From March to May of 2012, we prospectively performed FFDM and DBT on women who were referred to our institution with clinical signs and symptoms of breast cancers, or who had suspicious findings on screening mammography or ultrasonography (US). Of the 219 consecutive women who received FFDM and DBT, 16 patients were excluded who either had a surgical clip in the breast or had a history of vacuum assisted breast biopsy as either could provide information regarding lesion location. As a result, a total of 203 women were enrolled in this study. The mean age of the patients was 49.8±11.2 years (range, 22–78 years). Among these patients, 107 (52.7%, 107/203) were asymptomatic but exhibited abnormalities on screening mammography (n=68) or US (n=39). Abnormal findings on FFDM in 68 patients included: dense breast tissue in three, mass in 20, calcification in 20, asymmetry in 17, distortion in one, mass with calcification in one, mass with distortion in one, asymmetry with calcification in three, calcification and distortion in one, and left breast mass and right breast calcification in one. Thirty-nine patients who had an abnormality detected on US showed either a benign mass, benign calcification, asymmetry, or no abnormal findings on mammography. The other 96 patients (47.3%, 96/203) showed clinical signs and symptoms of breast lesions including palpable lumps (n=85), pain (n=6), nipple discharge (n=3), nipple retraction (n=1), and Paget's disease (n=1). Of the 203 patients in this study, 98 patients had previously been evaluated elsewhere for known breast cancer. Clinical and histopathologic data were obtained from the electronic medical records and standard histopathologic report.

Breast composition of the patients according to the Breast Imaging Reporting and Data System (BI-RADS) recorded during the initial clinical interpretation were as follows: almost entirely fatty tissue (BI-RADS composition of a) (n=8, 3.9%), scattered fibroglandular densities (BI-RADS composition of b) (n=38, 18.7%), heterogeneously dense tissue (BI-RADS composition of c) (n=100, 49.3%), and extremely dense tissue (BIRADS composition of d) (n=57, 28.1%). For statistical analysis, breast density was dichotomized as fatty (BIRADS density categories a and b) versus dense (BIRADS density categories c and d).

FFDM and DBT images of both breasts in the craniocaudal (CC) and mediolateral oblique (MLO) positions were obtained using a commercially available device (Selenia Dimensions System; Hologic, Bedford, USA) in the combo mode by one dedicated technologist. DBT acquisition was performed with the X-ray tube moving over a 15° arc, while the breast remained compressed in the conventional manner, allowing a series of ultra-low dose images. After acquisition, data from the projection images were combined to create a full 3D-image set of 1 mm slices of the breast. Radiation doses used for DBT with FFDM were approximately twice that used for routine FFDM alone. Mean glandular dose per examination was 1.71±0.51 mGy for each CC view and 1.65±0.61 mGy for each MLO view for DBT, and 1.68±0.68 mGy for each CC view and 1.62±0.75 mGy for each MLO view for FFDM.

Six radiologists who specialize in breast imaging and with experience ranging from 2 to 15 years in reading mammograms participated in this reader study. The readers were blinded to the patient history, clinical reports, histopathologic information, and results obtained with other modalities. Prior to the study, each reader received additional training using DBT images from 50 known cases (35 cancers and 15 benign lesions). Those images had been obtained at our institution for other research purposes and were not included in this particular study. After training was completed, the readers independently interpreted two-view FFDM alone, two-view DBT alone, and combined FFDM and DBT datasets. To minimize learning bias, a 12-week interval elapsed between each reading session. The first and second reading sessions each consisted of half of the two-view FFDM alone and half of the two-view DBT alone images, randomized, and presented in sequential order. The third reading session consisted of all of the images from the combined FFDM and DBT dataset. The readers were instructed to detect the tumors, if any, and to record the location of each lesion. Identifiers were used to confirm that the readers had correctly identified a cancer. After identifying the lesions, the readers provided a number from 0 to 100 to rate the likelihood of malignancy and designated a forced BI-RADS category of 1, 2, 3, 4 or 5 to indicate the most likely outcome based on the appearance of the lesion. Cases assigned to BI-RADS categories of 1, 2, or 3 were considered to be normal or benign, and those assigned to BI-RADS categories of 4 or 5 were considered to be abnormal.

