| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Technical Developments |
1 From the Departments of Radiology (K.G.A.G., E.E.D., S.H.M., L.J.S.K.) and Pathology (J.L.P.), The Netherlands Cancer Institute, Antoni van Leeuwenhoek Hospital, Plesmanlaan 121, 1066 CX Amsterdam, the Netherlands. From the 2000 RSNA scientific assembly. Received September 25, 2001; revision requested December 10; revision received March 1, 2002; accepted April 2. Address correspondence to K.G.A.G.
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMaterials and Methods Results DiscussionREFERENCES |
|---|
© RSNA, 2002
Index terms: Breast, MR, 00.121412, 00.12143 • Breast neoplasms, diagnosis, 00.30, 00.311, 00.32 • Breast neoplasms, MR, 00.121412, 00.12143 • Cancer screening • Genes and genetics
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results DiscussionREFERENCES |
|---|
Differences in image interpretation guidelines contribute to the different specificities reported (9,13). Various investigators have focused on objective and quantitative rules to standardize the interpretation of contrast-enhanced MR images on the basis of temporal or morphologic characteristics of contrast material uptake (14–18). To date, most methods are based on the rating of features such as spiculation, washout, and peripheral enhancement by radiologists, followed by the merging of these ratings with an automated classifier (17–20). With this approach, the classification is objective, but careful construction of rating guidelines (21) may be necessary to avoid substantial inter- and intraobserver variations in ratings. Unlike in computerized analysis of mammograms (22–27), only a few investigators (28–30) have pursued both automated rating and automated classification of features to optimize the objectivity and the consistency of interpretation of contrast-enhanced MR images.
The goals of this study were directed toward optimization of the efficacy of MR screening programs for asymptomatic disease in women with increased lifetime risk of breast cancer. The first goal was to train and validate a lesion analysis system to distinguish accurately and consistently between malignant and benign lesions. The second goal of this study was to select a screening-specific operating point at which to attain a previously selected combination of accuracy in exclusion of malignancy of a lesion and in avoidance of biopsy of benign lesions.
|
Materials and Methods |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results DiscussionREFERENCES |
|---|
Indications for examination included findings at contrast-enhanced MR breast screening (in 30 women with 39 benign lesions and one woman with one malignant lesion), a question of extent and multifocality of the lesion before breast-conserving surgery (in 32 women with 37 malignant lesions), or an unclear diagnosis based on findings in conventional mammograms or ultrasonographic (US) images (in one woman with one benign lesion and two women with two malignant lesions). In this study, a minimum risk assessment of 15% was used for MR screening of women at increased lifetime risk for breast cancer. Breast lesions from populations other than a screening population were included in the current study to allow initial training of our system with sufficient characteristics of malignant lesions despite the low prevalence of malignancy in our screening population.
All premenopausal women underwent imaging between the 5th and 15th day of their menstrual cycle, and they underwent repeat imaging if large motion artifacts were apparent. The distribution of lesion size, as estimated from the MR images, is shown in Figure 1.
|
|
MR Imaging Technique
MR imaging was performed with a 1.5-T system (Somatom; Siemens Medical Systems, Erlangen, Germany) with fast low-angle shot, or FLASH, three-dimensional MR imaging with the patient prone and both breasts in a double-breast array coil. One series of precontrast MR images was obtained before power injection (4 mL/sec) of 0.1 mmol of gadoteridol (Prohance; Bracco-Byk Gulden, Konstanz, Germany) per kilogram of body weight and was followed up by acquisition of five series of postcontrast images at intervals of approximately 90 seconds. Acquisition of the first series of postcontrast MR images was initiated at 45 seconds after the start of contrast material injection. The following MR imaging parameters were used: T1-weighted sequence, repetition time msec/echo time msec of 8.1/4.0, reconstructed in-plane matrix of 256 x 256 pixels, isotropic in-plane resolution of 1.2 x 1.2 mm2 or 1.4 x 1.4 mm2, section thickness of 1.6 or 1.4 mm, and no fat suppression. Data for all images were transferred to the lesion analysis system by using the digital imaging and communications in medicine, or DICOM, protocol.
