Michael A. Jacobs, PhDPeter B. Barker, DPhilDavid A. Bluemke, MD, PhDCindy Maranto, RTCheryl Arnold, RTEdward H. Herskovits, MD,

PhDZaver Bhujwalla, PhD

Index terms:Breast neoplasms, MR, 00.12141Cancer screeningComputers, diagnostic aidMagnetic resonance (MR), tissue

characterization

Published online10.1148/radiol.2291020333

Radiology 2003; 229:225–232

Abbreviations:IAD � intraset euclidean distanceIED � interset euclidean distanceISODATA � iterative self-organizing

dataROC � receiver operating characteristic

1 From the Russell H. Morgan Depart-ment of Radiology and Radiological Sci-ence, Johns Hopkins University Schoolof Medicine, Traylor Bldg, Room 217,712 Rutland Ave, Baltimore, MD 21205(M.A.J., P.B.B., D.A.B., C.M., C.A.,E.H.H., Z.B.); and F. M. Kirby ResearchCenter for Functional Brain Imaging,Kennedy Krieger Institute, Baltimore,Md (P.B.B.). Received March 22, 2002;revision requested June 11; final revisionreceived January 13, 2003; acceptedFebruary 24. M.A.J. supported by NIHT32 CA09630 and by Breast CancerSpore P50 CA88843. P.B.B. supportedby NIH R21CA/RR91798. D.A.B. sup-ported by NIH RFA CA96012. Addresscorrespondence to M.A.J. (e-mail:mikej@mri.jhu.edu).

Author contributions:Guarantors of integrity of entire study,M.A.J., Z.B.; study concepts, M.A.J.,D.A.B., Z.B.; study design, M.A.J.,D.A.B., P.B.B., Z.B.; literature research,M.A.J., D.A.B.; clinical studies, M.A.J.,D.A.B., C.A., C.M.; data acquisition,C.A., C.M., M.A.J.; data analysis/inter-pretation, M.A.J., D.A.B., P.B.B., Z.B.;statistical analysis, M.A.J., E.H.H.; manu-script preparation, M.A.J., P.B.B., D.A.B.,Z.B.; manuscript definition of intellec-tual content, M.A.J., Z.B.; manuscriptediting, M.A.J., P.B.B., D.A.B., Z.B.;manuscript revision/review and finalversion approval, all authors© RSNA, 2003

Benign and Malignant BreastLesions: Diagnosis withMultiparametric MR Imaging1

PURPOSE: To both develop and use a tissue signature method for the identificationand classification of breast lesions and healthy breast tissue with magnetic reso-nance (MR) imaging.

MATERIALS AND METHODS: Thirty-six patients underwent breast MR imaging(T1- and T2-weighted imaging and three-dimensional T1-weighted imaging withand without contrast material enhancement), followed by biopsy or mastectomyand histopathologic analysis. Tissue cluster analysis was performed by using theiterative self-organizing data technique to identify glandular, adipose, and lesiontissue signature vectors. Glandular and lesion tissue vectors were characterized byangular separation from the reference adipose tissue vector. Differences in angularseparation of histologically proved benign and malignant lesion groups were eval-uated with an independent t test. The usefulness of the angular separation model fordistinguishing benign from malignant lesions was evaluated with nonparametricreceiver operating characteristic curve analysis.

RESULTS: The model enabled successful identification and characterization ofbreast lesion tissue clusters in all patients; 18 lesions were benign, and 18 weremalignant. Angular separation � SD was 17.8° � 6.1° between adipose tissue andmalignant lesions and 29.0° � 11.2° between adipose tissue and benign lesions.Angular separations of benign lesions and malignant lesions were significantlydifferent (P � .002), with a specificity of 78% and sensitivity of 89% at a cutoff valueof 21°. Significant differences in angular separation from adipose tissue also werefound between glandular tissue and lesion tissue (P � .001) and, in glandular tissue,between patients with benign lesions and those with malignant lesions (P � .04).The area under the receiver operating characteristic curve was 0.84.

CONCLUSION: Multispectral analysis of conventional breast MR images based onthe iterative self-organizing data model and on measurement of angular separationbetween tissue signature vectors may enable automated lesion identification andclassification.© RSNA, 2003

Magnetic resonance (MR) imaging is increasingly used for the diagnosis and characteriza-tion of suspected breast lesions identified with mammography. Contrast material–en-hanced MR imaging has high sensitivity (approximately 90%) but lower specificity (ap-proximately 37%–86%) for breast cancer detection (1–7) because certain benign lesionsexhibit static or dynamic enhancement characteristics similar to those seen in breastcancers (8). For example, fibroadenomas may enhance at variable rates from slow to fast(9–14), considerably overlapping the enhancement rates of malignant lesions such asductal carcinoma.

Attempts to increase the specificity of breast MR imaging by using computer-aideddiagnosis, or CAD, have been reported previously. The methods used have includedstatistical measurement based on semiautomated segmentation (15–17), analysis of con-trast material uptake and washout curves (5,18,19), normalized slope measurements ofcontrast material uptake (4), and neural network analysis (20) of contrast enhancementcurves. In addition, a model comprising contrast-enhanced three-dimensional imaging

Breast Imaging

225

Ra

dio

logy

and pharmacokinetic analysis was vali-dated in experimental breast tumors (21)and clinical subjects (14). These methodshave demonstrated specificities rangingfrom 39% to 83%.

The purpose of our study was to bothdevelop and use a tissue signature methodfor the identification and classification ofbreast lesions and healthy breast tissue. Wehypothesized that quantitative measure-ments of multiple MR signal intensities,when combined in a multiparametricmodel, would enable automated differen-tiation between different tissue types, suchas adipose, glandular, and lesion tissues inthe breast.