After the independent reader study, two dedicated breast radiologists who did not participate in the first part of the study reviewed the FFDM and DBT images together using the reference location of cancers identified on US or magnetic resonance imaging. They determined mammographic lesion size, distance from the nipple to the lesion, lesion location quadrant, lesion type, and breast composition based on BIRADS classification. The lesion type was coded separately as to whether it had the characteristics of mass, microcalcification, asymmetry, or architectural distortion.

Results of histopathologic analysis performed on samples obtained via biopsy or surgery, if available, or a follow-up period of more than 2 years in those cases not biopsied were used as the reference standard. Cases with malignant biopsy results were considered positive. Cases with benign biopsy results or cases that did not undergo biopsy but did not have evidence of malignancy during 2 years of follow-up imaging were considered negative. Lesion matching between FFDM/DBT and histopathology was performed off-site by two radiologists provided with the histopathologic reports and the standardized templates used in the image review. Surgical histopathology and core specimens were reviewed by one pathologist with 25 years of experience in breast histopathology.

The clinical, radiological, and histopathological findings of the subsequent surgical specimens, if available, for the 206 lesions were reviewed. Diagnostic performances of each reader with the three modalities were compared using jackknife alternative free-response receiver operating characteristic (JAFROC) curves constructed from the likelihood of malignancy rating for FFMD alone, DBT alone, and combined FFDM and DBT images. Diagnostic sensitivities and specificities were calculated from BI-RADS scores. The McNemar test was used to compare individual sensitivity and specificity for each modality. JAFROC analysis was used to analyze free-response receiver operating characteristic (FROC) data acquired using the multiple-reader multiple-case protocol. JAFROC figure of merit (FOM) was adjusted for the clustering effect of multiple readers as a random factor. Analysis of overall sensitivity and specificity across all readers for each modality was performed using a generalized estimating equation with an exchangeable correlation structure.

To identify lesion variables which could affect the cancer detectability of the combined modality, associations between detectability and the patient's clinical factors (age and symptoms) as well as lesion variables (mammographic lesion size, distance from the nipple to the lesion, lesion location quadrant, lesion type, breast composition based on BI-RADS classification, and histopathology) were assessed using ordinal logistic regression analysis. The detectability was assumed to be proportional to the number of readers who detected the lesions. To determine the most important factors affecting the detection performance of the combined modality, the association between the patient's clinical factors, lesion variables, and the number of readers who detected the lesions were examined using a cumulative logit model based on multinomial distribution with the proportional odds assumption. The proportional odds assumption was tested, and multicollinearity among the predictors was controlled through stepwise variable selection. All statistical analyses were performed using SPSS version 19.0 (IBM Corp., Armonk, USA), SAS version 9.2 (SAS Institute Inc., Cary, USA), and JAFROC (http://www.devchakraborty.com); p-values <0.05 were indicative of statistical significance.

Of the 206 lesions identified in 203 women, 129 (63.0%) were malignant and 77 (37.0%) were benign. Histopathologic confirmation was available for 188 lesions, this included all lesions identified as carcinoma (n=129) as well as 59 benign lesions. The duration of US follow-up for lesions with benign histologic findings was 22.5 to 23.8 months (median, 23.1 months); lesion stability was confirmed in all cases. Eighteen lesions were presumed to be benign without histologic confirmation, and there was no evidence of malignancy during the follow-up period of more than 2 years.

The mean tumor size of the invasive cancers was 2.1±1.3 cm (range, 0.1–7.7 cm), and that of ductal carcinoma in situ (DCIS) was 3.1±1.8 cm (range, 0.5–6.0 cm). Malignant histologic findings included 104 infiltrating ductal carcinomas (IDCs) with or without DCIS, 3 mucinous carcinomas, 2 invasive papillary carcinomas, 1 medullary carcinoma, 2 invasive lobular carcinomas, 1 invasive mixed ductal and lobular carcinoma, and 16 DCIS. Benign lesions which were histologically confirmed included fibrocystic change (n=18), fibroadenoma (n=28), intraductal papilloma (n=5), columnar cell change (n=2), sclerosing adenosis (n=1), radial scar (n=1), phyllodes tumor (n=1), benign mammopathy (n=1), atypical lobular hyperplasia (n=1), and lobular carcinoma in situ (n=1).

