions of single-reader cancer detection rate were estimated separately for 10 single-reader screening mammogram volumes.
To compare breast cancer detection rates of mammography and a hypothetical new modality observed in trials, one of us (Y.J.) calculated the probability distribution (in replicated trials) of observed multireader mean cancer detection rates from the probability distribution of observed single-reader cancer detection rates (Appendix). For such calculation, single-reader cancer detection rates will be uncorrelated across modalities if cancer detection rates for two modalities are measured from two different groups of readers and two different cohorts of patients. However, if trial investigators compare cancer detection rates from the same group of readers or from the same patient cohort, then the single-reader cancer detection rates will be correlated across modalities. This correlation will give rise to higher statistical power in trials. However, the data needed for estimating this correlation do not exist until after a new modality has already been used extensively in clinical practice.
Therefore, we substituted the correlation between mammography and itself for the correlation between mammography and a new modality, and we assumed that the former correlation would be at least as strong as the latter correlation. One of us (D.L.M.) measured the correlation between mammography and itself (which also depends on the single-reader mammogram volume) from two single-reader cancer detection rates (rate 1, mammograms obtained in 1996, 1998, 2000, or 2002; rate 2, mammograms obtained in 1997, 1999, or 2001) for 392 readers who each contributed at least 500 screening mammograms to the calculation of each rate. These two rates were chosen to minimize the effect of gradual longitudinal changes. Analysis of larger case volumes was unreliable owing to small numbers of qualified readers.
To calculate the probability of trial outcomes, we postulated that, compared with standard mammography, a new modality would depict one additional cancer for every reader per 1000 screening examinations. Given this postulate, one of us (Y.J.) calculated, for a multireader trial comparing a new modality against mammography, (a) the probability of observing no increase (or observing a decrease) in the cancer detection rate; (b) the probability of observing at least half of the postulated improvement (ie, one-half or more additional cancers detected per 1000 screening examinations); and (c) the power for detecting a statistically significant increase in the cancer detection rate.
Statistical power was defined as the probability of correctly rejecting the null hypothesis of no difference with a critical value of .05, given the alternative hypothesis of an increase of one additional cancer detected per reader per 1000 screening examinations. The calculation (Appendix) was performed numerically with a custom-written computer program that was validated by means of analytical calculation of hypothetical normally distributed observed single-reader cancer detection rates. (These calculations were exact in that they produced discrete but the exact same results that the analytic calculations produced, with the exception of negligible numerical errors.) Normality was not assumed in the calculation except for validation of the methods.
|
RESULTS
|
|---|
Cancer Detection Rates
The mean observed single-reader cancer detection rate among 510 readers who read at least 500 screening mammograms during the study was 3.91 cancers, the standard deviation was 1.93, and the range was 0.25–13.75 cancers per 1000 screening mammograms (Fig 1). When these radiologists were grouped by case volume, the mean observed single-reader cancer detection rate for most groups was higher than the mean rate of 3.91 cancers per 1000 examinations, especially among those who read between 2000 and 6000 mammograms during this study (more than four cancers per 1000 screening mammograms), but it was lower than the overall mean for radiologists who read either 500–1000 or more than 10 000 mammograms (Table 1). The standard deviation decreased with increasing case volume, with the smallest standard deviation among radiologists who read 8000 or more mammograms during the study period (Table 1). For the entire study population, the overall breast cancer detection rate for all methods combined and the detection rate for screening mammography were 5.09 and 3.94 cancers per 1000 examinations, respectively.
View larger version (10K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 1: Histogram of single-reader cancer detection rates for 510 radiologists who read at least 500 screening mammograms during the study. The mean cancer detection rate is 3.91 cancers (standard deviation, 1.93) per 1000 screening mammograms.
