Sonographic Appearance of Invasive Ductal Carcinoma of the Breast According to Histologic Grade

Transcription

1 Women s Imaging Original Research Blaichman et al. Efficacy of Sonographic Features in BI-RADS Lexicon Women s Imaging Original Research Jason Blaichman 1 James C. Marcus Tahra Alsaadi Mona El-Khoury Sarkis Meterissian Benoít Mesurolle Blaichman J, Marcus JC, Alsaadi T, El-Khoury M, Meterissian S, Mesurolle B Keywords: breast ultrasound, BI-RADS lexicon, grade, invasive ductal carcinoma DOI: /AJR Received June 6, 2011; accepted after revision January 3, J. Blaichman and J. C. Marcus contributed equally to this work. 1 All authors: Division of Radiology, Cedar Breast Clinic, McGill University Health Center, Royal Victoria Hospital, 687 Pine Ave W, Montreal, QC, H3H 1A1, Canada. Address correspondence to B. Mesurolle (benoit.mesurolle@muhc.mcgill.ca). WEB This is a Web exclusive article. AJR 2012; 199:W402 W X/12/1993 W402 American Roentgen Ray Society Sonographic Appearance of Invasive Ductal Carcinoma of the Breast According to Histologic Grade OBJECTIVE. The purpose of this study was to compare the efficacy of the sonographic features in the BI-RADS lexicon for predicting malignancy grade of invasive ductal breast carcinoma in women assigned a BI-RADS category of 4 or 5. MATERIALS AND METHODS. Two radiologists retrospectively evaluated 299 consecutive cases of grades 1 3 invasive ductal breast carcinoma presenting as a mass in consensus by using the BI-RADS sonographic lexicon. Histologic grade was established on surgical specimens. Effect sizes were calculated via the Goodman and Kruskal tau, an asymmetric measure of strength of nominal association, and results were interpreted in terms of proportional reduction in error. RESULTS. Thirty-eight lesions (13%) were grade 1, 153 (51%) were grade 2, and 108 (36%) were grade 3, with the majority of all masses showing an irregular shape (84%) and hypoechoic echotexture (82%). Of the sonographic features examined, malignancy grade was best predicted by posterior acoustics (τ = 0.13, p < 0.001), lesion boundary (τ = 0.05, p < 0.001), and margin (τ = 0.04, p = 0.001). Among grade 3 lesions, there were significantly more lesions with posterior enhancement (53 vs 27.6; adjusted standardized residuals (z res ) = 7; p < 0.001), abrupt interfaces (68 vs 51.2; z res = 4; p < 0.001), and microlobulated margins (12 vs 5.8; z res = 3; p = 0.001) than would be expected. CONCLUSION. Malignancy grade was slightly to moderately predicted by margin, lesion boundary, and acoustic sonographic features. In particular, grade 3 invasive ductal breast carcinomas were more likely than expected to display microlobulated margins, abrupt interfaces, and posterior enhancement. T he use of breast ultrasound has shown remarkable promise in distinguishing benign from malignant solid lesions [1]. Recognizing this, several groups have proposed classification schemes, such as that of Stavros et al. [2]. More recently introduced, the sonographic BI-RADS lexicon has shown its usefulness in differentiating benign from malignant solid masses [3]. Breast carcinomas of varying histologic grades have been shown to display different morphologic features; however, only two previous studies [4, 5] have actually focused on comparing the sonographic appearance of invasive ductal breast carcinomas according to grade. In a retrospective study involving 120 diagnosed invasive ductal breast carcinoma patients, Lamb et al. [4] investigated the relationship between imaging characteristics and histologic grade. Contrary to expectations that more malignant masses would show spic- ulated mass margins and acoustic shadowing, they found that higher-grade tumors were significantly more likely than lower-grade ones to display poorly defined margins and posterior acoustic enhancement. Rotstein and Neerhut [5] carried out a retrospective analysis of 181 diagnosed grade 3 invasive ductal breast carcinomas and similarly found that only 30% of these lesions displayed the classic feature of acoustic shadowing. Although of significant importance, these studies did not use the BI-RADS classification to categorize sonographic features, making comparison of their findings with future research studies difficult. Furthermore, Lamb et al. made no attempt to quantify the strength of their effect sizes and Rotstein and Neerhut examined grade 3 lesions in isolation. To better understand the relationship between sonographic features and malignancy grade, we carried out a retrospective study of invasive breast carcinomas using the BI-RADS W402 AJR:199, September 2012