Lesion Analysis System
The lesions were analyzed retrospectively by a medical researcher (E.E.D.) and an imaging physicist (K.G.A.G.), who independently used the lesion analysis workstation developed at our institution. This system is an extension of collaborative work between three different institutions in Europe and the United States (29,33). The operators of the workstation were blinded to each other’s results and histologic outcome.
The reviewers each performed the following steps: Each lesion was localized by using the interactive detection function of the system and was verified on the basis of radiologists’ reports. The detection function offers fast exploration of the MR data by means of linked cursors in three reconstructed views (sagittal, transverse, and coronal) and with two types of processing (subtraction and washout) simultaneously (Fig 2).
|
Morphologic and temporal features (Table 2) were automatically rated with the system by using a region of interest with dimensions that were identical to those of the segmented lesion or were derived from it. Each feature was determined by means of analysis in three dimensions rather than in two dimensions; consequently, no "representative" section had to be chosen. In general, lesions with inhomogeneous uptake and spiculated boundaries have low values for smoothness of uptake. Lesions with well-defined margins have large mean values for margin sharpness, and lesions with partially well-defined margins have high variation in margin sharpness values. Each morphologic feature was derived from the first subtraction images (postcontrast MR images 1 minus the precontrast MR images) and from the subtraction images at the frame where the feature had the largest value (maximum across subtractions). Computation of the morphologic features has been described in detail (29). The temporal features of washout and signal-enhancing ratio have also been described (16,20).
|
Statistical analysis was performed (SPSS, version 9.0; SPSS, Chicago, Ill). The feature values were first tested for deviations from the normal distribution by using Q-Q plots and Kolmogorov-Smirnov tests. Log transformations were applied to improve the approximation to normal dispersion when necessary to avoid problems with subsequent statistical analysis. In an exploratory analysis, statistically significant indicators of malignancy were identified by means of the Wilks 
test. A P value of less than .05 was considered to indicate a significant ability to differentiate between benign and malignant lesion types.
Prospective performance of the lesion analysis system was estimated by means of leave-one-out cross validation, which involves training the system on all lesions but one, testing on the one lesion, and repeating the procedure until all lesions have been tested once. In addition to this commonly used technique, a conservative splitting validation was used, which involves random splitting of the database into an independent training set and a test set of equal size.
Linear discriminant analysis and stepwise selection (entry and removal threshold probabilities of F set to 0.10 and 0.15, respectively) were used to obtain a subset of features that makes a significant contribution to the classification into benign and malignant lesion types. All features listed in Table 2 were included in this selection process. Possible correlation between features was taken into account at this stage, but interaction terms were not entered because we expected the power of our statistics to be compromised in an attempt to avoid selection of terms that are predictive by coincidence. The likelihood of malignancy of each lesion was quantified by means of the Bayesian posterior probability (34), which was implemented in our lesion analysis system.
The statistical significance of the difference in results obtained by the two operators (ie, the interobserver variation) was assessed by using a paired-samples T test of the likelihood-of-malignancy values for each lesion. In addition, the statistical significance of the difference in the true-positive fraction (TPF) at selected false-positive fraction (FPF) was evaluated by means of receiver operating characteristic (ROC) analysis combined with univariate z score tests (35,36). The ROC software was provided by the University of Chicago (37,38).
Second goal: selection of screening-specific operating point.—A clinically relevant value of likelihood of malignancy was derived to serve as a threshold below which the system advises follow-up (benign lesion) or otherwise immediate biopsy (malignant lesion). In the following description, this threshold value will be referred to as the "operating point."
Parameterized ROC analysis (35) of the likelihood-of-malignancy values was performed to determine the sensitivity, or TPF, of the system for specificity, or 1 - FPF, values between 0% and 1%. Proper parameterization of each ROC curve was verified by means of visual comparison of the fitted curve with the unfitted data. Each point on the curve corresponds to a unique operating point. However, compared with the balanced training set used in the current study, the prevalence of benign lesions is much higher than that of malignant lesions in the actual screening population. To determine a clinically relevant operating point, each pair of (TPF, FPF) coordinates on the ROC curve was transformed to the corresponding pair of PPV (the fraction of lesions that is correctly assumed to be malignant) and NPV (the fraction of lesions that is correctly assumed to be benign) coordinates. The expected prevalence of malignancy (PM) in the screening population was taken into account in this transformation: PPV = TPF x PM/[(TPF x PM) + (1 - PM) x FPF], and NPV = [(1 - FPF) x (1 - PM)]/[(1 - FPF) x (1 - PM) + (1 - TPF) x PM].
