Complementary Value of Contralateral Parenchymal Enhancement on DCE-MRI to Prognostic Models and Molecular Assays in High-risk ER⁺/HER2⁻ Breast Cancer

The past decades have shown a continuous trend to individualization of the selection of chemotherapy. Markers such as patient age, tumor size, tumor receptor status, and nodal involvement are used to select chemotherapy using prognostic models such as the Nottingham Prognostic Index and Adjuvant! Online (1, 2). Nonetheless, variation still exists in treatment outcome in patients who have similar markers.

Overtreatment of breast cancer remains a concern in subgroups of patients, leading to adverse side effects such as chronic fatigue without survival benefit. Conversely, undertreatment may lead to recurrence and increased risk of ultimately succumbing to breast cancer. Hence, patient-specific treatment of breast cancer is desired.

More recently, further individualization of chemotherapy has been accomplished using molecular assays such as the 70-gene signature and the 21-gene recurrence score, addressed by Mammaprint and Oncotype DX, respectively. These assays have been shown to complement prognostic models (3–6). Potential reduction in overtreatment up to 22% has been reported in MINDACT and TAILORx trials (5, 6), but it is likely that subgroups of patients still exist in which overtreatment occurs.

Current research on prognostic markers focuses in large part on the primary tumor, while the breast stroma is relatively underexposed. Nonetheless, the stroma may play an important role in breast cancer risk prediction (7–9), treatment response (10–13), and assessment of outcome (14, 15). Genomic changes in the stroma may trigger cell transformations leading to malignancy (16). Perfusion of the stroma surrounding the tumor on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) has been associated with patient survival (10, 11). We recently found that pronounced contrast-enhancement of the stroma in the unaffected healthy breast on MRI has been associated with better survival of patients with estrogen receptor (ER)⁺/HER2⁻ breast cancer (17).

Given the interaction between tumor and stroma, their combination is likely to contain complementary information. Thus, markers of survival derived from the primary tumor may be complemented with those derived from healthy stroma. Nonetheless, to the best of our knowledge nothing is currently known about the potential gain of combining such markers.

Pursuing further individualization of chemotherapy, the aim of this study is to determine whether, and if so, to what extent, markers of stroma are able to identify patients at low risk of metastases or death from the group considered to be at high risk according to routine prognostic models derived from the tumor.

The study design is illustrated in Fig. 1. Patients with an ER⁺/HER2⁻ breast cancer who had a preoperative breast DCE-MRI were included. In these patients, the association between pronounced parenchymal MR contrast enhancement in the unaffected healthy breast and survival was discovered (17). In the current study, new data were added by analyzing the fresh-frozen tumor tissue and generating the molecular assays. Using long-term survival as an endpoint, the complementary value of MRI relative to these molecular assays and conventional prognostic models was explored. Furthermore, the robustness of the biomarker was evaluated. The following paragraphs describe these steps in detail.

Data of this single-institution (Netherlands Cancer Institute-Antoni van Leeuwenhoek Hospital) cohort study was acquired after approval of the institutional review board and written informed patient consent. The patients participated in the prospective Multimodality Analysis and Radiological Guidance IN breast conServing therapy study (MARGINS, 2000–2008); patients who were eligible for breast conserving therapy on the basis of conventional imaging and clinical assessment were consecutively recruited for an additional preoperative breast MRI. Proof of breast cancer was acquired using core biopsy or fine-needle aspiration. Treatment plans were established in consensus by a multidisciplinary team of breast cancer specialists. The patients with unilateral invasive ER⁺/HER2⁻ breast cancer were consecutively included for the current retrospective analysis.

MR images were acquired using a 1.5T scanner (Magneton, Siemens) with a dedicated double-breast array coil (CP Breast Array, four channels, Siemens). An unenhanced fast low-angle shot three-dimensional T1-weighted image was acquired. Gadolinium containing contrast (14 mL, Prohance) was administered at 3 mL/second using a power injector followed by 30-mL saline solution. Subsequently, four consecutive postcontrast series at 90-second interval were acquired. The imaging parameters were: repetition time 8.1 ms, echo time 4.0 ms, flip angle 20°, isotropic voxel size 1.35 × 1.35 × 1.35 mm³, and field of view 310 mm.