JAFROC FOMs using the likelihood of malignancy rating (%) for FFDM alone, DBT alone, and the combined modality for each reader as well as the overall FOM for all readers are listed in Table 1. The FOM of the combined modality was significantly higher than that of FFDM alone in four readers (p=0.045 for reader A; p=0.001 for reader B; p=0.002 for reader D; and p=0.003 for reader F). For readers B and D, the FOM of DBT alone was also higher than that of FFDM alone (p=0.014 for reader B; p=0.026 for reader D). For reader F, the FOM of the combined modality was higher than that of DBT alone (p=0.008). The overall FOM of combined DBT and FFDM was 0.827, which was significantly higher than that of FFDM alone (0.775) (p<0.001). The overall FOM of DBT alone was also higher than that of FFDM alone (0.807 vs. 0.775, p=0.027). There was no statistically significant difference between the FOM for the combined modality and that of DBT alone (p=0.144).

The sensitivity and specificity estimates of the three modalities for each reader as well as for all the readers combined are listed in Table 2. The sensitivity of the combined modality was significantly higher than that of FFDM alone for four readers (p=0.004 for reader A; p<0.001 for reader B; p<0.001 for reader D; and p=0.035 for reader F). For readers A and B, the sensitivity of DBT alone was also higher than that of FFDM alone (p=0.024 for reader A; p=0.015 for reader B). For reader D, the sensitivity of the combined modality was higher than that of DBT alone (p=0.035). Regarding specificity, two readers showed significantly higher specificity for the combined modality than for FFDM alone (p<0.001 for reader E; p=0.008 for reader F). The specificity of the combined modality was also higher than that of DBT alone for two readers (p=0.022 for reader E; p=0.009 for reader F). The overall sensitivity of the combined modality was significantly higher than that of FFDM alone (80.0% vs. 73.2%, p<0.001); and the overall sensitivity of DBT alone was also higher than that of FFDM alone (78.3% vs. 73.2%, p=0.007). However, differences in overall specificities between the three modalities were not statistically significant.

Among the 129 cancers (113 invasive cancers and 16 DCIS), 48 cancers (37 invasive cancers and 11 DCIS) that were missed on FFDM alone were detected using the combined modality by at least one reader (29 cancers by one reader, 13 cancers by two readers, 2 cancers by three readers, 2 cancers by two readers, and 2 cancers by six readers) (Figure 1).

Distribution of the lesions' characteristics (clinical, mammographic, and histopathologic characteristics) according to the number of readers who detected the lesions with combined DBT and FFDM are summarized in Table 3. Of 129 cancers, 98 lesions (76.0%) were detected by five or more readers, 9 lesions (7.0%) by three or four readers, and 12 lesions (9.0%) by one or two readers. Ten lesions (8.0%) were not detected by any reader. Ordinal logistic analysis revealed mammographic lesion size (p=0.009), patient's symptoms (p=0.003), mammographic identification of microcalcifications (p<0.001), architectural distortion (p=0.049), and the presence of masses (p<0.001) to be significantly associated with cancer detectability using the combined modality. Patient's age, distance from the nipple to the lesion, mammographic density, lesion location quadrant, and pathology were not significantly associated with detectability of the lesions. In multivariate analysis, mammographic findings of masses (odds ratio, 76.0; p<0.001) or microcalcifications (odds ratio, 61.8; p<0.001) were significantly associated with a higher cancer detection rate.

Features of the 10 breast cancers that were not detected on the combined modality by any reader are shown in Table 4. Among these lesions, there were two DCIS and eight IDCs. Of the 10 lesions, four showed focal asymmetry and one showed a subtle architectural distortion on an unblinded review by two dedicated radiologists, although none of the six blinded readers detected the lesions on prospective review. The lesions were all misinterpreted as focal asymmetric prominent parenchyma, even in the case with a fatty breast background (patient number 6). Five cases showed no abnormal findings even on retrospective unblinded review of combined DBT and FFDM. All five cancers that were occult on combined DBT and FFDM were obscured by dense parenchymal tissue, and BI-RADS breast density was interpreted as heterogeneously dense or extremely dense in each case (Figure 2).