MATERIALS AND METHODS

ISODATA Model

The multiparametric MR image dataset that was analyzed with the iterativeself-organizing data (ISODATA) technique(22,23) consisted of T1-weighted images,fat-suppressed T2-weighted images, andthree-dimensional fat-suppressed T1-weighted images acquired before andduring contrast material enhancement(see MR Imaging) (Fig 1). These imagingsequences constitute the conventionalbreast MR imaging examination, andthey were selected for ISODATA analysisbecause each sequence provides differentcontrast to disclose different tissue types.For example, T1-weighted images providea clear visual separation of adipose tissuefrom glandular tissue, and T2-weighted fat-suppressed images facilitate the identifica-tion of fluid-filled structures such as cysts.Three-dimensional fat-suppressed T1-weighted images obtained prior to andduring contrast enhancement help iden-tify potential malignancies on the basis of

different enhancement caused by differentuptake of the contrast agent (9,24).

If we assume that each tissue type has acharacteristic signal intensity on each MRimage type, then each tissue type will forma cluster in a feature space the axes ofwhich represent the signal intensity of thattissue on MR images of that type. The cen-ter of each cluster can be described by atissue signature vector S�, calculated for theaverage signal intensity of all pixels thatare identified as part of the tissue type,such that S� � S1, S2, . . ., Sk, where Sk is themean signal intensity of the tissue type onthe kth image. The initial selection of nor-mal tissue signature vectors from the MRimages used in the ISODATA model isshown in Figure 2, A. The ISODATA tech-nique (22) is an unsupervised segmenta-tion method related to the K-means algo-rithm, with additional splitting andmerging steps that allow for the adjust-ment of cluster centers. Features of the ISO-DATA method include the ability to adjustthe number of clusters and the lack of aneed for initial training or for a prioriknowledge of the exact number of tissuetypes before segmentation. The modifiedISODATA algorithm that we used consistedof four main stages, summarized as follows(for a more complete description of the 14steps performed, see Jacobs et al [23]).

1. The criteria defining normal breasttissue parameters—in other words, adi-pose and glandular tissue signature vec-tors—and the initialization of the num-ber of clusters were input into theprogram. We first defined an adipose tis-sue signature vector that represented theadipose tissue depicted on the MR im-ages; a glandular tissue signature vectorwas likewise defined (Fig 2, A). The intra-set euclidean distance (IAD) and interset

euclidean distance (IED) � SD for eachtissue type were then computed by usingISODATA, information that was used forthe splitting and merging of the differentclusters. In this study, the number of ini-tial clusters was used to partition the fea-ture space into different segments withinwhich tissue cluster centers would be de-termined. By postulating a large numberof initial clusters, we ensured that all pos-sible combinations would be accountedfor. The expected number of clusters wasfive: adipose tissue, glandular tissue, ducts,and benign and malignant masses. Wetherefore set the number of initial clustersat 30—approximately an order of magni-tude greater than the expected number ofclusters or tissue types (ie, five).

2. The MR image feature space was par-titioned into random clusters.

3. Cluster centers were determined,and IAD and IED were calculated be-tween pixel vectors and cluster centers.IED was defined as the distance betweencluster centers and was calculated fromthe magnitude of the difference betweenthe two tissue cluster centers, such thatIEDij � �Sj � Si�, where Sj and Si are thecluster centers. IAD was defined as thevariance of each tissue cluster, such that

IADj �1Nj

�i�1

Nj

�Sji � Sj � ,

where Nj is the number of pixels in thejth cluster, Sj

i is the ith data point in thejth group, and Sj is the cluster center.

4. The clusters were then split andmerged on the basis of the projected IADand IED. The splitting and merging ofclusters allowed for adjustment to thenumber of clusters on the basis of thestructure of the data.

Figure 1. Representative sagittal MR image data set analyzed with the ISODATA algorithm. A, T1-weighted image acquiredwith a fast spoiled gradient-echo pulse sequence (200/4.4). B, Fat-suppressed T2-weighted image acquired with a spin-echopulse sequence (5,700/102). C, Three-dimensional fat-suppressed T1-weighted image acquired with a fast spoiled gradient-echo pulse sequence (20/4) prior to contrast enhancement by gadodiamide. D, Contrast-enhanced three-dimensionalT1-weighted image acquired with the same pulse sequence as in C.

226 � Radiology � October 2003 Jacobs et al

Ra

dio

logy

Steps 3 and 4 were repeated until thedata converged (ie, the SD of the clusterswas minimized) or the maximum num-ber of iterations was reached (Fig 2, B).

Study Subjects

A total of 36 patients (age range, 18–80years; median, 45 years) were included inthis study. The protocol was approved bythe Joint Committee on Clinical Investi-gation at the Johns Hopkins UniversitySchool of Medicine, and informed con-sent was obtained from all subjects.

Subjects were selected retrospectivelyfrom a cohort of 60 patients who were re-ferred consecutively over the course of 1year (from February 2000 to March 2001)for MR evaluation of suspected breast le-sions identified at mammography, ultra-sonography, and/or clinical examination.MR images of 11 patients could not beevaluated with the ISODATA method be-cause of technical difficulties such as im-proper field of view (n � 6), inadequate fat

suppression (n � 3), archival problems (n �1), and failure of image coregistration (n �1). From the remaining 49 patients whoseimages were technically adequate, the re-search coordinator (C.M.), who was notblinded to histologic findings, selected atrandom 18 patients who had received adiagnosis of benign breast lesion and 18patients who had received a diagnosis ofcarcinoma. This study design was based onstatistical considerations (ie, the need tohave equal numbers of patients in eachgroup). The cases were presented in ran-dom order to the MR imaging physicist(M.A.J.), who performed image analysiswhile blinded to histologic findings. AfterMR imaging, core or excisional biopsieswere performed in 20 patients, and theother 16 patients underwent lumpectomyor mastectomy.