The resulting (PPV, NPV) curve will be referred to as the "predictive curve." The (TPF, FPF) coordinates of the ROC 95% confidence band points—which are given by the software (37)—were also converted by using the equations for PPV and NPV to produce the confidence band for the predictive curve. Calculations were performed by implementing the equations in a spreadsheet program (Excel 2000; Microsoft, Redmond, Wash).
Because the analysis system is applied to only MR images in which enhancement of lesions is visible (images without enhancement do not raise suspicion), the estimation of prevalence of malignancy was also limited to the subset of enhancing lesions. The resulting PPV is similar to that for the entire population because the majority of images that show no enhancement indicate accurately the absence of disease (ie, a true benign finding), and true benign findings do not affect the TPF and the PPV. For similar reasons, however, the NPV for the entire population is higher than that for the subset of suspicious lesions, and the corresponding FPF is lower. The distinction between these two estimates of NPV will be indicated clearly in the remainder of this article by using the labels "subset of suspicious lesions" or "total population." The reason we refer to both estimates is that the use of the lesion analysis system is limited to the subset of suspicious lesions, whereas clinical effect is typically assessed in the total population.
The prevalence of malignant lesions in the total screening population (2%) was estimated preliminary from data for screening and those in the literature (10). The prevalence of benign lesions in the total population was estimated retrospectively from clinical screening data at our hospital (23%). Consequently, the prevalence of malignancy among suspicious lesions was set to 8%. The predictive curve was used to select an operating point that corresponds to clinically acceptable PPV where the NPV is at least 98% in the subset of suspicious lesions (ie, less than 2% chance of misinterpreting an enhancing malignant lesion as a benign lesion in the current screening round). The minimally accepted PPV in screening may vary in different regions. In this study, a value of 50% was used as a guideline (ie, one of two biopsies are of benign lesions). This guideline is in agreement with the assessment by other authors (39) and is comparable to that at mammography screening for women aged at least 50 years in the Netherlands (40–42).
|
Results |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results DiscussionREFERENCES |
|---|
The efficacy of each computer-rated feature, as shown in Table 3, was ranked from high to low ability to differentiate between benign and malignant lesions. Positive correlation indicates that higher values correspond with a higher likelihood of malignancy, whereas negative correlation indicates that higher values correspond with a lower likelihood of malignancy.
|
Stepwise selection resulted in the combination of one temporal and three morphologic features: washout, smoothness of uptake (maximum across subtractions), mean margin sharpness (first subtraction), and variation in margin sharpness (maximum across subtractions) (P < .001 for the combined features). Signal-enhancing ratio was excluded because of its strong correlation with washout. With cross validation and stepwise feature selection, the estimated prospective performance of the system was 0.95 ± 0.02 (area under the ROC curve [Az] ± 1 SD). This Az measure indicates the relative area under the ROC curve in Figure 3. A perfect performance without false-positive and false-negative findings would yield a relative area under the ROC curve of 1.0. Although variation in margin sharpness (maximum across subtractions) is a poor indicator of malignancy by itself (Table 3), it provides a statistically significant contribution in combination with the other selected features. On average, the margins of the malignant lesions were not only found to be more sharply delineated than those of the benign lesions, but they also showed less variation in sharpness.
|
Second Goal
Selection of screening-specific operating point.—Figure 4 illustrates that the maximum PPV of the system is estimated to remain greater than 50% at an NPV of at least 98% for prevalence values roughly between 0.5% and 10% (in the total screening population). The peak performance of the system is estimated to occur at a prevalence of approximately 1%. The predictive curve at the expected prevalence in our screening population (2%) is depicted in Figure 5. The estimated prospective performance at the operating point of PPV = 50% is summarized in Table 4. We emphasize that these are first estimates of prospective performance derived from the tuning process of the lesion analysis system.