|
|
Statistical Projections
Our statistical projections regarding the postulated increase of one additional cancer detected per reader per 1000 screening examinations with a new hypothetical modality (which corresponds to an estimated 25% increase in cancers detected) are summarized in Figure 2 for trials that involve different readers and different patient cohorts across modalities (ie, no correlation in cancer detection rates across modalities) and in Figure 3 for trials that involve the same readers but different patient cohorts (ie, correlation in cancer detection rates across modalities). For the analysis with no correlation, there was a 29%–40% probability that a single reader would not show an increase (or would show a decrease) in cancer detection rate with the new modality; reducing this probability to less than 20% would require at least three to eight readers in one study, depending on single-reader case volume. The probability was only 53%–61% that a single reader would observe at least half of the postulated increase from the new modality, and increasing this probability to 80% would require nine to more than 30 readers in one study, depending on single-reader case volume. To achieve 80% power for detecting a statistically significant increase would require at least 25 readers, each reading the images of 8000 or more screening examinations, or 91 readers, each reading the images of 1000–2000 examinations. Widths of the 95% confidence intervals for power in a 30-reader study ranged from 15% to 35%, with wider intervals generally corresponding to case volumes that had fewer readers from whom the observed single-reader cancer detection rate was estimated. The 95% confidence intervals straddled 80% power only for single-reader case volumes of 3000 or more and were below 80% power for smaller case volumes. Figure 3 shows the 95% confidence intervals for single-reader case volumes of at least 500.
View larger version (43K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2: Probabilities as a function of the number of readers in a trial for observing no additional cancers detected (left), one-half or more additional cancers detected per 1000 screening mammograms (middle), and a statistically significant increase in cancer detection rate (ie, statistical power) (right), given a postulated increase of one additional cancer detected per 1000 screening mammograms. Correlation of individual readers' cancer detection rates between the two arms of a trial is not included. Variability in the observed single-reader cancer detection rates caused the discrepancies between the postulated and observed changes in cancer detection rates. Curves are grouped by single-reader total screening mammogram volume; parenthetical numbers indicate the number of readers for each volume category. See text and Figure 3 for confidence intervals. Sample sizes refer to one arm of a two-arm trial.
|
|
View larger version (35K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 3: Modification of the probabilities shown in Figure 2 due to the effect of correlation in the observed single-reader cancer detection rates between two arms of a trial. The correlation was measured from two separate studies of screening mammography (see Statistical Analyses section). The Pearson correlation coefficient was 0.44. Shaded areas represent 95% confidence intervals. The confidence band with the effect of correlation taken into account (not shown) was 2%–4% wider than the corresponding confidence band without correlation.
|
|
The Pearson correlation coefficient for the observed single-reader cancer detection rate for mammograms read during the two time periods was 0.44 (
500 cases per reader per time period). Inclusion of this correlation reduced by 1%–6% the probability of observing no increase (or observing a decrease) in the cancer detection rate, increased by 4%–5% the probability of observing at least half the postulated increase for the new modality, and increased by 1%–13% the power for detecting a significant increase (Fig 3).
Using sample sizes of published trials and a hypothetical 70 000-patient-per-arm trial, we calculated the statistical power for trials of full-field digital mammography and computer-aided detection based on the postulated one additional cancer detected per reader per 1000 screening examinations for these new modalities. These power projections do not correspond to the actual power of the trials because, owing to a lack of necessary data, we postulated the benefit of the new modalities and calculated statistical power without including the effect of the correlation (increase in power) achieved by using the same readers or the same patient cohort across modalities. With these postulates, most published studies (Table 2) did not have sufficient numbers of patients and radiologists to achieve 80% power to demonstrate a statistically significant increase if the true increase in cancer detection rate were one cancer per 1000 screening examinations.
View this table:
[in this window]
[in a new window]
|
Table 2. Projected Statistical Power for Several Published Trials of Full-Field Digital Mammography and Computer-aided Detection
|
|
The Digital Mammographic Imaging Screening Trial (5), in which slightly over 49 500 patients at 33 clinical sites were accrued, with a minimum of two radiologists per site, would require about 85 radiologists to achieve 80% power for demonstrating the postulated one additional cancer detected per reader per 1000 screening examinations. Thus, it likely achieved sufficient statistical power with 153 radiologists participating and with patients matched across modalities (25). (Note also that this trial involved the use of receiver operating characteristic analysis, which is not commonly used in clinical trials and may offer greater statistical power for given numbers of examinations and readers than do analyses of detection rate.)