2 Efficacy of Sonographic Features in BI-RADS Lexicon lexicon. In particular, our aim was to determine whether applying the descriptors used by the BI-RADS lexicon would be effective in helping us distinguish grade 3 carcinomas from other grades in women assigned a BI- RADS category of 4 or 5. To accomplish this, we retrospectively described the sonographic features of grades 1, 2, and 3 breast carcinomas according to the BI-RADS lexicon; compared these features to assess any significant differences between the various grades; and assessed the efficacy of BI-RADS categorizations 4 and 5 in predicting the histologic grade. Materials and Methods Permission Because this was a retrospective analysis, approval by the institutional review board was not required. Permission was obtained from the hospital for review of the patients medical records. Patients Data were reviewed retrospectively for invasive ductal breast carcinoma in 274 consecutive women at our institution between April 2002 and February All patients had an ultrasound-guided core needle biopsy performed and subsequently underwent surgical excision of the breast mass. Twenty-two patients had two tumors, two patients had three tumors, and the rest of the patients had a single tumor. Sonographic Review Breast examination was performed using either of two high-resolution scanners with highfrequency linear-array MHz transducers (Sequoia scanner with 15L8w broadband transducer, Siemens Healthcare; and Aplio scanner, Toshiba with high-frequency matrix transducer PLT1204AX, Acuson). One of the authors, who did not read the images, generated a set of two orthogonal views of each mass and number-coded all pairs of images to maintain confidentiality. Images of these 300 breast masses were then randomly displayed on a PACS workstation (IntelePACS, version 3.7.1, Intelerad Medical Systems) and the sonographic features were evaluated according to the terminology of the fourth edition of the BI-RADS lexicon in a joint consensus reading by two of the authors, with 4 and 12 years of breast imaging experience. In addition, each image was assigned a final BI-RADS category [3]. No attempt was made to gauge the initial level of agreement for either the sonographic feature or BI-RADS final assessment categories. Although blinded to the grade of the lesions, the radiologists were aware that they were evaluating malignant masses and limited themselves to TABLE 1: Distribution of Grade by Feature Category Feature assigning cases to BI-RADS categories 4a c (suspicious abnormality, three tiers) or 5 (highly suggestive of malignancy). One of the 300 images was assigned a BI-RADS category of 3 (probably benign) and excluded from further analyses. Evaluation of the masses was limited to sonographic features of shape (irregular, round, or oval), orientation (parallel or not parallel), margin (circumscribed, microlobulated, spiculated, angular, or indistinct), lesion boundary (abrupt interface or echogenic halo), echo pattern (hypoechoic, isoechoic, hyperechoic, complex, or anechoic), and posterior acoustic features (enhancement, none, shadowing, or combined pattern). Features such as alterations in the surrounding tissue and presence of calcifications were not included in the evaluation. For each feature category, the image interpreters were limited to selecting only one best descriptor from the BI-RADS lexicon; if the lesion could be described by more than one descriptor, the interpreters were instructed to choose the one most suspicious for malignancy. Frequency Grade 1 Grade 2 Grade 3 Total Overall Mass shape Oval Round Irregular Mass margins Circumscribed Microlobulated Indistinct Angular Spiculated Echo pattern Hyperechoic Isoechoic Hypoechoic Complex Anechoic Mass orientation Parallel Not parallel Acoustic Enhanced None Shadow Combination Lesion boundary Abrupt interface Echogenic halo BI-RADS 4a b c AJR:199, September 2012 W403