|
|
|
At operating point of PPV = 50% (TPF = 85.7%), false-negative findings were of infiltrating ductal carcinoma (lesion volumes of 0.2, 0.3, 0.5, and 2.9 cm3). None of these misclassified lesions were associated with a genetic predisposition or familial involvement. False-positive findings were transient enhancement (lesion volumes of 0.5 and 1.4 cm3) and fibroadenoma (lesion volume of 0.5 cm3). A systematic difference in values for likelihood of malignancy obtained by the first and second operators could not be found (P > .6). In fact, the differences in computed likelihood were less than 5% for 72 of the 80 lesions. The differences were not caused by discrepancies in interpretation of the likelihood of malignancy by the two operators but by variations in selection of the seed point during segmentation of the lesions. The effect of these variations is small. At the selected operating point, only one of the 80 lesions was characterized differently, which resulted in a difference in TPF between the two operators that was less than 1.5% (not significant [P > .9]).
|
Discussion |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results DiscussionREFERENCES |
|---|
It is always possible to achieve a PPV equal to the prevalence of malignancy in the population because if all lesions are called malignant, then both the TPF and the FPF equal 1 and the PPV becomes equal to the prevalence. Consequently, in a symptomatic population associated with high prevalence of malignancy, the follow-up of suspicious findings by means of biopsy is more easily accepted than it is in an asymptomatic population associated with low prevalence. For symptomatic indications, computerized analysis may still help achievement of consensus in triple assessment (ie, physical examination, imaging, and pathologic examination), but more benefit is expected in screening, where the annual prevalence of malignancy is considerably lower than the desirable PPV. A first estimate of PPV obtained in our clinic for asymptomatic screening is 20%. The lesion analysis system has been tuned to achieve the desirable 50% PPV guideline at high NPV in a consistent and reproducible manner. Although the 50% PPV may not be representative for clinics in all regions, the system is easily tunable to other PPVs along an estimated curve of performance.
MR Screening
Initial literature reports on the screening of asymptomatic women at increased lifetime risk indicate detection of breast tumors in a substantially earlier stage of malignancy when the women are under surveillance rather than when they are referred because of symptoms (10). In 109 asymptomatic women at high risk, Tilanus-Linthorst and colleagues (43) report 100% sensitivity (three of three tumors), 94% specificity (100 of 106 women), and no false-negative findings with a follow-up of 1 year (ie, NPV = 100%, PPV = 33%). In 105 asymptomatic women, Kuhl and colleagues (12) report 100% sensitivity (nine of nine tumors), 95% specificity (91 of 96 women), and no false-negative findings with follow-up of at least 1 year (ie, NPV = 100%, PPV = 64%). However, this relatively large PPV is difficult to compare with findings in other studies because the false-positive findings were taken from a smaller population (n = 105) than were the true-positive findings (n = 192) to guarantee at least 1 year follow-up of benign findings. In addition, the prevalence of malignancy (approximately 5% [nine of 192 women]) was found to be somewhat higher than that reported in other studies and that observed in our hospital (approximately 2%).
Differences in sensitivity and specificity reported in the examination of the asymptomatic population compared with that in the symptomatic population may be attributed to the small number of screening cancers accumulated thus far. In addition, differences in specificity may be caused by differences in the prevalence of normal findings (no enhancement) in the symptomatic and asymptomatic groups because normal findings are typically interpreted as true-negative findings until proven otherwise at follow-up. Other differences may be caused by discrepancies in follow-up guidelines for suspicious findings and differences in image interpretation (9,12). Nonetheless, first results in the screening of asymptomatic women at high risk indicate high certainty of excluding malignant disease. At a PPV of much less than 50%, however, fewer biopsies of benign lesions are desired without compromising the NPV.
Features
The margin of malignant lesions was found to be sharper (ie, better defined) than that of benign lesions (P < .01); this observation may seem inconsistent with observations of suspicious masses in mammograms. However, margins depicted in mammograms reflect differences in tissue density, whereas margins depicted in contrast-enhanced MR images reflect differences in uptake of contrast material. The higher speed of uptake and peripheral enhancement (44) of malignant lesions may explain the superior sharpness of their margins compared with that of benign lesions. In the present study, the mean margin sharpness in the first subtraction images appeared to be a better indicator of malignancy than was the mean margin sharpness maximized for all the subtraction images for all time frames (maximum across subtractions). Apparently, the difference in definition of benign and malignant margins decreases over time.