For a hypothetical study involving 70 000 patients per arm (Table 3), power varied from about 35% for trials that consisted of eight readers each with a case volume of 9000 examinations to about 84% for trials that consisted of 93 readers each with a case volume of 750 examinations, indicating greater power for trials that consist of more readers.
|
DISCUSSION
|
|---|
We found large variability in the observed cancer detection rates of screening mammography among individual radiologists. To a large extent, this variability is caused by a lack of precision in measuring the cancer detection rates of radiologists who have low case volumes. Insofar as precision is concerned, the variability in cancer detection rate reported here may not reflect differences in radiologists' interpretive skills; even if every radiologist were to detect, uniformly, the same fraction of cancers in screening populations, trials would still have interradiologist variability unless every radiologist read an extremely large case volume. However, our findings suggest that radiologists do not uniformly detect the same fraction of cancers in screening populations; otherwise, statistical power would have been the same for a given number of patients regardless of the number of readers, which is not consistent with our findings. Our findings also suggest that measuring the mean cancer detection rate for multiple readers yields a sample of interreader variability and therefore produces greater statistical power. Nevertheless, regardless of the origin, interreader variability has a detrimental effect on the outcome and statistical power of clinical trials that should not be ignored.
We found that it can be difficult to demonstrate significant improvement in the breast cancer detection rate for a new modality in clinical trials, because large numbers of patients and radiologists are required to achieve adequate statistical power. To detect a uniform increase in the cancer detection rate of one cancer per 1000 screening examinations in a two-arm trial with equal sample sizes in each arm would require approximately 93 radiologists and 70 000 patients per arm if each radiologist read a mean of 750 examinations, or 25 radiologists and 225 000 patients per arm if each radiologist read a mean of 9000 examinations. Detection of smaller, more realistic increases in the cancer detection rate would require even larger trials. Large patient samples are necessary for accrual of sufficiently large numbers of cancers. Large radiologist samples are an additional independent requirement for trial sample sizes.
The requirement of large numbers of patients and radiologists to obtain adequate statistical power suggests that it could be unduly burdensome to demonstrate improved cancer detection rates in clinical trials for a new breast imaging modality compared with the detection rates achieved by using the standard of care. It raises the question of whether it is effective, and indeed possible, to evaluate every new modality in randomized controlled trials with cancer detection rate as the end point. Unfortunately, our analysis results do not suggest an alternative solution, except that of using the same patient and radiologist cohorts across modalities, which increases statistical power by increasing correlation. Some increase in power can occur from matching low-case-volume radiologists across modalities; more increase can be expected from matching high-case-volume radiologists and matching patients across modalities. Statistical methods can also be used to adjust for differences in patient cohorts and increase the power (16).
The requirement of large numbers of patients and radiologists to achieve adequate statistical power also suggests that introducing new better modalities into clinical practice can be difficult. If routine audits of individual-practice performance parameters become commonplace, then one will be able to monitor whether practices that use new better modalities demonstrate consistently higher cancer detection rates than do comparable practices (with similar patient demographics and cancer incidence) that use conventional mammography. However, contrary to naïve expectations, the large variability associated with ascertainment of the breast cancer detection rate can make it difficult to observe an increase in the cancer detection rate in individual practices over short periods of time, giving rise to the possibility of inconsistent data. Inconsistent data on new modalities could cause confusion and controversy for consumers, payers, and public health policy makers and could even lead to erroneous calls for abandonment of new and better modalities.
Our study had a number of limitations. First, our analysis involved the use of observed single-reader cancer detection rates, which are more difficult to measure precisely than are multireader mean cancer detection rates. Our approach was necessary for estimating the outcome and statistical power of multireader trials involving an arbitrary number of readers without carrying out such trials.
Second, the distributions of single-reader cancer detection rates were estimated from 510 radiologists in the United States; thus, the results may be specific to the patient and radiologist populations of this study and not necessarily applicable to other populations. However, meaningful inferences can be drawn from this analysis because (a) the data used in this analysis were from seven regional breast cancer mammography registries in the United States, which represent a large sample of screening mammography in the United States, and were obtained from women with demographic characteristics similar to those of women in the general U.S. population (20); (b) the statistical analysis was tolerant of uncertainties in estimating the distribution of single-reader cancer detection rates, because it involved multiple summations of distributions for which the central-limit theorem applies; and (c) the 95% confidence intervals for power estimates were reasonably narrow.