3 Blaichman et al. Pathology Histologic grade was collected through retrospective review of charts. Tumor grade was established on surgical excision. Histologic grading of invasive ductal carcinoma was based on the Elston and Ellis [6] modification of the Bloom and Richardson system. Fig. 1 Invasive ductal carcinoma grade 1 in 55-year-old woman. Breast sonogram shows 1.1-cm mass displaying irregular shape, angular margins, nonparallel orientation, echogenic halo, hypoechoic echotexture, and posterior acoustic shadowing. This mass was categorized as BI-RADS 5. Statistical Analysis We predicted the malignancy grade multivariately via an unconditional ordinal logistic regression model with sonographic features as the predictors and patient age (mean-centered) and lesion size as the covariates (Appendix 1). All subsequent analyses of predictors were done univariately and nonparametrically. The Goodman and Kruskal tau test, an asymmetric chi-square based measure of strength of nominal association, was used to evaluate the degree to which various sonographic features could predict the breast carcinoma histologic grade (Appendix 2). Tau is bounded from 0 to 1, with a value of 0 meaning that the feature category is of no help in predicting grade and a value of 1 implying that the feature perfectly predicts the grade. The efficacy of the final BI-RADS categorization at predicting histologic grade was assessed via Somers d test, an asymmetric but directional measure of ordinal association. Somers d is bounded from 1 to 1, with a value of 0 indicating that the raters BI-RADS category assignment was of no help in predicting grade and a value of 1 or 1 indicating that the categorization perfectly predicted grade (the sign indicates the direction of the association). For an explanation of why the measures of association tau and d were chosen in lieu of the more commonly encountered Pearson correlation, please refer to Appendix 3. Given the multiple features evaluated (eight, including final BI-RADS categorization), familywise alpha was controlled (capped at α = 0.05) by the Holm-Bonferroni method (Appendix 4). Statistically significant features (at α HB < 0.05) were further analyzed via adjusted standardized residuals (z res ) to highlight which departures (Table 1) from expected cell frequencies were responsible for the observed univariate effects. Results Clinical, Sonographic, and Pathologic Findings Our study population consisted of 274 women (mean age, 60.9 [SD] 12.8 years; age range, years; median, 61 years). The masses observed on ultrasound had a mean size of 1.5 cm with an SD of 0.9 cm. The distribution of tumor grade for the study population was 13% (38/299), 51% (153/299), and 36% (108/299) for grades 1, 2, and 3, respectively. Surgical Grade and Sonographic Features Multivariate analyses We predicted surgical grade via an ordinal logistic regression model with sonographic features as the predictors Fig. 2 Invasive ductal carcinoma grade 1 in 73-year-old woman. Breast sonogram shows 0.4-cm mass, displaying round shape, microlobulated margins (arrow), nonparallel orientation, abrupt interface, hypoechoic echotexture, and neutral posterior acoustic features. This mass was categorized as BI-RADS 4a. Fig. 3 Invasive ductal carcinoma grade 2 in 65-year-old woman. Breast sonogram shows 1.1-cm mass (arrows) displaying irregular shape, indistinct margins, nonparallel orientation, echogenic halo, hypoechoic echotexture, and posterior acoustic shadowing. This mass was categorized as BI-RADS 5. Fig. 4 Invasive ductal carcinoma grade 2 in 78-year-old woman. Breast sonogram shows 1.2-cm mass displaying irregular shape, spiculated margins (arrow), nonparallel orientation, abrupt interface, hypoechoic echotexture, and posterior acoustic shadowing. This mass was categorized as BI-RADS 5. W404 AJR:199, September 2012