Histologic Analysis of Lesions

After MR imaging, biopsies were per-formed on lesions identified on mammo-

grams and MR images and at physicalexamination. Tissue specimens from mas-tectomy were cooled to approximately5°C. Specimens were sliced at 5–10-mm in-tervals, in the same plane in which the MRimages were acquired.

Each lesion was classified as benign ormalignant on the basis of histologic anal-ysis. A lesion was considered benign ifthe histologic report specified fibrocys-tic changes, including formation of cysts orincreased fibrous tissue, with benign pat-terns of ductal or lobular distortion andbenign cellular changes such as are seen intypical or atypical lobular or ductal hyper-plasia. A lesion was considered malignantif it was diagnosed histologically as ductal,lobular, or undifferentiated carcinoma.

MR Imaging

For mammographically identified le-sions, the mammogram and mammo-graphic report were made available to theradiologist (D.A.B.) for correlation with

Figure 2. MR image data from the same patient as in Figure 1. A, Diagram of signature vectors defined for each normal tissuetype (adipose and glandular) and assigned to the tissue cluster that most closely resembles the vector elements for that tissuetype in the ISODATA algorithm. B, Representative ISODATA theme map of the signature vectors defined for each tissue clustershows normal adipose tissue (dark blue) in most of the breast. Regions that are light blue, green, and yellow representglandular tissue, and pink and white areas represent tumor tissue. C, Diagram of the three-dimensional feature space formed bythe combination of T1-weighted images and T2-weighted fat-suppressed images with contrast-enhanced three-dimensionalimages. The angular separation model, with the distribution of tissue clusters in the three-dimensional feature space, is shown.Angles were calculated as the dot product between the normal tissue cluster and the abnormal tissue cluster by using each cluster’stissue signature vector (�1, �2). Each axis represents the signal intensity distribution for each MR image type.

Volume 229 � Number 1 Diagnosis of Breast Lesions with MR Imaging � 227

Ra

dio

logy

MR images. For abnormalities identifiedat physical examination, a marker wasplaced on the breast prior to MR imaging.

MR imaging was performed with a 1.5-TMR imager (GE Medical Systems, Wauke-sha, Wis) and a dedicated phased-arraybreast coil (MRI Devices, Milwaukee, Wis),with the patient lying prone and the breastin the holder to reduce motion. The imag-ing protocol included a fat-suppressed T2-weighted spin-echo pulse sequence (rep-etition time msec/echo time msec, 5,700/102) and a T1-weighted fast spoiledgradient-echo sequence (200/4.4) with afield of view of 18 � 18 cm, matrix of256 � 192, section thickness of 4 mm,and 1-mm gap between sections. In addi-tion, a three-dimensional fat-suppressedT1-weighted fast spoiled gradient-echopulse sequence (20/4, image matrix of512 � 160, section thickness of 2 mm) wasperformed before and during contrast en-hancement obtained with intravenous ad-ministration of 0.2 mL/kg (0.1 mmol/kg)gadodiamide (Omniscan; AmershamHealth, Princeton, NJ). The contrast agentwas hand injected over a period of 10 sec-onds, followed by a 20-mL normal salinesolution flush, with MR imaging begin-ning immediately after completion of theinjection. Total imaging time for the entireprotocol was less than 20 minutes.

MR Image Preprocessingand Analysis

MR image preprocessing and analysiswere performed at a workstation (Ultra-

SPARC60; Sun Microsystems, MountainView, Calif). Images were preprocessedwith image analysis software (Eigentool;Image Analysis Lab, Henry Ford Hospital,Detroit, Mich) (25–28). Subimaging ofthe breast from the background wasachieved by using thresholding and mor-phologic operations (29) to reduce com-putational time. After subimaging, aninhomogeneity correction method wasapplied to the MR image data set (30).Finally, the images were restored by us-ing a nonlinear restoration filter to re-duce image noise while preserving edgesand partial volume effects (31). To ac-count for possible patient motion, mis-alignment between sequences, and dif-ferences in section location or thickness,coregistration of the breast MR imageswas accomplished by a method describedpreviously (30).

Quantitative MR Measurements

Tissue clusters were defined by usingmultiparametric ISODATA segmentationand identified according to tissue signa-ture vector (32). Tissue signature vectorswere defined for the adipose and glandu-lar tissues in each patient (Fig 2, A) andwere used as input values for the ISO-DATA algorithm. Other tissue types wereautomatically determined by the ISO-DATA routine, which output a thememap with different colors representingthe different tissue types (Fig 2, B). Adi-pose tissue depicted on the theme mapproduced by multiparametric ISODATA

segmentation was visually verified by aradiologist (D.A.B.) according to its ap-pearance on T1- and T2-weighted images.The adipose tissue vector defined by ISO-DATA segmentation was used as the ref-erence vector for measurement of angu-lar separation of the other tissue clustersdefined by ISODATA and depicted on thetheme map. Classification of the differ-ent tissue types was determined on thebasis of angular separation between thetwo tissue signature vectors (Fig 2, C),which was calculated by using the dotproduct between the vectors. The resultsobtained with the ISODATA model foreach lesion were compared with the his-topathologic findings.

