We investigated automated quantitative measures of background parenchymal enhancement (BPE) derived from an early versus delayed post-contrast sequence in breast dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) for association with breast cancer presence in a case-control study. DCE-MRIs were retrospectively analyzed for 51 cancer cases and 51 controls with biopsy-proven benign lesions, matched by age and year-of-MRI. BPE was quantified using fully-automated validated computer algorithms, separately from three sequential DCE-MRI post-contrast-subtracted sequences (SUB1, SUB2, and SUB3). The association of BPE computed from the three SUBs and other known factors with breast cancer were assessed in terms of odds ratio (OR) and area under the receiver operating characteristic curve (AUC). The OR of breast cancer for the percentage BPE measure (BPE%) quantified from SUB1 was 3.5 (95% Confidence Interval: 1.3, 9.8; p = 0.015) for 20% increments. Slightly lower and statistically significant ORs were also obtained for BPE quantified from SUB2 and SUB3. There was no significant difference (p > 0.2) in AUC for BPE quantified from the three post-contrast sequences and their combination. Our study showed that quantitative measures of BPE are associated with breast cancer presence and the association was similar across three breast DCE-MRI post-contrast sequences.
Digital mammography is the standard clinical screening imaging modality for breast cancer. The masking effect for women with dense breasts reduces the sensitivity of mammography1. According to the American Cancer Society, a more sensitive modality, breast magnetic resonance imaging (MRI), is recommended as a supplementary examination to digital mammography2, for women considered at elevated risk of developing breast cancer, including typically women with known pathogenic BReast CAncer susceptibility gene 1/2 (BRCA1/2) mutation(s) and their untested first-degree relatives as well as those with a 20% to 25% lifetime risk or higher as calculated by risk models which account for age of affected relatives. Annual surveillance with breast MRI has been shown to reduce incidence of advanced-stage breast cancer in women with pathogenic BRCA1/2 mutations3.
Breast dynamic contrast-enhanced MRI (DCE-MRI) characterizes 3D anatomic properties of fibroglandular (i.e., dense) tissue (FGT) and the use of a contrast agent provides a sensitive way to characterize in-vivo physiologic and biologic activities in breast tissue4 that are relevant to breast cancer risk5. Contrast enhancement of normal breast tissue (other than breast tumors) in DCE-MRI is known as background parenchymal enhancement (BPE), which has been reported to be associated with breast cancer risk6,7,8, when visually assessed by the four qualitative Breast Imaging-Reporting and Data System (BI-RADS) categories (i.e., minimal, mild, moderate, or marked)9. Although clinically useful, BI-RADS-based BPE assessment is subjective, with high intra- and inter-reader variability10, 11, and there is particular difficulty distinguishing the categories of mild and moderate. Fully-automated computerized quantitative methods for objective and reproducible measurement of MRI BPE is in great clinical need.
In current clinical practice, screening breast MRI uses the same protocols as diagnostic breast MRI, including a pre-contrast and several (i.e., 3) post-contrast sequences acquired sequentially after injection of contrast agent9, 12, 13. Recently, abbreviated breast MRI protocols are emerging for breast cancer detection and lesion characterization14,15,16. In screening context the abbreviated MRI aims at developing a quicker and more cost-effective screening protocol16, by reducing certain MR sequences in current full DCE-MRI protocols17. The characteristics of BPE computed from the multiple post-contrast sequences in breast DCE-MRI merit further investigation. The purpose of this study was to investigate automated quantitative measures of BPE derived from an early versus delayed post-contrast sequence in breast DCE-MRI for association with breast cancer presence in a case-control study.
Dataset and imaging protocols
This retrospective study was Health Insurance Portability and Accountability Act (HIPAA) compliant and received Institutional Review Board (IRB) approval by the University of Pittsburgh, Human Research Protection Office (HRPO). Informed consent from patients was waived due to the retrospective nature of this study. All experiments and analyses were conducted in accordance with the Good Clinical Practice guidelines and regulations. The study cohort has been described in detail elsewhere18. To recap very briefly, we studied 102 women in a case-control setting, including 51 unilateral breast cancer cases (gold standard: pathology test) and 51 matched controls (by age [±3 years] and year-of-MRI [±1 year]) who had a single unilateral biopsy-proven benign lesion. Each participant had a suspicious unilateral abnormality rated as BI-RADS 4 or 5 in a diagnostic setting by digital mammography, ultrasound, and/or clinical exam from January 2009 to December 2011 at our institution; they consented to undergo bilateral breast MRI prior to percutaneous core and/or surgical biopsy. The cancer cases were not known to have cancer until after the MRI and pathology confirmation. No breast cancer was diagnosed for controls during an average of 3.7 years of follow up (range 1.4–5.5 years) since their MRI acquisitions.