The temporal features were computed in a region of interest that was limited to 80% of the segmented lesion with the largest feature values. Smaller percentage values had no obvious effect on the performance of the system. The 80% value was chosen to reduce possible influence of the partial volume effect while maintaining a sufficient number of voxels in smaller lesions.
The features selected in a pilot study that was performed with a much smaller independent database (29) are in agreement with a subset of the features selected in the current study. In the pilot study, smoothness of uptake and variation in margin sharpness resulted in an area under the ROC curve of 0.96. In the current study, temporal features were added, a larger independent set of lesions from a different imager and clinic were used, and semiautomated segmentation was used rather than manual segmentation. Consequently, two additional features were selected, which resulted in an area under the ROC curve of 0.95. These observations are an encouraging indication of the generalizing behavior of the computerized analysis.
Performance and Limitations
To achieve highly reproducible results, semiautomated segmentation of lesions has been combined with computerized rating and classification of features. The results indicate a clinically acceptable PPV (approximately 50%) at an NPV of at least 99% in the total screening population. Moreover, variations in results obtained by two independent operators of the system were found to be negligibly small.
The lesion analysis system is intended to be a tool that will provide radiologists with objective and consistent guidelines to attain prior selected PPVs and NPVs derived from statistics in past cases. Detection of suspicious lesions is accomplished exclusively by radiologists, although the interactive detection function of the system is currently in clinical use as an add-on feature to the existing image reading system. The efficacy of contrast-enhanced MR imaging in the detection of lesions in a screening population lies outside the scope of our current study, but it is one of the main focus points of multiinstitutional trials on MR screening. The decision to focus on computerized characterization before computerized detection follows from the known high sensitivity of contrast-enhanced MR imaging for invasive lesions. Because of the limited sensitivity of contrast-enhanced MR imaging to depict ductal carcinoma in situ (45,46), it is especially important that any enhancement related to ductal carcinoma in situ is interpreted accurately and consistently.
Cross validation of the system resulted in correct labeling of all cases of ductal carcinoma in situ in our database—although there were only a limited number—as malignant disease. Because automated rating and classification are currently tested independently from the determinations of the radiologists in our clinic, the interaction between the radiologists and the system has not yet been examined. We will investigate this topic in the future.
The current study was limited because most benign lesions were found at screening of asymptomatic women at increased lifetime risk, whereas the majority of malignant lesions were found in symptomatic patients. In this preliminary study, we combined the two populations to allow initial training of the system. Because of the low prevalence of malignancy in the screening population, approximately 2,000 asymptomatic women would need to be screened to allow detection of 40 malignant lesions. Provision of comparably sized sets of benign and malignant lesions is crucial to train the system to recognize the typical differences in characteristics. We emphasize that the combining of populations requires careful analysis and interpretation of data, while taking the limitations into account.
First, the distribution of sizes of benign and malignant lesions in the asymptomatic screening population is currently unknown and may be different from that in a symptomatic population. For example, the differences in size between the benign and malignant lesions in our database may be due in part to the fact that most malignant lesions were symptomatic (and were therefore more likely to be detected at a larger size), but they may also be a result of the inherently large difference in growth rates between benign and malignant lesions, as would also be true in an exclusively screening population. Given this uncertainty, size was not used as a feature in the current study, nor were any features used that were significantly correlated with size.
Second, differences may exist in MR image characteristics between sporadic cancers and those associated with a genetic predisposition. Differences in histologic phenotype, such as pushing margins, have been reported in cancers associated with BRCA1/2 (47), but the statistical significance of these observations on contrast-enhanced MR screening images is yet to be determined in larger sets of malignant tumors from both sporadic and high-risk populations. It is possible that pushing margins will improve the depiction of margin sharpness and decrease the variation in sharpness along the margin. If so, identification of these cancers by means of the analysis system will be facilitated, but such an effect can be assessed only in actual prospective application. We could not find an indication for increased failure of the system to characterize the 11 malignant lesions in our database that were associated with genetic predisposition or familial involvement.