To identify the combination of MRdata that would provide optimal visualseparation of breast tissue types in fea-ture space, ratios of IED to IAD were cal-culated between pixel vectors and clustercenters from the different MR image datasets. Whereas IED is the distance betweeneach tissue cluster center, and IAD is thevariance within each tissue cluster, theratio of IED to IAD is a measure of sepa-ration (33,34) between the different tis-sue clusters in feature space. The four MRimage data sets considered for optimalseparation were (a) T1- and T2-weightedimages; (b) T1- and T2-weighted imagesand fat-suppressed (nonenhanced) T2-weighted images; (c) T1- and T2-weightedimages and three-dimensional fat-sup-pressed contrast-enhanced T1-weighted

TABLE 1Patient Age, Lesion Size, and MR Imaging and Histologic Findings in Patients with Benign Lesions

Patient No.PatientAge (y)

LesionSize (cm)

MRFinding* Histologic Finding

1 63 1.2 Positive Benign fibrous breast tissue with microcalcifications2 41 4.8 Positive Benign breast tissue with fat necrosis3 45 1.6 Positive Apocrine metaplasia and cystic dilatation4 46 2.0 Positive Fibroadenoma5 74 2.0 Positive Residual atypical ductal hyperplasia, apocrine metaplasia, and intraductal papilloma6 44 4.0 Positive Fibrocystic change7 46 2.2 Positive Stromal fibrosis and elastosis. Florid ductal epithelial hyperplasia and sclerosing adenosis8 38 2.0 Positive Fibroadenoma9 48 1.8 Positive Fibroadenoma

10 46 3.6 Positive Fibrocystic change and adenosis11 53 1.0 Positive Benign breast tissue with fibrocystic change12 42 0.9 Positive Benign breast tissue with fibrocystic change, ductal hyperplasia, and adenosis with

microcalcifications13 45 0.5 Positive Benign breast tissue14 41 2.5 Positive Markedly atypical ductal hyperplasia with microcalcifications15 32 3.5 Positive Benign breast tissue with fibrocystic change16 44 1.0 Positive Benign breast tissue with fibrocystic change17 44 1.2 Positive Ductal ectasia and intense inflammatory infiltrate18 59 1.2 Positive Benign breast tissue with fibrocystic change and microcalcifications

Mean 47.3 2.1SD 9.7 1.2

* Positive indicates that contrast enhancement was observed in the lesion on MR images acquired after administration of gadodiamide.

228 � Radiology � October 2003 Jacobs et al

Ra

dio

logy

images; and (d) T1- and T2-weighted im-ages and three-dimensional fat-suppressedcontrast-enhanced and nonenhanced T1-weighted images.

Statistical Analysis

Differences in angular separation be-tween benign and malignant lesion groups

were evaluated by using an independentStudent t test. Statistically significant dif-ference was considered to be indicated byP � .05. In addition, the ability of theangular separation model to discriminatebetween benign lesions and malignantlesions was evaluated with nonpara-metric receiver operating characteristic(ROC) curve analysis.

RESULTS

The histopathologic and MR imagingfindings for the patient population inthis study are shown in Tables 1 and 2.On gadolinium-enhanced T1-weightedimages, all of the breast lesions were con-trast enhanced; contrast enhancementwas the criterion used to define a lesion.

The multiparametric ISODATA modelenabled successful segmentation of nor-mal and abnormal breast tissues and dif-ferentiation between adipose and glan-dular tissue types in all cases. Figure 3shows a representative multiparametricISODATA segmentation from a 56-year-old patient with infiltrating ductal carci-noma confirmed at histologic analysisafter mastectomy. The ISODATA segmen-tation of adipose and glandular tissues isshown in Figure 3, A. Representative his-tologic photomicrographs of tissue fromthe lesion area are shown in Figure 3,C and D. The angular separation of tumortissue from adipose tissue was 19.6°,whereas the separation of glandular tis-sue from adipose tissue was 7.1°. Overall,

TABLE 2Patient Age, Lesion Size, and MR Imaging and Histologic Findings in Patients with Malignant Lesions

Patient No.PatientAge (y)

TumorSize (cm)

MRFinding* Histologic Finding

1 46 1.0 Positive Infiltrating mammary carcinoma with one focus of ductal carcinoma2 52 1.3 Positive Infiltrating well-differentiated ductal carcinoma with atypical ductal hyperplasia3 56 1.1 Positive Infiltrating ductal carcinoma and benign fibroadipose tissue4 80 1.0 Positive Infiltrating mammary carcinoma with ductal and lobular features5 50 1.0 Positive In situ ductal carcinoma with microcalcifications6 54 2.5 Positive Invasive poorly differentiated mammary carcinoma with associated focal ductal carcinoma in situ7 47 1.0 Positive In situ ductal and infiltrating poorly differentiated ductal carcinoma with microcalcifications8 58 1.4 Positive Infiltrating lobular carcinoma, lobular carcinoma in situ, and foci of ductal carcinoma9 40 2.5 Positive Ductal carcinoma, solid and cribriform types

10 60 2.2 Positive Infiltrating ductal carcinoma11 62 2.0 Positive Infiltrating ductal carcinoma12 34 5.0 Positive Carcinoma13 62 6.0 Positive Differentiated ductal carcinoma14 70 3.6 Positive Infiltrating ductal carcinoma15 63 1.0 Positive Infiltrating ductal carcinoma with mucinous features16 52 1.3 Positive In situ and infiltrating mammary carcinoma17 34 3.0 Positive In situ carcinoma18 59 2.6 Positive Infiltrating ductal carcinoma

Mean 54.4 2.2SD 11.8 1.5

* Positive indicates that contrast enhancement was observed in the lesion on MR images acquired after administration of gadodiamide.