4 Efficacy of Sonographic Features in BI-RADS Lexicon and patient age (mean-centered) and lesion size as the covariates (Figs. 1 5). For each feature, the modal category was assigned as the reference group. No attempt was made to look at interactions among predictors. Categories with expected cell frequencies less than 1 were discarded (one case in which the margins were circumscribed and two cases in which the echo pattern was hyperechoic), leaving 296 cases. After this, there were no features for which more than 20% of cells in the two-way contingency tables had expected frequencies of less than 5. The proportional odds assumption was satisfied (χ 2 14 = 15.93, p = 0.32); therefore, a nominal model was not explored. The full model differed significantly from the constant-only model (χ 2 14 = 104.5, p < ), indicating that the features, as a set, reliably distinguished surgical grades. The variance accounted for was moderate to large, with Somers d = 0.59, as was the probability of correctly predicting the surgical grade classification, with c = 0.8. Ignoring the cumulative structure of the predictions (i.e., using the modal prediction), the model had a positive predictive value (PPV) of 63 / 87 = 0.72 for grade 3, 125 / 205 = 0.61 for grade 2, and 1 / 4 = 0.25 for grade 1, for an overall success rate of 189 / 296 = These values compare very favorably to the PPVs of using biopsy grade to predict surgical grade in our sample: 55 / 64 = 0.86 for grade 3, 116 / 180 = 0.66 for grade 2, and 23 / 42 = 0.55 for grade 1, with an overall success rate of 194 / 286 = 0.68 (14 cases had only a surgical grade). Within the model, the covariate lesion size was a significant predictor of grade (χ 2 1 = 13.01, p = ). With an odds ratio of 2.15 (95% CI, ), the odds of a mass being judged grade 3 instead of grade 1 or 2 or judged grade 2 or 3 instead of 1 doubled for each unit increase in lesion size. Among the sonographic features, the only significant predictor of grade was posterior acoustics (χ 2 3 = 28.15, p < ). Compared with shadowing, the odds of appearing in a higher grade category were 2.95 ( ) for a combined pattern, ( ) for enhancement, and 2.71 ( ) for none. Univariate analyses Table 2 summarizes the results of the univariate nonparametric statistical analysis used to determine whether individual sonographic features can accurately predict the histologic grade of a tumor. In turn, Table 3 provides the conditional probability of a malignancy grade given the presence of a feature category. Fig. 5 Invasive ductal carcinoma grade 3 in 40-year-old woman. Breast sonogram shows 1.5-cm mass displaying irregular shape, microlobulated margins, parallel orientation, abrupt interface, hypoechoic echotexture, and posterior acoustic enhancement. This mass was categorized as BI-RADS 4c. Mass shape Mass shape was largely ineffective at predicting tumor grade (τ = 0.02, p = > α HB = 0.01). Most masses (84%) were irregular in shape (Figs. 1 5). Mass orientation Mass orientation was ineffective at predicting tumor grade (τ = 0.01, p = 0.249) (Fig. 5). Mass margins Mass margins slightly predicted histologic grade (τ = 0.04, p = < α HB = 0.008). Among the 109 grade 3 lesions, the distribution of margin categories differed significantly from what would be expected if margin did not predict grade, with significantly more microlobulated margins than would be expected (12.0 vs 5.8, z res = 3, p = 0.001) and significantly fewer indistinct margins (4.0 vs 16.7, z res = 4.2, p < 0.001). Among the 153 grade 2 lesions, there were significantly fewer microlobulated margins than would be expected (42 vs 51, z res = 2.2, p = 0.014) and significantly more indistinct margins (30.0 vs 23.5, z res = 2.1, p = 0.018). When margins were categorized as microlobulated instead of indistinct, the odds were 8.57 times ( ) greater that the mass was grade 3 instead of grade 2 and times ( ) greater that the mass was grade 3 instead of grade 1 (Figs. 1, 3, 4). Lesion boundary Lesion boundary slightly predicted histologic grade (τ = 0.05, p = < α HB = 0.007). Among the 108 grade 3 lesions, there were significantly more abrupt interfaces than would be expected (68.0 vs 51.2, z res = 4, p < 0.001) and correspondingly fewer echogenic halos (68.0 vs 51.2, z res = 4, p < 0.001). Among the 153 grade 2 lesions, there were significantly fewer abrupt interfaces than would be expected (54.0 vs 71.9, z res = 4.1, p < 0.001) and correspondingly more echogenic ones (68.0 vs 51.2, z res = 4, p < 0.001). When boundaries were categorized as abrupt instead of echogenic, the odds were 3.04 times ( ) greater that the grade was 3 instead of 2 and 1.66 times ( ) greater odds that the grade was 3 instead of 1. Echo pattern Mass echotexture did not effectively predict histologic grade (τ = 0.01, p = 0.210). The majority (82%) of masses appeared