The largest tumor diameter on MRI was measured in three orthogonal planes by a dedicated breast MR imaging radiologist with more than 10 years of experience (C.E. Loo), after which the largest of the three measurements was recorded.

ER status was positive if more than 10% of the cells were stained positive (18). HER2 status was negative if scored less than 3 on IHC and if in situ hybridization did not demonstrate gene amplification (18). Histologic grade was assessed using the Bloom and Richardson method (19). Axillary load was divided into three groups: no, 1–3, or >3 lymph nodes positive for malignancy.

The investigated prognostic models based on clinicopathology were the Nottingham Prognostic Index (1) and the most recent (2008) Dutch clinical guidelines for chemotherapy (18). The Nottingham Prognostic Index was defined as (0.2*S)+N+G, where S is the size of the tumor in cm; N is 1 for 0, 2 for 1–4, and 3 for >4 lymph nodes positive for malignancy; and G is the histologic grade. Patients with a Nottingham Prognostic Index above 3.4 were considered to be at high risk (1).

According to the Dutch clinical guidelines, node-negative patients were indicated for chemotherapy (and thus considered at high risk) if they were younger than 35 years (except if they had grade I tumors less than 1.1 cm), older than 34 years with tumor size 1.1 to 2 cm and grade II or III, or older than 34 years with tumor size larger than 2 cm. Patients with positive lymph nodes were also indicated for chemotherapy (18).

The investigated molecular assays were the 70-gene signature and the 21-gene recurrence score (3, 4). From hereon, the official molecular assays will be referred to as Mammaprint and Oncotype DX, and the reconstructed molecular assays as the 70-gene signature and 21-gene recurrence score, respectively.

Fresh-frozen tumor samples corresponding to the lesions in the MR images were collected from the tissue bank of the Netherlands Cancer Institute/Antoni van Leeuwenhoek Hospital. RNA and DNA was extracted using QIAcube and AllPrep DNA/mRNA/miRNA Universal Kit (Qiagen). Extraction was limited to samples that had 30% or more tumor tissue. Samples with low RNA quality, RNA Integrity Number < 6 from Bioanalyzer 2100 (Agilent), or low amount of RNA were not further processed. RNAseq libraries were prepared using Truseq stranded mRNA Library Prep Kit (Illumina). These were sequenced on a Hiseq 2500 (Illumina) with 65-bp single-end reads. After cutting adapter sequences, RNA sequence reads were mapped to the human genome (GENCODE 23) with STAR 2.5.0a to quantify the RNA per gene (20).

The 70 probes described in the 70-gene signature were mapped to the genome to assess which probe corresponds to which gene: 63 probes were mapped to 59 genes, 4 to multiple genes, 1 to an intron, and 2 between genes. RNA sequencing data were normalized and log-transformed (21). The 70-gene signature score was calculated as the Spearman correlation of the 63 probes with the average good-prognosis profile (3). The 70-gene signature score was dichotomized using a threshold resulting in similar proportions of predicted labels on the lymph-node negative patients in our data set and the data set of van't Veer and colleagues (3). Patients with a 70-gene signature score below this threshold were considered to be at high risk.

Log2-transformed counts were scaled by gene length. A reference expression value was calculated per sample by taking the average expression of the reference genes from Paik and colleagues (4), except for the ribosomal gene RPLP0, which is not measured by our RNAseq protocol. Normalized gene expression was calculated as 15-(log2(10)-(expression-reference)). The risk score was calculated on the normalized gene expression using the formula from Paik and colleagues (4). The 21-gene recurrence score was split into low-, intermediate-, and high-risk based on the thresholds described by Paik and colleagues (4).

The automatic processing of the MR images is described elsewhere in detail (17). In short, field inhomogeneities were corrected (22), the breast area was segmented on the MR images (23), and the stromal tissue was segmented in the contralateral breast (24).