Following the standard clinical breast MRI protocol at our institution, all DCE-MRIs were acquired by a 1.5T scanner (GE Signa EXCITE, GE Health, Nutley, NJ) in bilateral axial view, using a dedicated 7-channel surface array breast coil (InVivo, Gainesville, FL, USA). The DCE sequences were fat-suppressed and included a pre-contrast and three post-contrast sequences. Bolus injection of the contrast agent, ProHance (Bracco Diagnostics, Princeton, NJ), at 0.1 mmol/kg, 3 cc/sec was followed by a 20 cc saline flush. The first post-contrast sequence acquisition was centered at 90 seconds after contrast material injection. The temporal resolution of the post-contrast sequence acquisition was about 3 minutes for each sequence. Three corresponding subtraction sequences (i.e., SUB1, SUB2, and SUB3) were generated by subtracting the pre-contrast from each of the three post-contrast sequences, respectively. Major imaging parameters were: matrix 512 × 512; field of view 28–34 cm, slice thickness 2 mm; flip angle 10°, repetition time (TR) 5.68 msec, echo time (TE) 2.736 msec. The slice number of the bilateral axial scan ranged from 68 to 160 depending on the size of the breasts.
A total of 153 breasts (51 unilaterally cancer-free breasts from the cancer cases and 102 breasts bilaterally from the controls) were studied for two breast-wise analyses. The main analysis (Comparison A) was to compare MRI BPE measures from the contralateral breasts of both cancer cases and controls. A robustness analysis (Comparison B) compared BPE from the biopsy-proven benign breast of controls to the contralateral breasts of cancer cases. Note that the ipsilateral side of the cancer cases was excluded from BPE analysis at this stage, as our current BPE quantification algorithm has not yet been validated in processing cancer-affected breasts.
MRI BPE quantification
Previously published fully-automated computer algorithms19,20,21 were adapted (Fig. 1) to process breast DCE-MRI scans. Breast-wise BPE measures were computed separately from each of the three SUB sequences (the SUBs were aligned by rigid registration), with the goal of comparing the effect of BPE estimated from an early versus delayed post-contrast sequence. For breasts with benign findings, the benign lesion was included as part of “normal” tissue (as compared to cancerous tissue) in quantifying BPE for the robustness analysis. First, the breast region was separated from the other body parts (e.g., chest cavity) included in the breast MR images19 and the absolute total volume of the breast was computed (|Breast| in cm3). BPE was quantified within the breast region. The extent of MRI contrast enhancement was measured by a voxel-wise intensity enhancement ratio20:
where I denotes the corresponding voxel intensity value in the pre-contrast, post-contrast, and subtraction images. A breast-wise BPE measure (|BPE|) was summarized as the absolute total volume (cm3) of enhancing voxels whose enhancement ratios (R%) had an equal or greater value than a predefined cutoff threshold, R%cutoff. Here, an R%cutoff value of 20% was selected referring to a previous risk-reducing intervention study performed on an independent breast MRI dataset (100 MRI scans)20; as a robustness analysis, BPE was also quantified by using R%cutoff = 30% and R%cutoff = 40%. In addition, a percentage-based breast-wise BPE measure was computed as Because of the nature of full automation, our automated BPE quantification algorithm is reproducible, generating exactly the same results for a given MRI scan.
MRI FGT quantification and mammographic density estimation
For comparison purposes, breast density (fibroglandular tissue) measures were obtained. MRI FGT contents were segmented from the pre-contrast sequence of the MRI scan using a published fully-automated method21 (Fig. 1), generating two quantitative FGT measures: absolute total volume (|FGT|; unit: cm3) and relative percentage . Standard clinical assessment of mammographic density by BI-RADS density categories was retrieved from mammography reports of the digital mammograms acquired within 6 months prior to the analyzed MRI scans.