hypoechoic (Figs. 1 5). Posterior acoustic features Posterior acoustics effectively predicted histologic grade (τ = 0.13, p < 0.001, < α HB = 0.006). Among the 108 grade 3 lesions, there were significantly more cases with posterior enhancement than would be expected (53.0 vs TABLE 2: Degree to Which Sonographic Feature Predicts Histologic Grade Feature τ SE p α HB a Mass shape Mass orientation Mass margins Lesion boundary Echo pattern Posterior acoustics Note τ = Goodman and Kruskal tau, α HB = Holm-Bonferroni alpha level (with k = 8). Dash indicates p > a Effect size is statistically significant if p < α HB ; k includes two tests of the BI-RADS (not shown). AJR:199, September 2012 W405

5 Blaichman et al. 27.6, z res = 7, p < 0.001) and significantly fewer ones with shadow acoustics (13.0 vs 37.1, z res = 6.1, p < 0.001). Among the 153 grade 2 lesions, there were significantly fewer cases with posterior enhancement than would be expected (21.0 vs 38.8, z res = 4.7, p < 0.001) and significantly more cases with shadow acoustics (67.0 vs 52.0, z res = 3.7, p < 0.001). Among the 38 grade 1 lesions, there were significantly fewer cases with posterior enhancement than would be expected (2.0 vs 9.2, z res = 3, p = 0.001) and significantly more cases with shadow acoustics (22.0 vs 12.9, z res = 3.3, p < 0.001). When acoustics were categorized as enhanced instead of shadowed, there were 13.0 times ( ) greater odds that the grade was 3 instead of 2 and times ( ) greater odds that the grade was 3 instead of 1 (Fig. 5). TABLE 3: Probability of Malignancy Grade Given the Presence of a Feature Category Feature Probability No. Grade 1 Grade 2 Grade 3 Unconditional Mass shape Oval 20 a Round Irregular Mass margins Circumscribed 1 Microlobulated Indistinct Angular Spiculated Echo pattern Hyperechoic 2 Isoechoic Hypoechoic Complex 10 Anechoic 0 Mass orientation Parallel Not parallel Acoustic Enhanced None Shadow Combination Lesion boundary Abrupt interface Echogenic halo BI-RADS 4a 9 4b c Note Dash indicates cells with expected frequency < 5. a Not shown in order to prevent errors in inference. Total Surgical Grade and BI-RADS Categorization Raters final BI-RADS assessment did not effectively predict histologic grade (d = 0.09, p = 0.103). Given the foreknowledge that they were evaluating malignant masses, raters were to limit themselves to assigning cases to only four BI-RADS categories: 4a c (suspicious abnormality, three tiers) or 5 (highly suggestive of malignancy). For investigating whether a truncated scale would have more predictive power, the data were reanalyzed with the final assessments dichotomized between 5 and those rated 4c or lower. With the truncated scale, BI-RADS category did effectively predict histologic grade (d = 0.12, p = 0.041), but this finding is likely a falsepositive given its critical importance (α HB = < p = 0.041). Moreover, it predicted histologic grade in the wrong direction, with fewer of the 108 grade 3 lesions than expected rated 5 (55 vs 63.2, z res = 2.0, p = 0.023) and correspondingly more rated 4 (53 vs 45.8, z res = 2.0, p = 0.023). Discussion Our study was primarily focused on determining which BI-RADS lexicon sonographic features best characterize high-grade invasive breast carcinomas in women assigned a BI-RADS final assessment category that would lead to a biopsy (i.e., 4a c and 5). Further, we wished to evaluate whether the final assessment category itself could be an effective predictor of histologic grade. Evaluation of lesion shape was addressed in an indirect fashion previously by Rotstein and Neerhut [5] by using the depth-towidth ratio as a proxy. Exclusively examining grade 3 lesions, they reported that the majority were round (with a depth-to-width ratio of 1) (33%) or ellipsoid (with a depthto-width ratio less than 1) in shape (40%). In our study, however, the majority of highgrade (82%) as well as intermediate- (87%) and low-grade (80%) invasive ductal carcinomas displayed an irregular shape, a malignant criteria prompting the radiologists to classify the masses as at least BI-RADS 4. Although shape may be a strong indicator of malignancy, we found it ineffective at distinguishing between histologic grade (τ = 0.02). We found mass margins to be somewhat effective at discriminating between malignancy grades (τ = 0.04). Although angular and spiculated margins two well-established malignant criteria [1] were present in equal frequency among the three different categories, microlobulations and indistinct W406 AJR:199, September 2012