Figure 3. MR image data from a 56-year-old woman with infiltrating ductal carcinomaconfirmed by histologic analysis following mastectomy. A, Theme map obtained with mul-tiparametric ISODATA segmentation at an angular separation threshold of 19° demonstratesclear delineation of adipose tissue (blue) from glandular tissue (light green to yellow). B,Sagittal T1-weighted contrast-enhanced digital subtraction image acquired with a FSPGRpulse sequence (20/4). C, D, Histologic photomicrographs (hematoxylin-eosin stain, originalmagnification in C, �2; in D, �40) on which the tumor tissue appears in pink and white. Thehistologic morphology of the lesion was consistent with that of infiltrating ductal carcinoma.

Volume 229 � Number 1 Diagnosis of Breast Lesions with MR Imaging � 229

Ra

dio

logy

for malignant lesions (n � 18), the aver-age angular separation between adiposetissue and tumor tissue was 17.8° � 6.1°(mean � SD). For benign lesions (n � 18),the average angular separation betweenadipose tissue and lesion tissue was29.0° � 11.2°.

Angular separation of lesion tissuefrom adipose tissue was significantly dif-ferent between benign and malignant le-sions (P � .002, df � 26, t statistic ��3.64). Angular separation from adiposetissue also was significantly different be-tween glandular tissue and lesion tissue(P � .001, df � 22, t statistic � �7.41). Inaddition, there was a significant differ-ence in glandular tissue signature vectorsbetween patients with benign lesions andthose with malignant lesions (P � .04,df � 30, t statistic � 2.04) (Table 3). Fig-ure 4 shows the ROC curve used to eval-uate the ability of the ISODATA-basedangular separation model to differentiatebenign lesions from malignant lesions.The area under the ROC curve is 0.84.The sensitivity and specificity of severalrepresentative angular separation thresh-olds are presented in Table 4.

The MR data set that included T1-weighted and T2-weighted images andcontrast-enhanced and nonenhancedthree-dimensional T1-weighted imageshad the largest IED/IAD ratio—2.54 forthe lesion tissue cluster and 1.51 for theglandular tissue cluster (Table 5). This re-sult indicates that this image data set pro-vided the best separation in featurespace. The large IED/IAD ratio also sug-gests that this data set may be optimal forachieving increased separation betweentissue types and decreased variancewithin each tissue type. Two data sets(T1- and T2-weighted images combinedwith contrast-enhanced three-dimen-sional T1-weighted images, and T1- andT2-weighted images combined with non-enhanced three-dimensional T1-weightedimages) had similar IED/IAD ratios for thelesion tissue cluster (1.85 and 1.80, respec-

tively) and glandular tissue cluster (1.31and 1.32, respectively). The T1- and T2-weighted image data set exhibited the low-est IED/IAD ratio for the lesion tissue (ie,1.68) and glandular tissue types (ie, 1.19).

DISCUSSION

The results of this study indicate that aquantitative, multiparametric approachto breast MR imaging may enable theautomated characterization of breast le-sions as benign or malignant. The appli-cation of the angular separation model toMR image analysis provides quantitativeinformation that may assist the radiolo-gist in determining whether a lesion isbenign or malignant. Further evaluationwill be necessary to determine whetherthis quantitative assessment improvesthe accuracy of clinical diagnosis ofbreast cancer.

Other investigators previously reportedthat T1- and T2-weighted images are notsufficient for distinguishing benign frommalignant tissue (1,10,35) and that con-trast-enhanced MR image data allow forbetter visualization of the tumor extentand of multifocality in breast masses sus-pected to be malignant (4,9). The results ofour study indicate that quantitative analy-sis of all MR data is useful for tissue classi-fication as benign versus malignant. Al-though T1- and T2-weighted images alonemay not be useful for differentiating tissuetypes, the combination of these imageswith contrast-enhanced three-dimensionalT1-weighted images increases the distancebetween tissue clusters and minimizes thespread of values within a single cluster.This assertion is supported by the IED/IADratios calculated for the different imagedata sets; the multiparametric data set con-sistently provided a higher IED/IAD ratiothan the other data sets.

T1-weighted images can provide excel-lent visual separation of adipose tissuefrom glandular tissue, and T2-weighted

images can depict cysts, necrosis, hemor-rhage, and some fibroadenomas (1,10,35).Previous investigators, such as Kuhl et al(35), suggested that the incorporation ofT2 weighting and contrast material en-hancement into the breast MR imagingprotocol could increase diagnostic specific-ity. Most malignant breast lesions are pro-foundly enhanced after administration of acontrast agent, and the enhancement pat-tern may aid in the differential diagnosis ofthe lesion (5,18). However, some benignlesions, such as fibroadenomas, tend to en-hance to the same extent as carcinomas.To overcome this difficulty, Degani et al(21) used the contrast enhancement pat-terns on MR images obtained at three dif-ferent time points during contrast materialuptake to produce a color-coded thememap that characterized tumor heterogene-ity in terms of microvascular permeabilityand extracellular fraction. The use of thisthree-time-point imaging and mappingmethod was shown to improve the accu-

TABLE 3Angular Separations of Glandular Tissue and of Lesion Tissue from the AdiposeTissue Reference, Calculated for Malignant and Benign Lesions

Lesion

Angular Separation ofGlandular Tissue from

Adipose Tissue (degrees)

Angular Separation ofLesion Tissue from

Adipose Tissue (degrees)

Mean SD Mean SD

Malignant (n � 18) 5.6* 2.4 17.8† 6.1Benign (n � 18) 7.8 3.5 29.0 11.2

* P � .04 indicates a significant difference between malignant and benign lesions.† P � .002 indicates a significant difference between malignant and benign lesions.

