| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
1 Department of Radiology, University of Wisconsin Hospital and Clinics,
E3/311-3252 Clinical Sciences Center, 600 Highland Ave., Madison, WI
53792-3252.
2 Department of Biological Regulation, Weizmann Institute of Science, P. O. Box
26, Rehovot, Israel 76100.
Received September 7, 2001;
accepted after revision June 5, 2002.
Funding for the first 18 patients in this study was from the Weizmann
Institute of Science, Rehovot, Israel. Funding for the remaining patients was
from the United States Israel Binational Science Foundation, Jerusalem,
Israel, grant #98-00461. The work at the Weizman Institute was supported by
Sir David Alliance, CBE, United Kingdom, and by the Fred and Andrea Fallek
Professional Chair in Breast Cancer Research.
Abstract
 Top Abstract IntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
SUBJECTS AND METHODS. Fifty-seven women with 68 pathologically confirmed breast lesions were imaged (45 masses, 23 microcalcifications). Seven consecutive 2-min three-dimensional gradient-recalled echo acquisitions were performed after suitably timed gadopentetate dimeglumine injections. We derived a composite parametric image from three judiciously selected time points (three-time-point method), using a model-based kinetic algorithm. In this composite image, color brightness and hue signify contrast uptake and washout characteristics related to the product of microvessel surface area and permeability, as well as to the extracellular volume fraction. The reviewer was provided with the slice location of lesions, but with no other radiographic or clinical information. The reviewer then classified the lesions as benign or malignant using a 5-step receiver operating characteristic scale.
RESULTS. Observers using the three-time-point method correctly diagnosed 27 of 31 malignant and 31 of 37 benign lesions (sensitivity, 87%; specificity, 84%). The area under the receiver operating characteristic curve was 0.911. False-negative results were found for three patients with low- to intermediate-grade ductal carcinoma in situ and one patient with 5-mm invasive ductal cancer. For the 45 solid lesions, sensitivity and specificity were 96% and 82%, respectively.
CONCLUSION. Application of the three-time-point method permitted, in most cases, differentiation of malignant and benign lesions, even in the presence of complex breast enhancement patterns. Sensitivity for solid tumors was higher than for ductal carcinoma in situ.
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
Prior animal and pilot human clinical studies by Degani et al. [15] and Furman-Haran et al. [16] have resulted in the development of a parametric method for breast MR imaging diagnosis. The method is model-based [17] and is predicated on the premise that for a diagnosis using breast MR imaging to be successful across imaging platforms and among various physicians, a level of standardization is required. The hypothesis is that recording and processing data from three high-spatial-resolution images obtained at the properly chosen time points can be used to accurately differentiate benign and malignant breast lesions. Furman-Haran et al. [18] have shown that high spatial resolution is required for accurate diagnosis using the parametric method. We present results from a collaborative effort to perform the first clinical study of the three-time-point method in a typical heterogeneous population found in the United States.
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
Practical breast MR imaging entails a compromise between image quality and temporal resolution, and thus a limited number of time points are obtained to represent the postcontrast-injection SI time course. The algorithm we propose detects the SI at each image location, pixel by pixel, for one unenhanced time point and two contrast-enhanced time points (hence, the term "three time point"). The algorithm then codes the SI changes among the three time points using color intensity and color hue as follows: First, color intensity codes the rate at which the SI changes between the first and second time points, with a resolution of 256 intensities in which dark colors signify slow change and bright colors signify rapid change. Second, color hue is a measure of contrast agent washout and is coded depending on the SI change between images recorded at the second and third time points. If the SI increases from the second to the third time point, that location is coded blue; if it stays constant (± 10%), it is coded green; and if it decreases, it is coded red.