6 Efficacy of Sonographic Features in BI-RADS Lexicon margins differed between grade 3 lesions and those of the other grades because highgrade lesions displayed more microlobulated and fewer indistinct margins. This high prevalence of microlobulated margins with low prevalence of indistinct margins likely reflects the lack of desmoplastic reaction within surrounding tissue [7]. Previously, the relative lack of desmoplastic reaction in high-grade tumors compared with low- and intermediate-grade tumors has been invoked to explain why the former are more likely to present as well-defined lesions, whereas the latter tend to induce spiculation [8]. Direct comparison of our margin results to those of earlier studies is difficult because the earlier studies used different descriptors. Rotstein and Neerhut [5] characterized the margins of their 181 grade 3 masses evaluated by ultrasound as 87% aggressive ( spiculated, microlobulated, or angular ), 11% as nonaggressive ( well-defined smooth ), and 2% as indeterminate. Although well-defined smooth clearly maps to circumscribed and indeterminate to indistinct, the authors do not provide the breakdown of their aggressive category. Mapping our criteria onto the scale of Rotstein and Neerhut, our 108 grade 3 lesions are 95% aggressive, 1% nonaggressive, and 3% indistinct. Although the difference in the prevalence of aggressive lesions in the two studies is significant, the difference ( ) in subcategories may not be. Similarly, although the difference ( 0.15 to 0.05) in the prevalence of nonaggressive grade 3 lesions in the two studies is significant, it is doubtful that our findings are comparable because only two of the 299 masses we evaluated across all grades were well circumscribed. Given the decade-long difference between the onset of data-collection in our respective studies, it is certainly possible that the higher frequencies and better spatial resolution of our ultrasound apparatus led us to misclassify fewer margins as well circumscribed. The argument is the same for the study by Lamb et al. [4], which collected data from 1996 to 1997, identified 104 margins as either poorly defined (90%) or well defined (10%), and found no significant difference in margin definition across grades. Of all the sonographic features, posterior acoustics were by far (τ = 0.13) the strongest predictor of histologic grade. Compared with grade 1 and grade 2 lesions, grade 3 lesions were more likely to display posterior enhancement and less likely to show posterior shadowing, which are classically considered to be good sonographic predictors of malignancy [2]. When reanalyzed using our methodology, the data of Lamb et al. [4] also show that acoustics are a strong predictor of grade (τ = 0.10, p = 0.002). Grade 3 lesions were more likely than expected to display posterior enhancement (18.0 vs 11.2, z res = 3.2, p < 0.001) and less likely than expected to show posterior shadowing (14.0 vs 20.4, z res = 2.6, p < 0.005). Grade 2 lesions were less likely than expected to display posterior enhancement (2.0 vs 8.0, z res = 3.0, p = 0.001). Grade 1 lesions were more likely than expected to show posterior shadowing (12.0 vs 6.9, z res = 2.7, p < 0.004). Similarly, Rotstein and Neerhut [5] found an unexpectedly low occurrence of shadowing among grade 3 masses. Several theories have been put forward to explain the finding of posterior acoustic enhancement in grade 3 lesions, a feature normally associated with benign breast lesions. Kobayashi [7] suggested that the cellular content of a breast mass may be proportional to its enhancement. We would therefore expect higher-grade lesions to display more rapid proliferative rates and thus higher cellularity, leading to increased enhancement. A study by Gozzi et al. [9], however, reported that it was the organization of the tumor tissue, in particular the number of histologic interfaces between the cellular and fibrotic components of the tumor, that ultimately determines the posterior acoustic features of the tumor. The lack of a desmoplastic reaction may also explain