TABLE 4Sensitivity and Specificity of theModel at Six Threshold Angle Valuesfor Prediction of Benign versusMalignant Lesions

Threshold Angle(degrees)

Sensitivity(%)

Specificity(%)

17 33 9418 50 8919 67 8921 89 7822 94 7223 94 67

Figure 4. Graph of ROC curves for ISODATA-based analysis of angular separation accordingto lesion tissue type and for all tissues. The areaunder the curve for lesion tissue type was 0.84.The diagonal line indicates an area under thecurve of 0.50 (ie, no separation between tissuetypes).

230 � Radiology � October 2003 Jacobs et al

Ra

dio

logy

racy of diagnosis of fibroadenoma (14).This area of breast imaging has been re-viewed recently (36).

In our study, malignant lesions hadsmaller angular separations from adiposetissue than did benign lesions, but themechanisms underlying these differencesin angular separation have yet to be de-termined. MR imaging is highly sensitiveboth to localized changes in tissue watercontent and to changes in the generalenvironment produced by the exchangeof water with the surrounding tissue.Changes in tissue water content can re-flect physiologic and morphologic alter-ations that occur during the growth oftumors. The use of extracellular gadolin-ium-based contrast agents improves thedetection of most malignant breast le-sions by means of observed lesion vascu-larity, architecture, and permeability.The timing of the contrast material injec-tion and of breast imaging influences thesignal intensity on MR images.

Malignant breast lesions tend to ex-hibit a washout pattern on dynamic con-trast-enhanced images that differs fromthat of benign lesions, which usually ex-hibit either persistent or plateau-type en-hancement (5). However, benign lesionsand malignant lesions can have the samecontrast enhancement pattern (5,14).The time delay between contrast materialinjection and image acquisition may in-fluence the angular separation betweenthe lesion tissue and adipose tissue, re-sulting in the smaller angle seen in ma-lignant lesions; this hypothesis requiresfurther evaluation in future studies.

Of the four MR image data sets used inour ISODATA segmentation model, themultiparametric set that included all fourtypes of MR images consistently pro-duced better tissue cluster separation asdefined by the IED/IAD ratio than theother MR data sets. In the sets lackingcontrast-enhanced image data, the lesion

was not optimally segmented from thesurrounding tissues. The T1- and T2-weighted image data set had the lowestIED/IAD ratio. Our study results indicatethat the inclusion of contrast-enhancedthree-dimensional T1-weighted imagesallowed for better segmentation of thebreast tissue.

Breast tumors, whether benign or ma-lignant, tend to have longer T1 and T2relaxation times than does normal breasttissue (37,38). Malignant tumors haveshorter T1 and T2 relaxation times thando benign lesions, although there may besome overlap in relaxation times be-tween the two lesion groups (37,38). Thisdifference in relaxation times may be re-lated to increased water content in be-nign lesions and may have produced theincreased angular separation observed inthis study.

The results of previously publishedstudies indicate that MR imaging hashigh sensitivity but lower specificity forthe detection of breast lesions. In most ofthese studies, time-intensity curves wereused to measure the dynamic enhance-ment of signal intensity over time, afterinjection of a contrast agent. Some inves-tigators, using time-intensity curves alone,have demonstrated a sensitivity of ap-proximately 91% for lesion detectionwith MR imaging, with specificities rang-ing from 37% to 86% (5,39–43).

Adams et al (15) evaluated the com-bined use of T1- and T2-weighted imageswith intermediate-weighted and Dixon-opposed breast MR images and super-vised segmentation classifiers in a three-dimensional feature space. Supervisedsegmentation requires a training data setto teach classifiers how to recognize dif-ferent tissue types. In addition, Adams etal evaluated shape analysis, in whichcompactness and frequency of lesionboundary characteristics are considered.They found shape analysis important for

separating benign from malignant tumorson the basis of T1-weighted, T2-weighted,and Dixon-opposed MR images. Lucas-Quesada et al (16) constructed a two-di-mensional feature space from contrast-enhanced and nonenhanced MR imagesand employed two semiautomated meth-ods for segmentation of breast tumors.The two methods gave similar segmenta-tion results, with accuracy of 76%–84%(according to the authors’ criterion);however, the segmented lesions were notclassified as benign or malignant.

Gilhuijs et al (17) reported the use of amodel for computer-aided diagnosisbased on the extraction of different fea-tures from dynamic contrast-enhancedimages, including the uptake of contrastmaterial, sharpness of lesion definition,shape of the lesion, and radial gradient.Features were evaluated individually andin combination by means of ROC analy-sis. The combination of radial gradientwith sharpness gave the best results (ROCarea, 0.86–0.87 for two-dimensionalanalysis and 0.92–0.96 for three-dimen-sional analysis), with specificity of 77%and sensitivity of 100%. Lucht et al (20),using a neural network classifier andtime-intensity curves derived from breastMR images, achieved a sensitivity of 84%and specificity of 81%.

To our knowledge, our study is the firstevaluation of a four-dimensional mul-tiparametric MR image data set consist-ing of T1- and T2-weighted images andcontrast-enhanced and nonenhancedthree-dimensional T1-weighted images.The results achieved with our classifica-tion method are comparable to those ofother, similar studies. For example, withan angular separation of 21°, sensitivityof 89% and specificity of 78% wereachieved. At larger angles, sensitivity of94% and specificity of 72% were realized.These data demonstrate the potential ofthe angular separation model for cor-rectly identifying and classifying breastlesions. The advantage of the model de-scribed in this article is that it may beadapted to include five dimensions (eg,dynamic contrast enhancement) or more(eg, shape analysis [44], diffusion coeffi-cients, or sodium imaging [45]). The goalof subsequent multiparametric analysiswill be to determine the minimum num-ber of parameters that objectively andquantitatively yield optimal sensitivityand specificity.