Third, the postulated increase of one additional cancer detected per 1000 screening examinations is large, given the low breast cancer prevalence in average-risk screening populations. In our study population, the breast cancer detection rates for all methods combined and for screening mammography were 5.09 and 3.94 cancers per 1000 examinations, respectively, suggesting that one additional cancer detected per 1000 screening examinations is near the upper limit. Smaller increases in cancer detection rate would require even larger studies than those reported here.
Fourth, the postulated uniform increase in cancer detection rate is a simplistic possibility of how the use of new modalities may improve the cancer detection rate. Fifth, the results of this study are based largely on mathematical modeling and therefore need to be validated independently in observational studies. Finally, we did not analyze recall rate or specificity in this study. Imprecision in ascertainment and interreader variability for these measures will affect trial outcomes in a manner similar to imprecision in ascertainment and interreader variability for cancer detection rate reported herein.
In summary, low breast cancer prevalence in an average-risk screening population and interradiologist variability cause large variation in observed cancer detection rates, which can mask improvements in the cancer detection rates achieved with new breast imaging modalities in clinical trials and clinical practices. This masking effect may partially explain some of the lack of positive feedback on the initial clinical use of new technologies such as computer-aided detection, even when the technologies have been shown in preclinical retrospective trials to be beneficial. Investigators must consider interradiologist variability when calculating statistical power for trials of new screening modalities.
|
APPENDIX
|
|---|
For studies involving different readers across modalities (ie, excluding correlation), the calculation proceeded as follows: (For simplicity, we use in our illustrations the assumption that the single-reader cancer detection rate is normally distributed, even though we made no such assumption in the actual calculations.) First, we calculated the distribution of multireader mean observed cancer detection rates from the distribution of single-reader observed cancer detection rates. If the single-reader cancer detection rates were normally distributed, then the multireader mean cancer detection rate would be normally distributed with the same mean and with variance reduced by a factor equal to the number of readers in the study. Therefore, the multireader mean cancer detection rate would be less variable with more readers. For distributions that were not normal, this calculation required self-convolution of the single-reader cancer detection rate distribution (26), which we implemented by using a standard Fourier technique with which convolution (in the real domain) was replaced with multiplication (in the frequency domain).
Second, we calculated the distribution of multireader mean observed cancer detection rates for a postulated increase of one additional cancer detected per 1000 screening mammograms for every radiologist. This distribution was identical to the first distribution of mean cancer detection rates, with the exception that the mean of the second distribution was higher by an amount equal to the postulated increase. Last, we calculated the distribution of observed increases in the cancer detection rate from convolution of the two mean observed cancer detection rate distributions. If the multireader mean cancer detection rate were normally distributed, then the observed increase would also be normally distributed, with a mean equal to the postulated increase in the cancer detection rate (one cancer per 1000 examinations) and a variance twice as great as that of the multireader mean cancer detection rate. Therefore—and this is not necessarily intuitive—the observed increases in cancer detection rate were more variable than were the observed multireader mean cancer detection rates. The analytical expectations of normally distributed cancer detection rates were used to validate the method of numerical calculation that was used to calculate the results reported here.
For studies to measure the cancer detection rate of two modalities for the same readers (ie, considering correlation), we first calculated the probability of the observed single-reader increase in cancer detection rate before calculating the probability of the mean observed multireader increase. The distribution of observed single-reader increases in cancer detection rate was calculated from a bivariate joint-probability distribution of observed single-reader cancer detection rates measured from the two time periods. A postulated increase of one additional cancer detected per 1000 screening mammograms was added marginally to one of the two cancer detection rates in this distribution. Then, the probability of the observed single-reader increase in cancer detection rate was calculated by integrating this bivariate joint distribution along a diagonal direction. Finally, the distribution of observed multireader mean increases in cancer detection rate was calculated from the distribution of observed single-reader increases by means of the Fourier technique. The 95% confidence intervals (for calculations including and excluding the effect of correlation) were based on 1000 bootstrapping samples of the observed single-reader cancer detection rates.
|
ADVANCES IN KNOWLEDGE
|
|---|
- We provide a quantitative estimate of the interradiologist variability in observed breast cancer detection rates for screening mammography in typical clinical practices in the United States.