Of course, for a given SI time curve, the color coding of each individual point will be highly dependent on the selection of the time points for SI measurement. The optimal points are predicted by forming a calibration map specific to the set of physical imaging parameters (e.g., TR, TE, flip angle) and average published estimates of the contrast agent pharmacokinetic parameters in humans and an average T1 of the breast at the imaging magnetic field strength [23]. In a three-time-point calibration map the x-axis corresponds to the EVF, and the y-axis, to K. The aim is to choose imaging time points that divide the K-EVF plane so as to provide the best discrimination between benign and malignant lesions. The iteration method by which the optimal imaging times are chosen is shown in Figure 1A,1B,1C,1D. For our imaging parameters, the predicted optimal time points were 0, 2, and 6 min.
|
|
|
|
Despite our initial hope that three time points would clearly separate the data, some lesions displayed bright green pixels (rapid and significant SI increase without washout), so that their characteristics were worrisome in regard to EVF and K (i.e., borderline malignant). To more clearly identify these pixels, we formed another parametric image, this time based on data at 0, 4, and 8 min after injection. For these images, the same criteria were then used for interpretation as were used for the 0-, 2-, and 6-min images. It was hoped that this strategy would result in increased sensitivity, with no or minimal loss of specificity. For all patients, we tabulated the effect that this additional information had on making our final diagnosis.
Patient Recruitment
Our mammography section typically obtains 12,000 mammograms and performs
200-300 biopsies per year. The women in our study were randomly selected from
patients seen in our radiology department for palpable masses, mammographic or
sonographic abnormalities thought by the radiologist and surgeon to require
biopsy. However, we did not accept women whose prebiopsy studies indicated a
high likelihood of a cyst because most mildly complex cysts are benign. We
believed that the inclusion of these patients would have diluted the intent of
this study—to test a method for distinguishing benign from malignant
lesions in the more challenging cases. During the recruitment period
(September 1, 1998—October 20, 2000), 62 women were recruited and imaged
using the three-time-point method.
After having the procedure for breast MR imaging explained to them, all patients signed an informed consent approved by the human subjects committee of the institution performing the clinical trial. Of these 62 volunteers imaged using the three-time-point protocol, data for five patients was discarded for the following reasons: election of follow-up rather than biospy in three patients, incomplete correlation with the biopsy specimen in one patient, and motion in one patient. We present results based on the three-time-point method to prospectively make a diagnosis in 57 women who had a total of 68 pathologically confirmed lesions.
MR Imaging
Imaging was performed on a 1.5-T scanner (Sigma; General Electric Medical
Systems, Waukesha, WI), using a dedicated receive-only breast coil (Phased
Array Breast Coil; MRI Devices, Waukesha, WI). A three-dimensional
gradient-echo acquisition was performed using the following parameters: TR/TE,
15/4.2; flip angle, 30°; field of view, 16-18 cm; matrix size, 256 x
256; excitations, 1; slice thickness, 2.2-3.0 mm. Seven consecutive
acquisitions of 56 slices each (interpolated from 28 slices) requiring
slightly more than 2 min per acquisition were obtained over 14 min and 45
sec.
Gadodiamide (Omniscan; Nycomed Laboratories, Princeton, NJ) was injected, using a dose of 0.1 mmol/kg, 3 min after the beginning of the scanning series (i.e., 1 min after the start of the second scanning sequence). Contrast agent was administered at 2 mL/sec, followed by 15 mL of saline flush, also delivered at 2 mL/sec, using an automated pump (Spectris MR Injector; Medrad, Indianola, PA).
We used the following terminology: The 0-min image refers to the first of the seven acquisitions obtained during an imaging sequence, before contrast injection. The 2- and 6-min images refer to the time after contrast injection and not to the time since initiation of the seven-acquisition sequence. Thus, we actually used the third and fifth acquisitions within the seven acquisition series in the three-time-point calculation when speaking of the 2- and 6-min images.
Image Interpretation and Data Analysis
The seven acquisitions were first examined by the radiologist at the
clinical site on a standard workstation supplied by the scanner manufacturer,
using dedicated software (Functool; General Electric Medical Systems). The
radiologist's criteria were typically those used in daily interpretation of
clinical MR imaging examinations (this radiologist typically interpreted 10-15
breast MR imaging examinations per month). These criteria combine morphologic
and time-curve analyses in a manner similar to that described by Kuhl et al.