this finding in part [4]. At present, the true explanation remains uncertain. With respect to echogenicity, no significant difference was noted among the different grades, and most lesions (82%) appeared hypoechoic. As reported in a study by Schrading and Kuhl [8], we observed a single high-grade carcinoma appearing as fibroadenomalike: the single case assigned as BI-RADS III, which was excluded from further analysis. Interestingly, lesion boundary was found to predict grade (τ = 0.05), with abrupt interfaces being more frequently associated with high-grade masses. This could be related to the pushing margins classically observed in these high-grade lesions, which again likely relate to the lack of desmoplastic reaction induced by these lesions [10]. However, this result is tempered by recent studies that suggest only fair interobserver agreement among radiologists in the assessment of lesion boundaries of malignant masses [11]. A secondary objective of our study was to investigate whether the BI-RADS final assessment categories of 4a through 5 could predict the histologic grade of invasive ductal breast carcinoma in women who would be recommended for biopsy. Although the categorization was ineffective when all three tiers of category 4 were used as distinct levels (probably due to their low interrater reliability [11]), aggregating the scale into just categories 4 and 5 led to good discrimination (d = 0.12); this finding should be treated with skepticism, however, because the critical α HB = < p = Worse, the scale was poorly calibrated with too many category 4 lesions predicted to be grade 3 and too few category 5 lesions. From the data, a likely possibility is that some sonographic features that are good predictors of malignancy and heavily influence the final assessment category assigned, such as irregular shape or posterior shadowing, are not good predictors of actual tumor grade. Our study is not without limitations. Because the study was retrospective, our findings only necessarily apply to women with confirmed malignant lesions, whereas we would ideally like to draw conclusions about women in a prospective manner. Furthermore, our study could not fully simulate actual clinical conditions because the analyses were carried out on selected images rather than performed during real-time sonographic evaluation of the lesions. Moreover, the radiologists were required to select one BI-RADS criterion only, whereas, in normal practice, more than one criterion may apply to a given lesion. The radiologists also knew that this study involved only histologically proven malignant lesions; this knowledge may certainly have led them to be biased in assigning higher grades to lesions than if that were not the case. Even if the aforementioned issues were eliminated, there is the ever-present concern of important interobserver variability in describing individual sonographic features as described in the BI-RADS lexicon. A recent study revealed that certain sonographic features showed better interobserver agreement than others [11]; our study did not take this variability into account. Despite all of these limitations, however, our study provides important information regarding the sonographic evaluation of breast lesions in the context of the current limitations of the BI-RADS lexicon. In conclusion, we have shown that malignancy grade can be slightly to moderately predicted by margins, lesion boundaries, and posterior acoustic sonographic features. AJR:199, September 2012 W407

7 Blaichman et al. In particular, grade 3 invasive ductal carcinomas were more likely than expected to display microlobulated instead of indistinct margins and abrupt instead of echogenic lesion boundaries and were more likely than expected to cause posterior enhancement instead of shadowing acoustics. These findings can likely be explained, at least in part, by their relative lack of a desmoplastic reaction. Further work is necessary to elucidate the full potential of sonography in the evaluation of malignant breast lesions. However, this study offers a little more insight into this important diagnostic tool. References 1. Skaane P, Engedal K. Analysis of sonographic features in the differentiation of fibroadenoma and invasive ductal carcinoma. AJR 1998; 170: Stavros AT, Thickman