| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Breast Imaging |
1 From the Department of Radiology (H-1307), Stanford University Medical Center, 300 Pasteur Dr, Stanford, CA 94305-5105 (R.L.B., D.M.I.); R2 Technology, Inc, Los Altos, Calif (K.F.O.); and the Department of Radiology, University of California, San Francisco (E.A.S.). From the 1999 RSNA scientific assembly. Received February 25, 2000; revision requested April 9; revision received August 14; accepted September 6. Address correspondence to R.L.B. (e-mail: birdwell@leland.stanford.edu).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODS RESULTS DISCUSSIONREFERENCES |
|---|
MATERIALS AND METHODS: A multicenter retrospective study accrued 1,083 consecutive cases of breast cancer detected at screening mammography. Prior mammograms were available in 427 cases. Of these, 286 had lesions visible in retrospect. The 286 cases underwent blinded review by panels of radiologists; a majority recommended recall for 112 cases. Two experienced radiologists compared prior mammograms in 110 of these cases with the subsequent screening mammograms (when cancer was detected), noting mammographic characteristics of breast density, lesion type, size, morphology, and subjective reasons for possible miss. The prior mammograms were then analyzed with a CAD program.
RESULTS: There were 110 patients with 115 cancers. On the prior mammograms with missed cancers, 35 (30%) of the 115 lesions were calcifications, with 17 of 35 (49%) clustered or pleomorphic. Eighty of the 115 (70%) were mass lesions, with 32 of 80 (40%) spiculated or irregular. For calcifications and masses, the most frequently suggested reasons for possible miss were dense breasts (12 of 35; 34%) and distracting lesions (35 of 80; 44%), respectively. CAD marked 30 (86%) of 35 missed calcifications and 58 (73%) of 80 missed masses.
CONCLUSION: Detection errors affected cases with calcifications and masses. CAD marked most (77%; 88 of 115) cancers missed at screening mammography that radiologists retrospectively judged to merit recall.
Index terms: Breast neoplasms, diagnosis, 00.32, 00.812 • Cancer screening, 00.32, 00.812 • Computers, diagnostic aid, 00.32, 00.812
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTS DISCUSSIONREFERENCES |
|---|
In contrast, interval cancers are not limited to those that can be seen in retrospect; they are typically defined as cancers in patients presenting with clinical findings before the next scheduled mammogram. This time interval is typically 1 year but could be 18–24 months, depending on the individual practice location (2,3). Interval cancers account for a mean of 7%–13% of the breast cancers in women aged 40 years and older undergoing annual screening (4). Compared with screening-detected cancers, these tumors tend to be larger, of higher grade, and more likely to have lymph node involvement (5,6).
This study was designed to assess cancers that were not prospectively diagnosed at screening mammography, did not go on to manifest clinically, and were detected on a later screening mammogram and proved at biopsy to be cancer. We use the term missed cancers for those cases in which a screening-detected cancer can be seen as a focal abnormality on a retrospective review of the prior mammogram, for which the majority of a blinded panel of five radiologists would have recalled the patient on the basis of a review of the prior mammogram.
Missed cancers can be due to suboptimal performance in perception of lesions and analysis of perceived findings (7). Methods suggested to decrease the number of missed cancers include training, experience, continuing education, prospective double reading, retrospective evaluation of missed cases, and computer-aided detection (CAD).
CAD is designed to provide visual prompts to the interpreting radiologist in specific areas on the image. Many different CAD programs have been developed to detect masses and microcalcifications (8–14). Several factors affect the reported performance of CAD, including lesion subtlety, the size and makeup of the training and study sets, and the type of validation method used (11). One group of investigators suggests that the use of CAD in screening mammography may result in increased effectiveness without an increase in the work-up rate (15).
The purpose of this study was twofold: (a) to determine whether unique mammographic findings and lesion characteristics could be identified in cases where screening-detected breast cancers were judged to have been present on a prior mammogram by the majority of a blinded panel of radiologists and (b) to determine the performance of CAD in marking the missed cancers on the previous screening mammograms.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTS DISCUSSIONREFERENCES |
|---|
All of the 1,083 current screening mammograms (those in which cancer was detected) were evaluated by a site radiologist who had knowledge of the location of the biopsy-proved cancer. The site radiologist used this information to mark the location of the lesion or lesions on transparent film overlays and to document the type of lesion (ie, microcalcifications, masses with or without calcifications, spiculated mass/architectural distortion, other). A total of 427 of the 1,083 cases (a) had available prior screening mammograms that had been obtained 9–24 months (mean, 14 months) before the current mammograms and (b) met the criteria for the initial study (15).