[13]. Because the radiologist
was involved in recruitment of the patients, he was aware of the imaging and
clinical data. The radiologist assigned to the imaging results a malignant or
benign category. In parallel to the qualitative interpretation, MR images were
sent by file transfer protocol from the laboratory performing the clinical
trial to the research laboratory for analysis by the three-time-point
algorithm. In cases of mammographic lesions, digital photographs of the
mammograms with the lesions circled were sent to the research site. For
sonographic lesions, the location and estimated size of the lesions were sent
for correlation with MR imaging. However, in neither case was the diagnosis of
the radiologist or clinical information provided. On the basis of the
mammograms and sonograms, the decision as to the location of the lesion in the
MR images was made by the radiologist and the researcher, with the radiologist
making the final determination in cases of disagreement.
Interpretation of images at the research site involved examining each of the parametric slices computed from imaging time points at 0, 2, and 6 min for a coherent group of pixels that could indicate a lesion. Prior experience with the three-time-point method has shown that when a lesion has greater than 15% red pixels, it is likely malignant; if less than 10% red pixels are present, the lesion is likely benign. Benign lesions typically show greater than 50% blue pixels and low color intensity (Fig. 2A,2B).
|
|
We generated an additional parametric map for all lesions, using data from acquisitions at 0, 4, and 8 min, already available as part of the seven-acquisition series. With the two new time points, if green pixels became red and reached the fraction threshold for cancers, the lesion was diagnosed as malignant (Fig. 3A,3B,3C). The final diagnosis was graded using the receiver operating characteristic categories: 1, very likely benign; 2, probably benign; 3, indeterminate; 4, possibly malignant; 5, very likely malignant
|
|
|
We planned to determine, in general terms, how these five categories came into alignment with the Breast Imaging Reporting and Data System (BI-RADS) [24] scoring system used for mammography. In calculating sensitivity and specificity, grades 4 and 5 lesions were called three-time-point positive, whereas grade 1, 2, or 3 lesions were called three-time-point negative. Because a grade 3 in clinical practice might align with a BI-RADS category 3, implying some type of follow-up requirement, increased medical costs, and, for some patients, potential anxiety, the number of times this score was offered was also tabulated.
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
Fifty-seven diagnoses were made by excisional biopsy; eleven, by fine-needle aspiration. The distribution of our pathologically confirmed diagnoses were as follows: 31 malignant lesions, 21 lesions with a diagnosis of invasive ductal cancer, eight lesions with a diagnosis of ductal carcinoma in situ (DCIS) (two were large lesions, the first measuring 8 cm in diameter and the second measuring 4 cm in diameter with a 3-mm focus of microinvasion), one tubular cancer, and one tubulobular cancer. Thirty-seven lesions were benign: of these, 14 were fibroadenomas. The remaining 23 lesions frequently showed mixed pathology, with the most frequent diagnoses being sclerosing adenosis and fibrocystic change.
Observers using the three-time-point method correctly diagnosed 27 of 31 malignant lesions (grade 4 or 5) and 31 of 37 benign lesions (grade 1, 2, or 3). The data were used to generate a receiver operating characteristic curve (Fig. 4) using the ROCFIT software package (Metz CE, University of Chicago, Chicago, IL). The area under the receiver operating characteristic curve (Az) was 0.911, and the standard deviation in Az was 0.036. Only one lesion was graded as indeterminate (grade 3): findings at pathology showed a benign intraductal papilloma. The reviewers' study for the three-time-point method yielded the following sensitivities and specificities, respectively: all lesions (n = 68), 87% and 84%; solid masses (n = 45), 96% and 82%; and microcalcifications (n = 23), 63% and 81%.
|
Four results were false-negative: one small solid lesion and three lesions showing micro-calcifications without mass. The three foci of microcalcifications were all pathologically diagnosed as DCIS: two intermediate grade (8 and 14 mm) and one low grade (four ducts). The typical three-time-point pattern associated with these false-negative results consisted of a mixture of bright green and blue pixels, with only occasional scattered red pixels (Fig. 5A,5B,5C). Five other foci of DCIS were correctly diagnosed as malignant: one was high grade (comedo), three were intermediate grade, and one was DCIS detected in a single duct, which was not otherwise specified by the pathologist. The misdiagnosed solid lesion measured 5 mm in diameter and had a pathologic diagnosis of invasive ductal cancer. In the same MR imaging slice, two larger lesions measuring 8 and 14 mm were correctly diagnosed as malignant.