Time series were registered to each other using deformable registration to compensate for patient motion (25). Late enhancement was calculated at each voxel location as the relative increase in signal intensity between the first post-contrast scan and the last post-contrast scan. The late enhancement values were sorted from lowest to highest, and the mean of the top 10% was calculated. This unitless value that can be compared between patients is from hereon called the contralateral parenchymal enhancement (CPE).

The distribution of CPE was analyzed with respect to the prognostic models and the molecular assays. The distribution of CPE was tested for normality using the Shapiro–Wilk test. CPE was split in tertiles into low, intermediate, and high enhancement. Differences in patient, tumor, and systemic treatment characteristics (age at diagnosis, largest tumor diameter on MRI, histological grade, axillary load, endocrine, and/or chemotherapy) between the CPE groups were tested using the analysis of variance for continuous normal characteristics, Kruskal–Wallis test for continuous non-normal characteristics, and χ² test for discrete characteristics. The agreement between the various prognostic models for the identification of high-risk patients was tested using Cohen κ.

To investigate potential bias in the availability of viable fresh-frozen tumor samples, patient and tumor characteristics as well as survival of patients for whom tissue was available were compared with those for whom tissue was absent.

The primary endpoint of the current study was to establish whether CPE was able to identify patients at low risk of metastases or death from the group considered to be at high risk according to routine prognostic models derived from the tumor. Baseline survival analyses are reported as reference. Survival analysis was performed using Cox proportional hazards models for overall survival (OS) and distant disease-free survival (DDFS) up to 10 years using the standardized definitions for endpoints by Hudis and colleagues (26). In short, patients had an OS event in case of death, and a DDFS event in case of distant metastases or death. Follow-up time in months was defined as the time between diagnosis and censoring at the last date of follow-up, regardless of whether they were scheduled for future follow-up, or whether they had been lost to follow-up. The median of the follow-up times was reported. The association of CPE with survival was analyzed between CPE groups and as a trend.

To take possible differences in aforementioned patient, tumor, and systemic treatment characteristics between CPE groups into account, these characteristics were adjusted between CPE groups using inverse probability weighting (IPW) based on propensity scores (27). IPW adjustment assigned weights to each patient, matching the distributions of the characteristics between CPE groups. In more detail, first the relation between each characteristic and the CPE groups was modeled using multinomial logistic regression. Continuous variables were modelled using restricted cubic splines. After this, the propensity score was calculated. For each patient, the propensity scores in the CPE groups sum to one. The final weighting per patient was defined as the inverse propensity score of the CPE group that patient belongs to. IPW adjustment avoids adding degrees of freedom in the Cox proportional hazards model, hence allowing for correction of confounders when few events occur. A potential drawback of IPW adjustment is the possibility of extreme weights for some patients, which can introduce bias (28). Weights of the IPW adjustment were updated with each analysis.

Missing values (largest tumor diameter on MRI: 4 patients, axillary load: 4 patients, histologic grade: 8 patients) were multiple imputed before IPW adjustment (29). All variables including outcome variables were used in the imputation (30, 31). The imputed values were exclusively used for the IPW adjustment. Results are shown pooled for the ten imputations. IPW-adjusted Kaplan–Meier survival curves were generated and 95% confidence intervals were assessed on the basis of 2,000 bootstrap resamples. Cumulative survival differences after 10 years were reported. Survival analyses were reported for unadjusted CPE, CPE adjusted for patient and tumor characteristics, and CPE adjusted additionally for systemic treatment. We focused on each group of patients at high risk according to each prognostic model separately. In each of these analyses, CPE tertiles and IPW were recalculated before survival analysis.

Competing risk analysis was performed to compare survival of breast cancer–related deaths and non-breast cancer–related deaths in the patient population (32). The distinction between breast cancer–related deaths and non-breast cancer–related deaths was based on the 10th revision of the International Classification of Diseases (33).

The robustness of CPE was assessed by altering the percentile threshold from 0% to 99% by increments of 5%. For each threshold, the mean of the voxels above that threshold was calculated. These CPE values were subsequently split on tertiles in low, intermediate, and high CPE groups. Survival analysis with IPW adjustment was performed and HRs were reported.