One of three designated radiologists (not the site radiologists) independently reviewed these 427 prior mammograms (n = 242, 103, 82) to determine in retrospect whether the subsequently detected cancers were visible. For the review, the designated radiologist compared the prior mammogram with the current mammogram, including the overlay created by the site radiologist indicating the location of the biopsy-proved lesion on the current mammogram. If the cancer was deemed visible, the designated radiologist created a second overlay marking the location of the lesion on the prior mammogram. Cancers were judged as visible in 286 (67%) of the 427 cases. These 286 prior mammograms with visible cancers were divided into four sets, each with approximately 75 cases. Each case set was enlarged by the addition of three groups of mammograms: (a) five cases in which no abnormalities could been seen on the prior mammograms, (b) 20 current cases in which an abnormality was confirmed by the site radiologist and judged to be relatively subtle, and (c) 20 randomly chosen negative cases from participating mammography facilities, confirmed to be negative by at least one subsequent negative mammographic examination.
To determine whether the retrospectively visible lesions should have been worked up further, a blinded review by four panels of radiologists, each with five members, was conducted. The experience of the panel radiologists included a mean of 17 years practicing radiology (range, 3–35 years), 50% with a primary focus in mammographic interpretation, with a mean of 300 screening mammograms interpreted per month (range, 40–1,000 cases). The current and normal cases were common to all four case sets. The mean sensitivity of the panel radiologists (measured on the current mammograms) was 84%, and the mean specificity (measured in the normal cases) was 81%. Detailed findings are reported elsewhere (15). The members of each panel independently assessed one of the four case sets according to the American College of Radiology’s Breast Imaging Reporting and Data System (BI-RADS) to determine for each case whether a finding was present that would be recommended for either additional imaging (BI-RADS category 0) or biopsy (BI-RADS categories 4 and 5). The study design dictated that the panel radiologists were only shown the prior mammograms; no earlier mammograms were presented, and only the age of the patient was provided. Each mammogram was assigned a consensus value of 0–5, reflecting the number of panel radiologists who assessed the visible prior mammograms as needing further evaluation in the location later proved to have cancer.
There were 112 cases in which three, four, or five of the five panel radiologists recommended further work-up. We define this subset of 112 cases as missed cancers using the reasoning that if a majority of radiologists in an independent blinded review interpreted the prior mammograms as abnormal, then the findings were visible abnormalities that should have prompted action. The other 174 cases where zero, one, or two of the five panel radiologists recommended further work-up represent a different group of mammographic findings. These are not considered to have been missed cancers according to our definition because 80% of them, even in retrospect, show no mammographic characteristics that would prompt further work-up (16).
Radiologist Case Review
Two radiologists who specialize in breast imaging (R.L.B., D.M.I.) together reviewed the 112 visible prior mammograms, excluding two cases; one was excluded because of an ambiguous lesion location on the overlays created by the designated radiologist, and the second was excluded because there was ambiguity as to whether or not an earlier breast biopsy had been performed. The 110 patients with findings visible on the prior mammograms comprise the study population. In these 110 patients, 115 cancers were diagnosed, of which 82 of 115 (71%) were invasive and 33 of 115 (29%) were ductal carcinoma in situ only. In the 91 patients for whom staging information was available, 32 (35%) had stage 0 cancer. One of the 33 lesions of ductal carcinoma in situ was found in a patient who had a synchronous invasive cancer. Forty-one (45%) of the 91 patients had stage I cancer, and 18 (20%) had stage II. No cases of stage III or IV cancer were reported. The mean age was 63 years (range, 38–85 years). Further breakdown of age on the basis of cohort decade showed that 1% of the women were 30–39 years old, 10% were 40–49, 25% were 50–59, 31% were 60–69, 29% were 70–79, and 4% were 80–89.