|
|
|
Six false-positive results were obtained: one 11-mm focus of fibrocystic change, one 3-mm focus of sclerosing adenosis, one 9-mm intraductal papilloma, one 10-mm focus of mixed pathology (fibrocystic change and fibroadenoma), one 3-mm intraductal papilloma, and one 9-mm fibroadenoma. Of 14 fibroadenomas, 12 were correctly diagnosed as benign (Fig. 6A,6B,6C), whereas two were thought to be malignancies.
|
|
|
We generated an additional parametric map using data from a second set of three-time-points (0, 4, and 8 min) already available as part of the seven-acquisition series. The purpose was to increase the sensitivity of the method for lesions with borderline parameters indicating cancer. Use of the additional images also added confidence to a diagnosis in case of artifact during the 0-, 2-, or 6-min images. In 61 of 68 patients, using interpretation criteria noted previously, we found that data from the first set of three time points (0-, 2-, or 6-min) was sufficient for a confident diagnosis. In the remaining seven cases, the 0-, 2-, or 6-min-derived images, enough green pixels became red to reach the fraction threshold for cancer. All of these lesions were therefore diagnosed as malignant. When correlated with pathology, five of seven lesions were malignant (two DCIS, two infiltrating ductal cancers, and one tubular cancer), and two were benign (sclerosing adenosis and fibrocystic change). In no lesion did the fraction of red pixels decrease and lead to misdiagnosis of a malignant lesion. Thus, our strategy for use of the additional time points was successful in that it increased sensitivity to a greater extent than it reduced specificity.
The radiologist, using qualitative clinical criteria based on wash-in and washout curves and morphology, achieved the following sensitivities and specificities, respectively: all lesions (n = 68), 93% and 82%; mass lesions (n = 23), 100% and 82%; microcalcifications (n = 23), 71% and 69%.
Seven patients (12.5%) may have benefited from having breast MR imaging: First, in four patients, a second or even third focus of malignancy was detected on MR imaging alone and changed the surgical approach (Fig. 7A,7B,7C). Second, in one patient, in whom the radiologist, based on mammographic results, suspected a 1-cm tumor, MR imaging revealed a 4-cm tumor, which was subsequently diagnosed as an infiltrating ductal cancer. The information was provided to a surgeon at an outside hospital. We are uncertain whether this information was used in extending the margins. Third, in one patient, in whom a well-marginated lesion with rapid contrast agent washout was seen, the surgeon declined imaging guidance when he removed the palpable lesion. However, when postsurgical pathologic examination described only benign breast tissue, the surgeon was urged to repeat sonography, which confirmed lack of excision. Subsequent imaging-guided excision resulted in a diagnosis of benign papilloma. Fourth, in one patient, after a failed mammographically guided localization of a mammographically vague and sonographically occult lesion, MR imaging—guided needle localization was used to excise the lesion and arrive at the diagnosis—invasive ductal cancer.
|
|
|
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
The decreased sensitivity of breast MR imaging for DCIS has been acknowledged [9, 10], and in this respect, the current three-time-point method has a similar shortcoming. In our population of 31 malignant lesions, of eight patients with DCIS, five lesions were correctly diagnosed as malignant. Although all enhancing DCIS showed a high fraction of bright green pixels, further cases must be studied to define new three-time-point—specific criteria for DCIS diagnosis.
Although we envisioned interpretation based on only three time points, in the early stage of our study, we incorporated a parametric image based on a second set of three time points. This strategy was designed to increase the sensitivity of the method in detecting cancers whose K and EVF characteristics were borderline (e.g., enhanced rapidly without washout). Of the seven lesions in which the second set of three time points was used to change an initial diagnosis, five lesions were cancers and two lesions were benign. Therefore, we now recommend routine incorporation of data from the additional images as part of the interpretation process.