- We demonstrate a detrimental effect of this interradiologist variability on the outcome and the statistical power of clinical trials in which breast cancer detection rates achieved with standard mammography and new breast imaging modalities are compared, when different patients are enrolled in the two arms of the trial.
- We show the effect of mammogram volume on the variability in observed breast cancer detection rates and the effect of this variability on the outcome and the statistical power of clinical trials.
|
ACKNOWLEDGMENTS
|
|---|
We thank the Breast Cancer Surveillance Consortium investigators, participating mammography facilities, and radiologists for the data they provided for this study (grant numbers U01CA63740, U01CA86076, U01CA86082, U01CA63736, U01CA70013, U01CA69976, U01CA63731, and U01CA70040). The list of Breast Cancer Surveillance Consortium investigators and the procedures for requesting data for research purposes are provided at http://breastscreening.cancer.gov.
|
FOOTNOTES
|
|---|
Author contributions: Guarantor of integrity of entire study, Y.J.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; manuscript final version approval, all authors; literature research, Y.J., R.A.S.; statistical analysis, Y.J.; and manuscript editing, all authors C.E.M. and R.A.S. are shareholders of and receive royalties from Hologic (Bedford, Mass). Y.J. and C.E.M. receive research funding from Hologic.
|
References
|
|---|
- Lewin JM, Hendrick RE, D'Orsi CJ, et al. Comparison of full-field digital mammography with screen-film mammography for cancer detection: results of 4945 paired examinations. Radiology 2001;218:873–880.[Abstract/Free Full Text]
- Lewin JM, D'Orsi CJ, Hendrick RE, et al. Clinical comparison of full-field digital mammography and screen-film mammography for detection of breast cancer. AJR Am J Roentgenol 2002;179:671–677.[Abstract/Free Full Text]
- Skaane P, Young K, Skjennald A. Population-based mammography screening: comparison of screen-film and full-field digital mammography with soft-copy reading—Oslo I Study. Radiology 2003;229:877–884.[Abstract/Free Full Text]
- Skaane P, Skjennald A. Screen-film mammography versus full-field digital mammography with soft-copy reading: randomized trial in a population-based screening program—the Oslo II Study. Radiology 2004;232:197–204.[Abstract/Free Full Text]
- Pisano ED, Gatsonis C, Hendrick E, et al. Diagnostic performance of digital versus film mammography for breast-cancer screening. N Engl J Med 2005;353:1773–1783.[Abstract/Free Full Text]
- Freer TW, Ulissey MJ. Screening mammography with computer-aided detection: prospective study of 12 860 patients in a community breast center. Radiology 2001;220:781–786.[Abstract/Free Full Text]
- Vyborny CJ, Kukec C, Jiang Y, Doi K. Experience with computer-aided detection in a low-volume mammography clinic. In: Peitgen HO, ed. Digital mammography 2002. Heidelberg, Germany: Springer-Verlag, 2002; 387–390.