In conclusion, multispectral analysis ofconventional MR images by means ofISODATA and measurements of angularseparation may provide an automatedmeans of breast lesion identification and

TABLE 5Comparison of IED/IAD Ratios among the Four Image Data Sets Studied,for Glandular and Lesion Tissue Clusters

MR Image Data Set

IED/IAD Ratio forTissue Cluster

GlandularTissue

LesionTissue

T1- and T2-weighted images 1.19 1.68T1- and T2-weighted images and fat-suppressed (nonenhanced)

T2-weighted images 1.32 1.80T1- and T2-weighted images and three-dimensional fat-suppressed

contrast-enhanced T1-weighted images 1.31 1.85T1- and T2-weighted images and three-dimensional fat-suppressed

contrast-enhanced and nonenhanced T1-weighted images 1.51 2.54

Volume 229 � Number 1 Diagnosis of Breast Lesions with MR Imaging � 231

Ra

dio

logy

classification. This approach may enablethe identification of specific tissue signa-tures characteristic of benign versus ma-lignant breast lesions.

Acknowledgments: We gratefully acknowl-edge Donald Peck, PhD, Physics and Engineer-ing Division head, Hamid Soltanian-Zadeh,PhD, senior staff scientist, and Lucie Bower, BS,systems analyst at the Image Analysis Lab,Henry Ford Health System, Detroit, Mich, forproviding the Eigentool image analysis soft-ware used in image processing, and the anon-ymous reviewers of the manuscript for theirconstructive comments.

References1. Heywang SH, Bassermann R, Fenzl G, et

al. MRI of the breast: histopathologic cor-relation. Eur J Radiol 1987; 7:175–182.

2. Gilles R, Guinebretiere JM, Shapeero LG,et al. Assessment of breast cancer recur-rence with contrast-enhanced subtrac-tion MR imaging: preliminary results in26 patients. Radiology 1993; 188:473–478.

3. Fischer U, Kopka L, Brinck U, Korabiow-ska M, Schauer A, Grabbe E. Prognosticvalue of contrast-enhanced MR mam-mography in patients with breast cancer.Eur Radiol 1997; 7:1002–1005.

4. Kelcz F, Santyr GE, Cron GO, MonginSJ. Application of a quantitative model todifferentiate benign from malignantbreast lesions detected by dynamic, gad-olinium-enhanced MRI. J Magn ResonImaging 1996; 6:743–752.

5. Kuhl CK, Mielcareck P, Klaschik S, et al.Dynamic breast MR imaging: are signalintensity time course data useful for dif-ferential diagnosis of enhancing lesions?Radiology 1999; 211:101–110.

6. Orel SG. High-resolution MR imaging forthe detection, diagnosis, and staging ofbreast cancer. RadioGraphics 1998; 18:903–912.

7. Kuhl CK. MRI of breast tumors. Eur Radiol2000; 10:46–58.

8. Weinreb JC, Newstead G. MR imaging ofthe breast. Radiology 1995; 196:593–610.

9. Heywang SH, Hahn D, Schmidt H, et al.MR imaging of the breast using gadolin-ium-DTPA. J Comput Assist Tomogr1986; 10:199–204.

10. Kerslake RW, Carleton PJ, Fox JN, et al.Dynamic gradient-echo and fat-sup-pressed spin-echo contrast-enhancedMRI of the breast. Clin Radiol 1995; 50:440–454.

11. Heywang-Kobrunner SH, Haustein J, PohlC, et al. Contrast-enhanced MR imagingof the breast: comparison of two differentdoses of gadopentetate dimeglumine. Ra-diology 1994; 191:639–646.

12. Gribbestad IS, Nilsen G, Fjosne HE,Kvinnsland S, Haugen OA, Rinck PA.Comparative signal intensity measure-ments in dynamic gadolinium-enhancedMR mammography. J Magn Reson Imag-ing 1994; 4:477–480.

13. Orel SG, Schnall MD, LiVolsi VA, TroupinRH. Suspicious breast lesions: MR imag-ing with radiologic-pathologic correla-tion. Radiology 1994; 190:485–493.

14. Weinstein D, Strano S, Cohen P, FieldsS, Gomori JM, Degani H. Breast fibro-adenoma: mapping of pathophysiologicfeatures with three-time-point, contrast-

enhanced MR imaging—pilot study. Ra-diology 1999; 210:233–240.

15. Adams AH, Brookeman JR, Merickel MB.Breast lesion discrimination using statis-tical analysis and shape measures onmagnetic resonance imagery. ComputMed Imaging Graph 1991; 15:339–349.

16. Lucas-Quesada FA, Sinha U, Sinha S. Seg-mentation strategies for breast tumorsfrom dynamic MR images. J Magn ResonImaging 1996; 6:753–763.

17. Gilhuijs KGA, Giger ML, Bick U. Comput-erized analysis of breast lesions in threedimensions using dynamic magnetic-resonance imaging. Med Phys 1998; 25:1647–1654.

18. Orel SG. Differentiating benign from ma-lignant enhancing lesions identified atMR imaging of the breast: are time-signalintensity curves an accurate predictor?Radiology 1999; 211:5–7.

19. Sinha S, Lucas-Quesada FA, DeBruhl ND,et al. Multifeature analysis of Gd-en-hanced MR images of breast lesions. JMagn Reson Imaging 1997; 7:1016–1026.

20. Lucht RE, Knopp MV, Brix G. Classifica-tion of signal-time curves from dynamicMR mammography by neural networks.Magn Reson Imaging 2001; 19:51–57.