Third, with a combined symptomatic and asymptomatic population, the PPV and NPV for an exclusive screening population cannot be determined in a straightforward manner. By transforming the ROC curve to a predictive curve based on a priori estimates of prevalence of malignancy in a screening population, the PPV and NPV in a screening population could be estimated from our training database. Although the actual prevalence may vary slightly in prospective application, the effects of such variation are not expected to result in large differences in performance of the system. Prospective estimates of PPV remained greater than 50% for a wide range of prevalence values.
To estimate the prospective performance of the system on the basis of the training population, a common test (cross validation) was used, and results were corroborated with conservative splitting validations. In addition, a simple classification method (linear-discriminant analysis) was used to minimize the risk of obtaining results that were specific for only the training population in our study. Nonetheless, prospective testing of the system on a larger set of multiinstitutional screening images remains necessary to validate its clinical performance and to fine-tune selection of the operating point. Currently, all lesions that are visible on MR images obtained in our clinic are added to the lesion analysis system.
Another topic of future investigation concerns the effect of the imaging technique. In the current study, a standard fast low-angle shot three-dimensional acquisition technique was chosen that balances temporal and spatial resolution at a voxel size of approximately 0.003 cm3. The size of the training population did not allow breakdown of the performance into multiple categories of lesion volume without the loss of statistical power. It is likely, however, that differentiation between very small benign and malignant lesions is more challenging than that between larger lesions. One way to compensate for this effect is to allow the analysis system to automatically adjust the operating point for small lesions to a different (but known a priori) trade-off between NPV and PPV. Automatic switching between operating points is feasible because the volume of the lesion is known after the segmentation process is complete.
In conclusion, the lesion analysis system has been tested successfully in a clinical environment. The combination of computer-rated washout, smoothness of uptake, mean margin sharpness, and variation of margin sharpness yields a significant contribution to characterization into benign and malignant lesion types. In the current study, first estimates of prospective performance indicate that the system is capable of excluding malignant disease with high confidence and reproducibility, at a clinically acceptable PPV in a screening setting. The ability to reproducibly select operating points on the basis of prior estimates of NPV and PPV may contribute substantially to the success of contrast-enhanced MR screening programs for asymptomatic women at increased lifetime risk, but prospective validation in an exclusive screening population remains necessary.
|
ACKNOWLEDGMENTS |
|---|
|
FOOTNOTES |
|---|
Author contributions: Guarantors of integrity of entire study, K.G.A.G., E.E.D., L.J.S.K.; study concepts, K.G.A.G., S.H.M., E.E.D.; study design, K.G.A.G., E.E.D.; literature research, K.G.A.G., E.E.D.; clinical studies, E.E.D., S.H.M.; experimental studies, K.G.A.G., S.H.M.; data acquisition, E.E.D., S.H.M.; data analysis/interpretation, K.G.A.G., J.L.P., L.J.S.K.; statistical analysis, K.G.A.G.; manuscript preparation, K.G.A.G., E.E.D., S.H.M.; manuscript definition of intellectual content, K.G.A.G., J.L.P., L.J.S.K., E.E.D.; manuscript editing and revision/review, K.G.A.G., S.H.M., E.E.D.; manuscript final version approval, all authors.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results DiscussionREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  S. Schrading and C. K. Kuhl Mammographic, US, and MR Imaging Phenotypes of Familial Breast Cancer Radiology, January 1, 2008; 246(1): 58 - 70. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
T. C. Williams, W. B. DeMartini, S. C. Partridge, S. Peacock, and C. D. Lehman Breast MR Imaging: Computer-aided Evaluation Program for Discriminating Benign from Malignant Lesions Radiology, July 1, 2007; 244(1): 94 - 103. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 E. Yeh, P. Slanetz, D. B. Kopans, E. Rafferty, D. Georgian-Smith, L. Moy, E. Halpern, R. Moore, I. Kuter, and A. Taghian Prospective Comparison of Mammography, Sonography, and MRI in Patients Undergoing Neoadjuvant Chemotherapy for Palpable Breast Cancer Am. J. Roentgenol., March 1, 2005; 184(3): 868 - 877. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
E. E. Deurloo, S. H. Muller, J. L. Peterse, A. P. E. Besnard, and K. G. A. Gilhuijs Clinically and Mammographically Occult Breast Lesions on MR Images: Potential Effect of Computerized Assessment on Clinical Reading Radiology, March 1, 2005; 234(3): 693 - 701. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