First, we measured the correlations between the imaging variables of mammographic density, FGT, and BPE using Spearman’s rank-order correlation coefficient. Second, we computed odds ratios of the BPE measures computed from different SUBs with breast cancer using univariate and multivariable conditional logistic regression, where the multivariable regression controlled for four covariates: menopausal status (pre or post), family history of breast cancer, ordinal mammographic density, and quantitative FGT. Family history of breast cancer was encoded as binary (positive if at least one first, second, or third-degree family member was diagnosed with breast cancer). A family-wise approach, i.e., false discovery rate (FDR)22, was applied to adjust for multiple tests at the 0.05 FDR level. Third, we assessed the predictive ability of BPE for distinguishing the cancer cases from controls using unconditional logistic regression analysis and area under the receiver operating characteristic curve (AUC), and DeLong’s test for comparing the AUCs computed from different SUBs. Last, we showed preliminary effects of BPE over four known risk factors (i.e., age, menopausal status, family history, and ordinal mammographic density) in predicting the cancers from controls, where the likelihood ratio test was used to measure the statistical significance on the difference of the AUCs in nested models. All statistical tests were two-sided (Wald tests for conditional logistic regression), with p < 0.05 considered statistically significant. Statistical analyses were performed using SAS software (version 9.3 SAS Institute, Cary, NC).
Patient and imaging characteristics
Table 1 shows detailed patient characteristics. Mammographic density and MRI measures are summarized in Table 2, separately for cases and controls. Mammographic density differed by >1 category only in one matched pair (scattered fibroglandular density for case, extremely dense for control). For amount of fibroglandular tissue, mammographic density and MRI FGT% were strongly correlated (Spearman’s correlation coefficient [SCC] was 0.60, p < 0.0001). At the same time, mammographic density and BPE% (from SUB1) were not correlated (SCC = −0.003, p = 0.97), indicating that BPE% may convey additional risk information as compared with the established risk factor of mammographic breast density. Comparing MRI measures of FGT and BPE, the absolute volume measures |FGT| and |BPE| were moderately correlated with SCC = 0.45 (SUB1), 0.50 (SUB2), and 0.50 (SUB3), respectively, all p < 0.0001, while the percentage measures FGT% and BPE% were not correlated with SCC = −0.03 (SUB1), 0.03 (SUB2), and 0.02 (SUB3), respectively, all p > 0.5. In terms of the correlations between bilateral breasts of a woman, MRI measures between the two breasts (benign lesion and contralateral) of controls showed a strong correlation, with SCC = 0.56 for BPE% (SUB 1) and SCC = 0.86 for FGT%, both p < 0.0001.
Associations of BPE measures with breast cancer presence
In this study cohort, none of mammographic density, |FGT|, or FGT% predicted case/control status in univariate or multivariable models (p > 0.4). BPE measured in different SUBs showed an association with breast cancer (Table 3). As seen in Comparison A (using contralateral breast of controls), univariate conditional logistic regression results reflected the odds of malignancy 1.5 times higher per 200 cm3 increase in |BPE| and 3.1 times higher per 20 percentage point increase in BPE% (SUB1). Compared to the univariate models, odds ratios for BPE measures were slightly higher in multivariable models controlling for menopausal status, family history of breast cancer, BI-RADS-based mammographic density, |FGT|, and FGT%. We found that the odds ratios were similar for the three subtracted sequences (SUB1, SUB2, and SUB3). After applying multiple test control for 6 comparisons at the 0.05 overall FDR, the 6 adjusted p-values remained statistically significant (all p < 0.031 for the 6 univariate analyses and all p < 0.021 for the 6 multivariable analyses).
In Comparison B (using benign breast of controls), we found overall similar results on the association of BPE with breast cancer (Table 3). We noticed that the odds for BPE% almost doubled the corresponding values in Comparison A. After applying the FDR adjustment, the p-values also remained statistically significant (all p < 0.038 for the 6 univariate analyses and all p < 0.032 for the 6 multivariable analyses).
Similar trends/results were observed for various robustness analyses as well. When excluding the six pairs for whom the cancer cases had additional high-risk factors as described above, odds ratios remained similar. Similar odds ratios were also observed when the 10 mismatched pairs in menopausal status were excluded from analysis. In menopausal status subgroup analysis, the trend still held for the 17 postmenopausal pairs (odds ratio = 13.6, p = 0.048) and the 24 premenopausal pairs (odds ratio = 2.4, p = 0.17), despite the smaller sample size.