- Gur D, Sumkin JH, Rockette HE, et al. Changes in breast cancer detection and mammography recall rates after the introduction of a computer-aided detection system. J Natl Cancer Inst 2004;96:185–190.[Abstract/Free Full Text]
- Irwig L, Houssami N, van Vliet C. New technologies in screening for breast cancer: a systematic review of their accuracy. Br J Cancer 2004;90:2118–2122.[Medline]
- Jemal A, Murray T, Ward E, et al. Cancer statistics, 2005. CA Cancer J Clin 2005;55:10–30.[Abstract/Free Full Text]
- Elmore JG, Wells CK, Lee CH, Howard DH, Feinstein AR. Variability in radiologists' interpretations of mammograms. N Engl J Med 1994;331:1493–1499.[Abstract/Free Full Text]
- Beam CA, Layde PM, Sullivan DC. Variability in the interpretation of screening mammograms by US radiologists: findings from a national sample. Arch Intern Med 1996;156:209–213.[Abstract]
- Kerlikowske K, Grady D, Barclay J, et al. Variability and accuracy in mammographic interpretation using the American College of Radiology Breast Imaging Reporting and Data System. J Natl Cancer Inst 1998;90:1801–1809.[Abstract/Free Full Text]
- Schmidt RA, Newstead GM, Linver MN, et al. Mammographic screening sensitivity of general radiologists. In: Karssemeijer N, Thijssen M, Hendriks J, van Erning L, eds. Digital mammography. Dordrecht, the Netherlands: Kluwer Academic Publishers, 1998; 383–388.
- Sickles EA, Wolverton DE, Dee KE. Performance parameters for screening and diagnostic mammography: specialist and general radiologists. Radiology 2002;224:861–869.[Abstract/Free Full Text]
- Elmore JG, Miglioretti DL, Reisch LM, et al. Screening mammograms by community radiologists: variability in false-positive rates. J Natl Cancer Inst 2002;94:1373–1380.[Abstract/Free Full Text]
- Elmore JG, Nakano CY, Koepsell TD, Desnick LM, D'Orsi CJ, Ransohoff DF. International variation in screening mammography interpretations in community-based programs. J Natl Cancer Inst 2003;95:1384–1393.[Abstract/Free Full Text]
- Barlow WE, Chi C, Carney PA, et al. Accuracy of screening mammography interpretation by characteristics of radiologists. J Natl Cancer Inst 2004;96:1840–1850.[Abstract/Free Full Text]
- Smith-Bindman R, Chu P, Miglioretti DL, et al. Physician predictors of mammographic accuracy. J Natl Cancer Inst 2005;97:358–367.[Abstract/Free Full Text]
- Sickles EA, Miglioretti DL, Ballard-Barbash R, et al. Performance benchmarks for diagnostic mammography. Radiology 2005;235:775–790.[Abstract/Free Full Text]
- Ballard-Barbash R, Taplin SH, Yankaskas BC, et al. Breast Cancer Surveillance Consortium: a national mammography screening and outcomes database. AJR Am J Roentgenol 1997;169:1001–1008.[Free Full Text]
- Breast Cancer Surveillance Consortium Web site. http://breastscreening.cancer.gov. Accessed April 12, 2005.
- Ernster VL, Ballard-Barbash R, Barlow WE, et al. Detection of ductal carcinoma in situ in women undergoing screening mammography. J Natl Cancer Inst 2002;94:1546–1554.[Abstract/Free Full Text]
- American College of Radiology. Breast Imaging Reporting and Data System (BI-RADS). Reston, Va: American College of Radiology, 2003.
- Pisano ED, Gatsonis CA, Yaffe MJ, et al. American College of Radiology Imaging Network digital mammographic imaging screening trial: objectives and methodology. Radiology 2005;236:404–412.[Abstract/Free Full Text]
- Papoulis A. Probability, random variables, and stochastic processes. New York, NY: WCB/McGraw-Hill, 1991.
This article has been cited by other articles:
|
|
|
|
 
D. B. Kopans, Y. Jiang, D. L. Miglioretti, C. E. Metz, and R. A. Schmidt
History Repeats
Radiology,
February 1, 2008;
246(2):
645 - 646.
[Full Text]
[PDF]
|
|
|
|
|
|
|
S. Ciatto, N. Houssami, D. Gur, R. M. Nishikawa, R. A. Schmidt, C. E. Metz, J. F. Ruiz, S. A. Feig, R. L. Birdwell, M. N. Linver, et al.
Computer-aided screening mammography.
N. Engl. J. Med.,
July 5, 2007;
357(1):
83 - 84.
[Full Text]
[PDF]
|
|
|
eLetters:
Read all eLetters
- Designing Breast Screening Imaging Trials
- Stefano Ciatto, MD
- Radiology Online, 25 Jun 2007
[Full text]
TOP
ABSTRACT
INTRODUCTION