21. Degani H, Gusis V, Weinstein D, Fields S,Strano S. Mapping pathophysiological fea-tures of breast tumors by MRI at high spa-tial resolution. Nat Med 1997; 3:780–782.

22. Ball G, Hall D. ISODATA, a novel methodof data analysis and pattern classification.Technical Report no. 4. Menlo Park, Calif:Stanford Research Institute, April 1965.

23. Jacobs MA, Knight RA, Soltanian-ZadehH, et al. Unsupervised segmentation ofmultiparameter MRI in experimental ce-rebral ischemia with comparison to T2,diffusion, and ADC MRI parameters andhistopathological validation. J Magn Re-son Imaging 2000; 11:425–437.

24. Kacl GM, Liu P, Debatin JF, Garzoli E,Caduff RF, Krestin GP. Detection of breastcancer with conventional mammogra-phy and contrast-enhanced MR imaging.Eur Radiol 1998; 8:194–200.

25. Jacobs MA, Knight RA, Windham JP, et al.Identification of cerebral ischemic lesionsin rat using eigenimage filtered magneticresonance imaging. Brain Res 1999; 837:83–94.

26. Peck D, Windham J, Emery L, Soltanian-Zadeh H, Hearshen D, Mikkelsen T. Cere-bral tumor volume calculations usingplanimetric and eigenimage analysis.Med Phys 1996; 23:2035–2042.

27. Soltanian-Zadeh H, Windham J, Peck D,Yagle A. A comparative analysis of severaltransformations for enhancement andsegmentation of magnetic resonance im-age scene sequence. IEEE Trans Med Im-aging 1992; 11:302–318.

28. Windham JP, Abd-Allah MA, ReimannDA, Froelich JW, Haggar AM. Eigenimagefiltering in MR imaging. J Comput AssistTomogr 1988; 12:1–9.

29. Soltanian-Zadeh H, Windham JP. A mul-tiresolution approach for contour extrac-tion from brain images. Med Phys 1997;24:1844–1853.

30. Jacobs MA, Windham J, Soltanian-ZadehH, Peck D, Knight R. Registration andwarping of magnetic resonance images tohistological sections. Med Phys 1999; 26:1568–1578.

31. Soltanian-Zadeh H, Windham JP, YagleAE. A multidimensional nonlinear edge-preserving filter for magnetic resonanceimage restoration. IEEE Trans Image Pro-cessing 1995; 4:147–161.

32. Jacobs MA, Bluemke DB, Bhujwalla Z,Marano C, Barker PB. A model to differ-entiate benign from malignant breast tu-mors using multiparametric MRI withhistological correlates (abstr). In: Pro-ceedings of the ninth meeting of the In-ternational Society for Magnetic Reso-nance in Medicine. Berkeley, Calif:International Society for Magnetic Reso-nance in Medicine, 2001; 563.

33. Tou J, Gonzales R. Pattern recognitionprinciples. Reading, Mass: Addison-Wes-ley, 1974.

34. Theodoridis S, Koutroumbas K. Patternrecognition. Vol 1. San Diego, Calif: Aca-demic Press, 1999.

35. Kuhl CK, Klaschik S, Mielcarek P, GiesekeJ, Wardelmann E, Schild HH. Do T2-weighted pulse sequences help with thedifferential diagnosis of enhancing le-sions in dynamic breast MRI? J Magn Re-son Imaging 1999; 9:187–196.

36. Orel SG, Schnall MD. MR imaging of thebreast for the detection, diagnosis, andstaging of breast cancer. Radiology 2001;220:13–30.

37. Merchant TE, Thelissen GR, de Graaf PW,Nieuwenhuizen CW, Kievit HC, Den Ot-ter W. Application of a mixed imagingsequence for MR imaging characteriza-tion of human breast disease. Acta Radiol1993; 34:356–361.

38. McSweeney MB, Small WC, Cerny V,Sewell W, Powell RW, Goldstein JH. Mag-netic resonance imaging in the diagnosis ofbreast disease: use of transverse relaxationtimes. Radiology 1984; 153:741–744.

39. Boetes C, Barentsz JO, Mus RD, et al. MRcharacterization of suspicious breast le-sions with a gadolinium-enhanced Tur-boFLASH subtraction technique. Radiol-ogy 1994; 193:777–781.

40. Gilles R, Guinebretiere JM, Lucidarme O,et al. Nonpalpable breast tumors: diagno-sis with contrast-enhanced subtractiondynamic MR imaging. Radiology 1994;191:625–631.

41. Harms SE, Flamig DP, Hesley KL, et al. MRimaging of the breast with rotating deliv-ery of excitation off-resonance: clinicalexperience with pathological correlation.Radiology 1993; 187:493–501.

42. Heiberg EV, Perman WH, Herrmann VM,Janney CG. Dynamic sequential 3D gad-olinium-enhanced MRI of the wholebreast. Magn Reson Imaging 1996; 14:337–348.

43. Daniel BL, Yen YF, Glover GH, et al.Breast disease: dynamic spiral MR imag-ing. Radiology 1998; 209:499–509.

44. Nunes LW, Schnall MD, Orel SG. Updateof breast MR imaging architectural inter-pretation model. Radiology 2001; 219:484–494.

45. Ouwerkerk R, Jacobs MA, Bottomley PA,Fajardo LL. A method for quantifying tis-sue sodium in breast tumors with shortecho time 23NA MRI (abstr). In: Proceed-ings of the tenth meeting of the Interna-tional Society for Magnetic Resonance inMedicine. Berkeley, Calif: InternationalSociety for Magnetic Resonance in Medi-cine, 2002; 2060.

232 � Radiology � October 2003 Jacobs et al

Ra

dio

logy