For all statistical tests, a two-tailed P <0.05 was considered significant. Statistical analysis was performed using R 3.3.0 (R Foundation for Statistical Computing).

A total of 435 patients were eligible for this study. These 435 patients were also used in the original study that identified the association between CPE and survival (17). Twelve patients (3%) were excluded because of previous surgery and 8 (2%) because MR image acquisition or registration failed. Patient and tumor characteristics of the remaining 415 patients are summarized in Table 1.

Fresh-frozen tissue was available for 276 of 415 (67%) patients; 39 (14%) were excluded because the tumor cell percentage was below 30%, 9 (4%) because of low RNA quality, and 1 (0.5%) because the tissue yielded less than 1 μg RNA. Patients from whom viable fresh-frozen tissue was available were on average three years older (P = 0.006), had a 3-mm larger tumor diameter (P = 0.011), and had a higher tumor grade (P = 0.006). Axillary load (P = 0.22) nor patient survival (OS: P = 0.70; DDFS: P = 0.91) were different.

Two-hundred-eighty-three of 415 patients (68%) were indicated for chemotherapy according to the Dutch clinical guidelines and 182 of 415 (44%) had a Nottingham Prognostic Index above 3.4. Of the patients with available tissue, 116 of 227 (51%) had a high-risk 70-gene signature and 107 of 227 (47%) patients had an intermediate or high-risk 21-gene recurrence score. The agreement between the various prognostic models for the identification of high-risk patients was slight to moderate. Slight to fair agreement was observed between the 70-gene signature and the clinicopathologic models (κ = 0.16 for the Dutch guidelines and κ = 0.28 for the Nottingham Prognostic Index). Slight agreement was observed between the 21-gene recurrence score and the clinicopathologic models (κ = 0.10 for the Dutch guidelines and κ = 0.17 for the Nottingham Prognostic Index). Moderate agreement was found between the Dutch guidelines and the Nottingham Prognostic Index (κ = 0.56) and between the 70-gene signature and 21-gene recurrence score (κ = 0.45).

The distributions of CPE in the prognostic models and molecular assays are shown in Fig. 2. Note that patients who died (Fig. 2, triangles) are overrepresented in the subgroups considered to be at high risk according to the prognostic models and having low or intermediate CPE. CPE did not follow a normal distribution (P < 0.001). CPE was not significantly correlated with the 70-gene signature (P = 0.362) or the 21-gene recurrence score (P = 0.469). The distribution of CPE was not significantly different between patients who died of breast cancer–related causes or other causes (P = 0.221).

Results of the survival analysis for OS and DDFS can be found in Tables 2 and 3, respectively. IPW-adjusted Kaplan–Meier curves for OS can be found Fig. 3.

The median time of follow up was 85 months (interquartile range, 69–108), 34 of 415 (8%) died, 42 of 415 (10%) died or had a distant metastasis. None of the propensity scores yielded extreme weights.

Patients with a Nottingham Prognostic Index above 3.4 [OS: HR (95% confidence interval (CI)) = 1.62 (0.81–3.26), P = 0.172, DDFS: HR (95% CI) = 1.39 (0.75–2.58), P = 0.295] or indicated for chemotherapy by the Dutch clinical guidelines [OS: HR (95% CI) = 1.01 (0.48–2.11), P = 0.984, DDFS: HR (95% CI) = 1.06 (0.54–2.08), P = 0.858] did not have a significantly worse survival. Patients with a high-risk molecular assay were significantly associated with OS [70-gene signature: OS: HR (95% CI) = 15.5 (2.07–117), P < 0.001, DDFS: HR (95% CI) = 10.1 (2.36–50) P < 0.001; 21-gene recurrence score: OS: HR (95% CI) = 2.76 (1.09–6.96), P = 0.032, DDFS: HR (95% CI) = 2.22 (0.97–5.08), P = 0.058].