In contrast to the earlier study (15), all of the mammograms reviewed in this part of the study were high-quality digital copy mammograms printed on a laser imager system (HQ969; Imation, Oakdale, Minn) at 100-µm resolution. To evaluate the quality of the copies, two radiologists (R.L.B., D.M.I.) performed a side-by-side comparison study of 20 original mammograms compared with copy mammograms. These 20 cases were selected to represent both masses and calcifications. The original and the copy film cases were each given a numerical rating of 1–5 (1 = unable to read, 3 = acceptable, 5 = good) and a narrative description of the relative visibility of the characteristics of the lesion. The mean quality rating for the original mammograms was 4.5 and for the copy mammograms 4.4. From the narrative descriptions, there were no cases in which it was believed that the copy quality hindered the evaluation of the cases.
The purpose of our review of the missed cancers was to assess mammographic characteristics and possible reasons for detection and interpretation errors that might have led to the false-negative reading of the 110 visible prior mammograms. This consensus retrospective review was done in a nonblinded fashion and included a side-by-side comparison of the current and the visible prior mammograms. No patient information, examination dates, or pathologic staging was available to the reviewers. The study cases were reviewed on a motorized mammogram viewer with the four-view current mammograms in the bottom row (with overlays showing the documented biopsy location of the lesions) and the four-view prior mammograms in the top row.
The mammographic characteristics that were evaluated included breast density, lesion type, size, location, and depth within the breast. The BI-RADS lexicon was used to describe the breast composition and the lesion type (17). In an effort to explain possible reasons why the lesions identified by a majority of the panel radiologists to merit further work-up were not prospectively read as abnormal on the prior mammograms, subjective factors were recorded. We recorded "detection" factors as those relating to issues of perception, for example, lesions located near the edge of the glandular tissue or the film edge and lesions depicted with sufficient subtlety so as to confound the process of detection. "Interpretation" factors were those cases in which the abnormality likely was perceived but was probably assessed as negative or benign, for example, clusters of calcifications too few in number to alert the need for action, readily detectable lesions having mammographic characteristics probably deemed as benign appearing, and lesion size too small to prompt work-up. Film technique was assessed as to whether positioning, compression, motion, or artifacts affected lesion visibility on one or on both views. In the subjective assessment of reasons for possible miss, any cases in which the factors were considered as possible hindrances to accurate film evaluation were recorded. Multiple factors could be recorded for each lesion.
CAD Case Review
The CAD system used in this study (R2 Technology V2.0; R2 Technology, Los Altos, Calif) consisted of a laser digitizer, a computer using proprietary signal-processing algorithms, and a customized motorized viewer with video display monitors. The original prior mammograms were digitized at 50-µm resolution with 12 bits of gray scale. The processing algorithm searched for features suggestive of microcalcifications (clusters of bright spots marked by solid triangles) and masses or architectural distortions (regions of high density with or without radiating lines marked by asterisks). Low-resolution depictions of the digitized mammograms are viewed on the motorized viewer display monitors. The CAD marks generated by the computer algorithm appear over the center of the region of interest on the low-resolution CAD images. The CAD system is designed such that in normal clinical use, the interpreting radiologist is prompted to reevaluate suspicious features marked by the CAD system after the mammograms have first been reviewed.
A determination was made by the two experienced radiologists about whether CAD correctly marked the 115 lesions on the 110 prior mammograms. The following criteria were used: (a) a marker of the correct type (triangle for calcifications, asterisk for masses) marked the lesion on one or both standard mammographic views; (b) if a lesion had both mass and calcification characteristics, then a marker of either type counted as a correct mark; (c) if a calcification lesion was spread out over a large area, then a marker indicating any part of the suspicious area was counted as a correct mark; and (d) for lesions noted as being visible on only one view, only a mark on that view was considered.
The sensitivity of the CAD system was computed as the ratio of the number of lesions correctly marked (as defined by the previous criteria) divided by the total number of lesions. Note that for patients with multiple lesions, each lesion was considered separately. Statistical tests of significance were performed by using contingency tables and a 
2 test for concordance. The CHITEST function (Microsoft EXCEL; Microsoft, Redmond, Wash) was used to calculate the P value. Statistical significance was inferred for P values less than .05. A logistic regression analysis was performed by using the JMP system (SAS Institute, Cary, NC) to confirm the conclusions of the univariate 2 test. An unpaired two-tailed t test was used to measure the significance of the difference in lesion size between the lesions that were and were not marked by CAD. The EXCEL program (Microsoft) was used to calculate the P value of the test. The calcification and mass lesions were measured separately.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTS DISCUSSIONREFERENCES |
|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Tables 4 and 5 list the recorded frequency of the subjective factors assigned to missed cancers on the visible prior mammograms and the performance of CAD for each of those factors. The two most common detection factors for masses were distracting lesions and lesions at the edge of the glandular tissue; CAD detected 27 of 35 (77%) and 29 of 34 (85%) of these cases, respectively. The most common interpretation factors were (a) lesions with lucent areas within the mass and (b) the area in question looked like normal tissue. CAD detected 38 of 48 (79%) and 20 of 28 (71%) of these cases, respectively.