D, Rapp CL, Dennis MA, Parker SH, Sisney GA. Solid breast nodules: use APPENDIX 1: Logistic Regression For logistic regression, Prentice and Pyke [12] showed that valid point estimators of the odds-ratio and their standard errors may be obtained by fitting the prospective model to retrospective data. APPENDIX 2: Tau Value Tau expresses the proportional reduction in error in predicting grade when a sonographic feature s category is known compared with when simply using the marginal distribution of grade (this is akin to percentage of variance in the grade explained by the feature). Additionally, tau was preferred to lambda, another asymmetric measure of nominal association, because of the large differences in the marginal counts of grade. APPENDIX 3: Measures of Strength of Association Between Two Variables Ultimately, the Goodman and Kruskal tau and the Somers d, like the more familiar Pearson correlation coefficient, are measures of the strength of the association between two variables. Unfortunately, the Pearson correlation coefficient is only a suitable metric when investigating continuous variables on an interval or ratio scale, such as patient age or lesion size. Sonographic features, grade, and BI-RADS categories, however, are noncontinuous variables that occupy more restricted levels of measurement. Sonographic features are categorical (also called nominal) variables they have no intrinsic ordering; grade and BI-RADS score are ordinal variables. Although there is a clear ordering to their categories, the distance between levels of a category cannot be measured. Given that we must use categorical and ordinal measures of association, the Goodman and Kruskal tau and the Somers d were chosen over other measures for two reasons: First, they are asymmetric the strength of effects depends on which variable is designated the predictor and which the outcome, and second, they support a proportional reduction in error interpretation (Appendix 2). APPENDIX 4: Holm-Bonferroni Method of sonography to distinguish between benign and malignant lesions. Radiology 1995; 196: American College of Radiology. Breast imaging reporting and data system (BI-RADS), 4th ed. Reston, VA: American College of Radiology, Lamb PM, Perry NM, Vinnicombe SJ, Wells CA. Correlation between ultrasound characteristics, mammographic findings and histological grade in patients with invasive ductal carcinoma of the breast. Clin Radiol 2000; 55: Rotstein AH, Neerhut PK. Ultrasound characteristics of histologically proven grade 3 invasive ductal breast carcinoma. Australas Radiol 2005; 49: Elston CW, Ellis IO. Pathological prognostic factors in breast cancer. I. The value of histological grade in breast cancer: experience from a large study with long-term follow-up. Histopathology 1991; 19: Kobayashi T. Diagnostic ultrasound in breast cancer: analysis of retrotumorous echo patterns correlated with sonic attenuation by cancerous connective tissue. J Clin Ultrasound 1979; 7: Schrading S, Kuhl CK. Mammographic, US, and MR imaging phenotypes of familial breast cancer. Radiology 2008; 246: Gozzi G, Cressa C, Bazzocchi M, Stanta G, Vidali C. Causes of attenuation of the sound waves in neoplasms of the breast: histologic and echographic correlation study [in Italian]. Radiol Med (Torino) 1986; 72: Tilanus-Linthorst M, Verhoog L, Obdeijn IM, et al. A BRCA1/2 mutation, high breast density and prominent pushing margins of a tumor independently contribute to a frequent false-negative mammography. Int J Cancer 2002; 102: Abdullah N, Mesurolle B, El-Khoury M, Kao E. Breast imaging reporting and data system lexicon for US: interobserver agreement for assessment of breast masses. Radiology 2009; 252: Prentice RL, Pyke R. Logistic disease incidence models and case-control studies. Biometrika 1979; 66: Holm S. A simple sequentially rejective multiple test procedure. Scand J Stat 1979; 6:65 70 In the Holm [13] Bonferroni method, the smallest p value must be less than α/k, the next smallest less than α/(k 1), with k decremented on each subsequent test until the current test s p value is rejected (α = alpha level; k = number of statistical tests performed). The choice of method was a matter of taste the results would be the same if another multiple comparison correction was used. W408 AJR:199, September 2012