BPE quantified from different SUBs on distinguishing cancer cases from controls
The AUCs of the measures of |BPE| and BPE%, individually or combined, for distinguishing cancer cases from controls, are shown in Table 4. AUCs were not significantly different (p > 0.2) for all three SUBs and their combination. The percentage BPE measure (BPE%) showed a greater AUC than the absolute volume measure (|BPE|), consistently in three SUBs, although the difference was not statistically significant (p > 0.05). Combining |BPE| to BPE% generated a non-significant increase in AUC (except for SUB3 where p < 0.05) compared to BPE% alone. The ROC curves of the combination of |BPE| and BPE% computed from the three SUBs were shown in Fig. 2.
BPE versus basic risk factors on distinguishing cancer cases from controls
As shown in Fig. 3, when using only four basic factors (i.e., age, menopausal status, family history, and ordinal mammographic density), the AUC of unconditional logistic regression in distinguishing cancer cases from controls was 0.578. When the combination of |BPE| and BPE% computed from SUB2 (we chose SUB2 because it achieved a greater AUC compared to SUB1 and SUB3) were added to the four basic factors, AUC was 0.673 in Comparison A and 0.687 in Comparison B, respectively, both with a statistically significant increase (p = 0.0029 in Comparison A and p = 0.0026 in Comparison B) relative to the AUC of using the four basic factors alone.
As an exploratory analysis for the 51 cancer cases, we found that none (all p > 0.15, using SCC, Wilcoxon rank-sum test, or Fisher’s exact test as appropriate) of the quantitative FGT, BPE measures (from any of the three post-contrast sequences), and basic risk factors (i.e., age, menopausal status, mammographic density, and family history) was associated with tumor size or tumor type (i.e., tumors involving IDC [n = 45] vs others [DCIS and ILC, n = 6]).
Qualitatively assessed breast MRI BPE using the BI-RADS categories has been studied on its relationship with breast cancer presence or risk6,7,8, 23. Two case-control studies6, 7 reported an association between qualitative BPE and breast cancer risk in high-risk screening cohorts. While a recent study8 showed an association of BPE with breast cancer presence, another study23 on non-high-risk patients found an opposite result. The difference in the nature of the study populations may account for the discrepant findings of these previous studies. In the present work, we performed a quantitative case-control study and showed that fully automated MRI BPE measures (both |BPE| and BPE%) are associated with breast cancer presence. Our findings are in line with several previous qualitative reader studies6,7,8, wherein a variety of odds ratios for BI-RADS-based BPE assessment were reported. Moreover, we showed that BPE measured in three different SUBs has a similar effect on its association with breast cancer presence. Our study therefore added a new contribution to the literature in examining the relationship of BPE measures with breast cancer presence/risk.
In our study cohort, neither BI-RADS-based mammographic density categories nor quantitative MRI FGT measures were correlated with breast cancer presence. In previous reader studies, MRI FGT showed a weak6 or null7 association with breast cancer risk. Together, it seemed that the association between BPE and breast cancer presence/risk may be independent of mammographic density and MRI FGT, but further validation is warranted for this finding.
Per the American College of Radiology (ACR) guidelines, the first post-contrast sequence is suggested for BI-RADS-based BPE assessment9. We quantified BPE levels in each of the three post-contrast sequences of the standard DCE-MRI protocol at our institution. Our results showed that the effects of association between BPE and breast cancer presence were similar for the three post-contrast sequences or SUBs: SUB1 had a slightly larger odds ratio than SUB2 or SUB3, and there is no statistically significant difference in terms of the AUC performance across the three SUBs. These findings imply that for computing quantitative BPE as a potential breast cancer presence/risk biomarker, any of the three post-contrast sequences may be equivalent in the context of breast MRI screening. More specifically, a single early post-contrast sequence (SUB1) may be adequate for use in a practical clinical workflow. This finding may be in line of an emerging abbreviated breast MRI screening protocol—the FAST MRI techniques16, 17 —which intended to use only the first post-contrast subtracted sequence (and the Maximum Intensity Projection images) for screening and had shown a similar screening performance with using the full DCE-MRI sequences17.
Currently, no standard value has been established for the intensity enhancement ratio threshold R%cutoff in quantifying BPE, as BPE has been assessed qualitatively in the past6,7,8,9, 23. In computer-aided breast cancer detection/diagnosis software, an R% cutoff threshold value of 30% or 50% was clinically observed24. The lower cutoff value of 20% reported here, may reflect specific/different properties associated with characterizing the contrast enhancement on normal breast tissues for studying breast cancer risk. In the robustness analysis of testing BPE measures quantified at R%cutoff = 30% and 40%, we found similar results with R%cutoff = 20%, indicating that our quantitative BPE measures may be a fairly robust biomarker for breast cancer, valid across a range of parameter values (i.e., 20%, 30%, and 40%) for R%cutoff. We noticed that the contrast-to-noise ratio of breast MR images is affected by DCE-MRI parameters and type of contrast agents. Therefore, the findings on the R%cutoff parameter for BPE quantification are subject to further investigation.