The two most common factors influencing nondetection of calcifications were dense breasts and distracting lesions. CAD detected 10 of 12 (83%) and eight of 10 (80%) of these cases, respectively. CAD results for interpretation errors for calcification cases were calcifications appeared benign (11 of 15; 73%) and too few calcifications (seven of nine; 78%). For cases in which film technique contributed to missing a lesion, CAD detected 14 of 16 (88%) of the calcification lesions and 22 of 37 (59%) of the masses.
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTS DISCUSSIONREFERENCES |
|---|
Many of the large studies outlining the characteristics of missed cancers include cancers in patients presenting with clinical findings between screening appointments. These so-called interval cancers are more likely to occur with metastases and have a poorer prognosis (3,5). Because we excluded such cases from our study, we will compare our results only with those published studies that also excluded interval (clinically detected) cancers, to allow a more accurate comparison of mammographic characteristics. This exclusion of studies that included interval cancers leaves only a small number of publications for direct comparison (1–3,18).
Similar to Harvey et al (18), we found missed cancers to be evenly distributed in mostly fatty and mostly dense breasts, 51% (59 of 115) and 49% (56 of 115), respectively, although Bird et al (1) found a greater percentage of missed cancers in dense breasts. Lesion location, size, and type, if they are reported at all, are diversely categorized in the literature.
The most common lesion types reported in studies of missed cancers are mass or density in 19%–64%, calcifications in 18%–28%, mass with calcifications in 2%, and architectural distortion in 4%–12% (1–3,18). Our study reports similar findings for these lesion types, except that we do not report architectural distortion as a separate lesion type. The study of Bird et al (1) was the only one to include the category of developing density, a finding they reported in 18% of the cases. This type of lesion can be identified only when two mammographic examinations obtained at different times are compared. Neither the panel radiologists nor the two experienced radiologists had any mammograms obtained earlier than the prior mammograms for review. This also meant that it was not possible to assess the stability of findings on the prior mammograms, a factor that may be used to dismiss a mammographic finding at the time of interpretation.
The most striking difference in the frequency distribution of lesion types involves the asymmetric densities reported by Harvey et al (18) as 53% of their missed cancer cases. A conversation with Jennifer A. Harvey, MD (February 2000), clarified that they assigned the term "asymmetric density" to lesions seen only on one view and to those lesions without discrete borders. Rather than a true variation in mammographic characteristics, the differences in lesion types may be simply due to differences in terminology.
Sickles (19) reported the mammographic features of 300 consecutive nonpalpable breast cancers. His findings of the mean age of the patient (57 years), tumor location (52% in upper outer quadrant), breast density (>50% in less dense breasts), and cases without regional or systemic metastases (>80%) were similar to our cases. There are some differences in the number of women in the 40–69-year-old age group (80%, as compared with 67% in our study), the number of cancers larger than 20 mm (2% vs 13%), and the number of patients presenting with classic signs of malignancy (39% vs our finding of 50%). We had a greater number of women over the age of 69 (32% vs 20% in Sickles [19]), which could account for some of these differences.
One hypothesis that has been suggested for the fact that some cancers are not prospectively detected is that they may have imaging characteristics different from those that are detected at screening (20). The fact that more of our cases demonstrated mammographic signs known to be associated with cancer (including linear branching calcifications and spiculated masses) is somewhat unexpected. To the contrary, Sickles (19) reported more than half of his 300 cases as having more subtle signs of possible cancer (this may simply be due to variations in the use of terminology), with a large number of those cases interpreted as having indirect signs of malignancy. Only two of our 115 lesions were characterized as a focal asymmetry or a single dilated duct. In the study of Sickles (19), the cases described as having the more subtle signs of cancer included all architectural distortion, all developing densities, all masses with indistinct margins, and all pleomorphic but not linear and branching calcifications (19). Also, the study by Sickles (19) was performed before BI-RADS terminology was introduced. Differences in the classification scheme for lesions limits comparison between the studies.