One of the authors, a radiologist experienced in breast MR imaging using subjective criteria similar to those of Kuhl et al. [13], achieved results similar to those achieved by the three-time-point method, although with perhaps a higher sensitivity. However, unlike the scientists making decisions using the three-time-point method, the radiologist, as director of the clinical effort, must have known the radiographic and clinical information associated with the patients. For example, Figure 5A,5B,5C illustrates a false-negative interpretation using three-time-point criteria. The radiologist, biased by the mammographic appearance of the microcalcifications, prospectively diagnosed this lesion as a malignancy. Thus, as was true for Muller-Schimpfle et al. [12] and Daniel et al. [9], a radiologist armed with relatively basic decision-making criteria performed as well as or better than one using a parametric automated method.
Evaluation by a radiologist of a breast with multiple enhancing areas using subjective criteria is impractical. For example, in one of the research patients referred for a palpable mass, three-time-point MR imaging showed at least 100 other smaller well-marginated enhancing foci (Fig. 6A,6B,6C), obscured on mammography by dense glandular tissue. Perhaps 10-20 of these foci exceeded 5 mm, the threshold above which most radiologists attempt to render a diagnosis. This case is one in which a parametric approach has its greatest application. Even for single lesions, the qualitative diagnostic accuracy of a radiologist in general practice may not equal that of a breast MR imaging specialist. For radiologists in general practice, a parametric diagnostic program may well be a valuable aid.
A multiplicity of approaches has been used in morphologic and contrast-enhancement magnitude and time course in differentiating benign and malignant breast conditions. Three recent reviews discuss this topic [14, 25, 26]. Some recent approaches are comparable to the three-time-point method. Kinkel et al. [21] used a three-time-point acquisition high-spatial-resolution method in which each acquisition took 5 min. Imaging was performed before and 2.5 and 7.5 min after gadopentetate dimeglumine administration. The reason for selection of these points was not provided but was probably based on clinical experience in detecting early wash-in and washout. In this retrospective study, washout was detected in 29 of 34 malignant lesions and 3 of 23 benign lesions, yielding a sensitivity and specificity of 85% and 87%, respectively. When morphology was added to the descriptors, sensitivity and specificity both increased to a remarkable 96%. Because the study of Kinkel et al. was not prospective, an estimated 90% sensitivity and 87% specificity were calculated for the clinical setting. These estimates are close to results achieved in our prospective study. Furthermore, although the three-time-point method does not currently incorporate morphologic descriptors into the decision-making process, high-spatial-resolution information is available on the unprocessed three-time-point images. The patient population of Kinkel et al. was different from ours: only three cases of DCIS (8% of their total versus 25% of our total of malignant cases) and only three fibroadenomas (13% of their total versus 38% of our total of benign cases) were present in their study population.
Daniel et al. [9] used a high-temporal-resolution imaging method and classified lesions with both K21 (a parametric descriptor associated with washout) and qualitative curve descriptors. Correlation of the results of this study with our results is somewhat difficult because threshold values were chosen with the clinical goal of purposely reducing false-positive results to achieve an extremely high negative predictive value. Furthermore, the region of interest chosen to calculate parameters and classify curves encompassed the entire lesion, whereas in the three-time-point method, there is point-by-point high resolution analysis of lesions rather than gross averaging of results. We speculate that the highly detailed spatial distribution of these color descriptors might be of use in distinguishing benign from malignant masses. For example, a large proportion of bright red pixels that form a peripheral ring may be considered a more sensitive indicator of malignancy than a similar number of pixels uniformly distributed throughout the tumor.
Currently, the three-time-point method has a few limitations. One technical limitation is that the method relies on high spatial resolution, and although our pixel size was 0.625 mm2 x 1.5 mm (interpolated from 3 mm), this size may not be sufficient for accurate diagnosis of 5-mm or smaller lesions.