In this work we performed two comparisons (Comparison A and Comparison B), aimed to test the effects of BPE quantified on a mixture of negative and benign-containing breasts. The associations found in both the two comparisons suggest robustness of BPE, in the sense that ultimately we would expect to derive BPE as an imaging biomarker from a wide range of “normal” or “non-diseased” breast tissue. That being said, it needs to point out that the inclusion of the benign lesion in quantifying BPE in benign-containing breasts may have introduced bias in Comparison B. Currently our algorithms lack the function of segmenting the benign lesions out in the breast, which prevented us from an additional robustness analysis by looking into the effects of BPE quantified from the breasts that have excluded the benign lesions.
The strengths of this study include 1) use of fully-automated computer methods yielding objective and reproducible BPE quantification, 2) use of MRI scans acquired with a fairly consistent imaging protocol, reducing complexity of dealing with varying MR imaging protocols and parameters, and 3) robustness analysis using a mixture of negative and benign-finding breasts. Our study has some limitations. This is a retrospective single-institutional study and our sample size is relatively small. Therefore further evaluation is warranted for assessing the generalizability of our findings on a larger breast MRI dataset (possibly a multi-center study), but this proof-of-concept study will be instrumental in guiding appropriate design of larger retrospective and/or prospective studies. We were not able to do a comparison of BPE between our quantitative assessment and the BI-RADS-based assessment, because the BI-RADS-based assessment of BPE was not yet fully implemented or standardized at our clinics during the time the study cohort had the MRI scans (2009–2011) and, therefore, not available for the majority of the MRI scans for analysis. However, this comparison will be feasible when more recent breast MRI scans are analyzed. In addition, the AUCs of 0.689 for BPE alone and 0.687 when combined with the four basic factors are in line with the reported AUCs (range 0.6–0.7525, 26) of existing risk models (such as the Gail model and the Claus model). Of note, as the incremental value of a biomarker to improve breast cancer risk prediction cannot be assessed conclusively in matched case-control studies27,28,29, the increased AUCs of BPE measures relative to the basic risk factors should be interpreted with caution; yet, they are still indicative of the preliminary value of BPE in comparing to existing risk factors. Future larger studies are warranted for a further evaluation of BPE’s prediction capability. Finally, we noticed that a group of premenopausal women had their MRI examinations outside the clinically recommended scanning window (second week of the menstrual cycle), which may have introduced variations in the BPE assessment because BPE levels vary between menstrual cycle weeks30,31,32. However, as indicated in previous work18, 31, the measured BPE outside the second week in the control group would have yielded a higher value than the actual levels, attenuating the difference of BPE between cancers and controls in the case-control analysis. Despite this attenuation, we still found a significant difference of the BPE measures between the cancer and control groups on all of the three SUBs.
In summary, this preliminary study showed that fully automated quantitative assessment of breast MRI BPE is associated with breast cancer presence and the association has a similar effect across the three sequential post-contrast DCE sequences. Our study supports further investigation of BPE as a potential biomarker of breast cancer risk, where a single early post-contrast sequence may be adequate for estimating breast cancer risk in breast cancer MRI screening. Quantitative MRI BPE is expected to ultimately improve breast cancer risk prediction33 and have direct benefits to enhance clinical utility of screening breast MRIs and to aid in more informed breast cancer risk management34.
Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This work was supported by a National Institutes of Health (NIH)/National Cancer Institute (NCI) R01 grant (#1R01CA193603), a Radiological Society of North America (RSNA) Research Scholar Grant (#RSCH1530), a Precision Medicine Pilot Award (#MR2014-77613) from the University of Pittsburgh Cancer Institute-Institute for Precision Medicine, a Biomedical Modeling Pilot Award from the Clinical and Translational Science Institute of the University of Pittsburgh, and a Competitive Medical Research Fund of the University of Pittsburgh Medical Center (UPMC) Health System. This project also used the University of Pittsburgh Cancer Institute Biostatistics Shared Resource Facility that is supported in part by award P30CA047904.