The majority of screening mammograms demonstrate no features worrisome for cancer. The interpreting radiologist must balance the need to detect the subtle but nonspecific findings that may indicate early cancer with the less important goal of avoiding a high recall rate. Double reading of screening mammograms has been advocated as a way to increase sensitivity, specificity, or both. Studies have shown that different readers overlook different findings and that having more than one person interpret the images from an examination increases the detection of breast cancer by 7%–15% (21–23). How the mammograms are double read is variable. The second reader can try to detect what the first reader missed (to increase sensitivity) (21), or he or she can try to dismiss what the first reader found (to increase specificity), or there can be a combination of both.
Advances in computer technology, ready adaptation of radiologic images to digital format, and Food and Drug Administration approval of a CAD system have increased interest in computers as "second readers" or prompters to assist the interpreting radiologist in the clinical setting (12). Bick et al (24) collected a data set of 105 consecutive screening-detected cancers with a large proportion of in situ or invasive cancers 1 cm or less in size. Prior work typically looked at larger and even palpable lesions (12–14). These investigators combined three CAD system detection schemes and reported that 88% (92 of 105) of the lesions were marked by CAD, with a mean of slightly less than five marks per mammographic image. In our study, the mean number of marks was 1.1 per mammographic image (a mean of 4.3 marks per four-view mammogram).
Current-generation CAD systems have a high detection rate for features known to be suspicious for breast cancer on screening mammograms, and these systems provide the potential for improved performance (12). Recent work (15) reported that the performance of CAD on 1,083 consecutive screening-detected cancers was 84% (906 of 1,083) overall, with 99% (400 of 406) of the calcification cases and 75% (506 of 677) of the mass cases being correctly marked. Because CAD is more sensitive for calcifications than masses, a study population more heavily weighted with calcification cases will inherently show higher overall CAD sensitivity. Further, when the CAD system was evaluated in the clinical setting, there was no increase in the radiologist’s work-up (or callback) rate before (8.3%) and after (7.6%) CAD installation. This study showed that CAD had the potential to reduce the 21% false-negative rate of the original interpreting radiologist by 77%.
We grouped lesion size into those smaller than 11 mm, those 11–20 mm, and those larger than 20 mm, in part to reflect the size differentiation used in the tumor-node-metastasis (TNM) staging system of the American Joint Committee on Cancer (25). Invasive cancers that are smaller than 11 mm, 11–20 mm, and larger than 20 mm are grouped accordingly as T1a or b, T1c, and T2, leading to differences in treatment options and prognosis. CAD performance for masses was not dependent on lesion size. However, for the 35 cases of calcifications, the five lesions not marked by CAD were significantly smaller (mean size, 7.2 mm) than those marked by CAD (mean size, 19.3 mm) (P = .007). The difficulty in accurately measuring the extent of calcification lesions on screening mammograms may affect the validity of this result.
Our study found that film technique indeed is a factor in detection misses. Bird et at (1) hypothesized that lesions located in the retroglandular regions were more often overlooked on the basis of the exposure factors used to optimize the glandular tissue density. Baines et al (26) reported that observer error and technical problems were responsible for delayed detection in 22% of the screening-detected breast cancers and in 35% of the interval breast cancers. We judged an equal number of the missed cancers (17%) to have been affected by inadequate positioning, compression, or motion artifact. Faced with unacceptable film quality, the radiologist must obtain additional mammograms. It is doubtful that suggested mechanisms to decrease false-negative interpretations, such as double reading, will have any effect on cancers missed because of poor film technique. CAD performance for masses and for calcifications when there were technical deficiencies was somewhat lower than when there were no such deficiencies. CAD should not be expected to compensate for poor image quality; it is designed to work only by using images that would be judged acceptable for interpretation without the use of CAD.
Detection factors relating to missed cancer differ in frequency on the basis of the lesion type. The dense breast was cited as a factor for lesion miss more often in cases of missed calcifications (34%) than for masses (14%). CAD performed similarly in the detection of calcifications (83% detected) and masses (82%) in those cases for which "dense breasts" were cited as a factor related to lesion miss.