Another potential technical limitation is that patient movement may cause problems with any parametric program that examines the time-dependent value of SI, including the three-time-point approach. In our clinical practice, in which we routinely view subtraction images, movement problems are minor and may complicate interpretation of images in less than 1% of patients. The scientific site involved in our research had the capability of simple in-plane motion correction. In practice, motion appeared to produce a scattered low-signal-intensity colored area, easily distinguished from true malignancy, and no reregistration was performed for any of the patients. Results from only one patient had to be discarded because of gross motion. Other researchers have developed methods for registration that could possibly improve future performance of any parametric method [27, 28].
Finally, as tested for our study, the three-time-point method was limited to examining only one breast at a time; this procedure is problematic if future use in screening of highrisk patients is to be considered. It is now possible, however, to reduce the TE and TR and to apply a 512 x 512 matrix with a field of view that covers both breasts. Whether this protocol would work has yet to be confirmed.
A biologic limitation of our study is that DCIS, especially low and intermediate grade, may not exhibit the three-time-point EVF and K characteristics that result in the early washout pattern of invasive cancer [29]. As we accumulate more cases of DCIS, we shall be closely studying the spatial distribution of pixels as well as their number to possibly detect other ways of making this difficult diagnosis.
Yet another potential limitation is that normal intrammammary lymph nodes may exhibit an enhancement profile similar to that of malignancy. However, because lymph nodes have a distinct morphology (fatty hilum with entering vessel), this profile should not be a problem for lymph nodes greater than 5 mm.
In conclusion, we have evaluated a model-based algorithm for diagnosis of gadolinium-enhanced breast MR imaging. This method precalculates an optimized protocol and analyzes the dynamic images at pixel resolution providing a standardized computer-aided diagnostic tool. Testing at 1.5-T showed overall sensitivity and specificity of 87% and 84%, respectively. This overall sensitivity was lowered by the fact that 25% of our malignancies were cases of DCIS, in which MR imaging is known to have decreased sensitivity—for solid lesions, sensitivity was 95%, and specificity was 82%. Although the experienced breast MR radiologist may achieve a similar accuracy for single lesions, this technique allows rapid reliable evaluation of multiple enhancing regions. Future research will quantify the internal distribution of color pixels in a lesion and incorporate this information and gross morphologic descriptors into a generalized neural network for comprehensive diagnosis of lesions encountered during breast MR imaging.
Acknowledgments
We thank Charles E. Metz, University of Chicago, for providing us with
software (ROCFIT) and assistance in calculation of the receiver operating
characteristic curve derived from our data; Ravi Karra and David Manthei, for
help in finding patient volunteers and for collating and archiving the data
involved in this study; and Kristi Anderson and Dawn Hay, for their help in
scheduling research patients into an already tight clinical schedule.
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
This article has been cited by other articles:
|
|
 |
|
  N. H. G. M. Peters, I. H. M. Borel Rinkes, N. P. A. Zuithoff, W. P. T. M. Mali, K. G. M. Moons, and P. H. M. Peeters Meta-Analysis of MR Imaging in the Diagnosis of Breast Lesions Radiology, December 1, 2007; 246(1): 116 - 124. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 T. E. Yankeelov, K. J. Niermann, J. Huamani, D. W. Kim, C. C. Quarles, A. C. Fleischer, D. E. Hallahan, R. R. Price, and J. C. Gore Correlation Between Estimates of Tumor Perfusion From Microbubble Contrast-Enhanced Sonography and Dynamic Contrast-Enhanced Magnetic Resonance Imaging J. Ultrasound Med., April 1, 2006; 25(4): 487 - 497. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. I. Wiener, K. J. Schilling, C. Adami, and N. A. Obuchowski Assessment of Suspected Breast Cancer by MRI: A Prospective Clinical Trial Using a Combined Kinetic and Morphologic Analysis Am. J. Roentgenol., March 1, 2005; 184(3): 878 - 886. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 M. Dadiani, R. Margalit, N. Sela, and H. Degani High-Resolution Magnetic Resonance Imaging of Disparities in the Transcapillary Transfer Rates in Orthotopically Inoculated Invasive Breast Tumors Cancer Res., May 1, 2004; 64(9): 3155 - 3161. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