Problems with the detection of masses appeared linked with their location at the edge of the glandular tissue and with the presence of other distracting lesions; the performance of CAD was 85% and 75%, respectively, for these factors. These detection problems may benefit from closer focused analysis of these areas of the film by the interpreting radiologist.
Of particular interest to us was the relatively frequent finding of cancers seen as masses containing internal radiolucency. The presence of fat within a mass is usually regarded as increasing the likelihood for a benign process; however, 59% of our missed-cancer masses had lucent areas within them, suggesting that the importance of this imaging feature needs to be downgraded relative to other imaging characteristics that indicate the presence of malignancy (shape, margins, associated findings).
All currently available CAD systems produce marks that point to areas not actually representing breast cancer. Our study of 110 mammograms containing 115 missed cancers showed a mean of 4.3 marks per four-view case, of which one-third marked the missed cancers. The majority of marks will indicate areas that the radiologist will choose to dismiss because no abnormal-appearing characteristics are judged to be present. However, the additional time involved to examine the low-resolution CAD images, reevaluate the screening mammograms accordingly, and decide which, if any, of the CAD marks indicate initially nondetected lesions that deserve further work-up must be assessed. The efficacy of CAD also must be compared with that of double reading by either the same or a different radiologist, as must the relative costs.
In summary, the mammographic characteristics of missed cancers are similar to those that are usually considered to be suspicious for breast cancer; a present-generation CAD system marked a high percentage of these cancers. CAD marked most of the cancers missed at screening that a majority of panel radiologists judged on blinded independent review to merit recall for further imaging. If radiologists using CAD prospectively also judge these cases to merit recall, then the use of CAD will substantially reduce the frequency of missed cancers. Because CAD is more sensitive in detecting microcalcifications than masses, enhancement in CAD algorithms should concentrate on improved methods to detect malignant masses. Additional research also is needed to determine the relative cost versus benefit of all second-reading approaches compared not only with conventional single-reader interpretation, but also with one another.
|
ACKNOWLEDGMENTS |
|---|
|
FOOTNOTES |
|---|
Abbreviations: BI-RADS = Breast Imaging Reporting and Data System, CAD = computer-aided detection, CC = craniocaudal, MLO = mediolateral oblique
Author contributions: Guarantor of integrity of entire study, R.L.B.; study concepts and design, R.L.B., D.M.I.; literature research, R.L.B., K.F.O.; clinical studies, R.L.B., D.M.I.; data acquisition, R.L.B., D.M.I., K.F.O.; data analysis/interpretation, all authors; statistical analysis, all authors; manuscript preparation, R.L.B., K.F.O.; manuscript definition of intellectual content, editing, revision/review, and final version approval, all authors.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTS DISCUSSIONREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  H. F. Boehm, T. Schneider, S. M. Buhmann-Kirchhoff, T. Schlossbauer, D. Rjosk-Dendorfer, S. Britsch, and M. Reiser Automated Classification of Breast Parenchymal Density: Topologic Analysis of X-Ray Attenuation Patterns Depicted with Digital Mammography Am. J. Roentgenol., December 1, 2008; 191(6): W275 - W282. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 F. J. Gilbert, S. M. Astley, M. G.C. Gillan, O. F. Agbaje, M. G. Wallis, J. James, C. R.M. Boggis, S. W. Duffy, and the CADET II Group Single Reading with Computer-Aided Detection for Screening Mammography N. Engl. J. Med., October 16, 2008; 359(16): 1675 - 1684. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 M. F. Chiang, R. Gelman, S. L. Williams, J.-Y. Lee, D. S. Casper, M. E. Martinez-Perez, and J. T. Flynn Plus Disease in Retinopathy of Prematurity: Development of Composite Images by Quantification of Expert Opinion Invest. Ophthalmol. Vis. Sci., September 1, 2008; 49(9): 4064 - 4070. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 H. Singh, S. Sethi, M. Raber, and L. A. Petersen Errors in Cancer Diagnosis: Current Understanding and Future Directions J. Clin. Oncol., November 1, 2007; 25(31): 5009 - 5018. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
D. Georgian-Smith, R. H. Moore, E. Halpern, E. D. Yeh, E. A. Rafferty, H. A. D'Alessandro, M. Staffa, D. A. Hall, K. A. McCarthy, and D. B. Kopans Blinded Comparison of Computer-Aided Detection with Human Second Reading in Screening Mammography Am. J. Roentgenol., |