Tertile cutoffs for CPE were 0.40 and 0.54. After adjustment for patient and tumor characteristics, patients in the high CPE group had a significantly better survival than those in the low CPE group (OS: P = 0.040; DDFS: P = 0.042, 10-year cumulative survival OS: 94% vs. 76%, difference = 18%, DDFS: 92% vs. 71%, difference 21%). This was also found for OS after additional adjustment for systemic treatment (P = 0.048); DDFS showed a trend (P = 0.058; 10-year cumulative survival OS: 94% vs. 77%, difference = 17%, DDFS: 92% vs. 72%, difference 20%).

One-hundred-and-eighty-two of 415 (44%) patients had a Nottingham Prognostic Index above 3.4. In this subgroup, 20 of 182 (11%) patients died and 23 of 182 (13%) died or had a distant metastasis. Tertile cutoffs for CPE were 0.39 and 0.54. After adjustment for patient and tumor characteristics, patients in the high CPE group had a significantly better survival than those in the low CPE group (OS: P < 0.001; DDFS: P = 0.001, 10-year cumulative survival OS: 98% vs. 73%, difference = 25%, DDFS: 98% vs. 80%, difference 18%). This was also found after additional adjustment for systemic treatment (OS: P = 0.006; DDFS: P = 0.008, 10-year cumulative survival OS: 98% vs. 74%, difference = 24%, DDFS: 98% vs. 80%, difference 18%).

According to the Dutch clinical guidelines, 283 of 415 (69%) patients were indicated for chemotherapy. In this subgroup, 24 of 283 (8%) died, 30 of 283 (11%) died or had a distant metastasis. Tertile cutoffs for CPE were 0.41 and 0.55. After adjustment for patient and tumor characteristics, patients in the high CPE group had a significantly better survival than those in the low CPE group (OS: P = 0.021; DDFS: P = 0.016, 10-year cumulative survival OS: 97% vs. 79%, difference = 18%, DDFS: 95% vs. 71%, difference 24%). This was also found after additional adjustment for systemic treatment (OS: P = 0.016; DDFS: P = 0.017, 10-year cumulative survival OS: 97% vs. 78%, difference = 19%; DDFS: 96% vs. 71%, difference = 25%).

A molecular assay could be constructed of 227 of 415 (55%) patients. In these patients, 116 of 227 (51%) had a high-risk 70-gene signature of whom 17 of 116 (15%) died and 21 of 116 (18%) died or had a distant metastasis. Tertile cutoffs for CPE were 0.37 and 0.47. After adjustment for patient and tumor characteristics, patients in the high CPE group showed a trend with overall survival when compared with those in the low CPE group (P = 0.054, 10-year cumulative survival OS: 91% vs. 58%, difference = 33%). This was also found after additional adjustment for systemic treatment (P = 0.088, 10-year cumulative survival OS: 90% vs. 59%, difference = 31%). There were no significant differences in DDFS.

Of the patients from whom a molecular assay could be constructed, 107 of 227 (47%) had an intermediate or high-risk 21-gene recurrence score of whom 11 (10%) died and 14 (8%) died or had a distant metastasis. Tertile cutoffs for CPE were 0.37 and 0.49. After adjustment for patient and tumor characteristics, patients in the high CPE group had a significantly better OS than those in the low CPE group (P = 0.030, 10-year cumulative survival OS: 97% vs. 75%, difference = 22%). After additional adjustment for systemic treatment, a trend was found (P = 0.059, 10-year cumulative survival OS: 97% vs. 77%, difference = 20%). There were no significant differences in DDFS.

The group of patients with high CPE (138 patients, 1 event) had significantly less breast cancer–related deaths than the group with low CPE (139 patients, 11 events, P = 0.004) after IPW-adjustment for patient, tumor, and systemic therapy characteristics (Supplementary Table S1). For non-breast cancer–related deaths, no significant association with CPE was found after IPW adjustment (Supplementary Table S1).

The effect of the percentile threshold on the hazard ratio of intermediate and/or high CPE with respect to low CPE was minimal (Supplementary Fig. S1). In the baseline survival analysis, the median HR (interquartile range) after IPW adjustment for patient and tumor characteristics was 0.29 (0.27–0.31) for intermediate CPE and 0.34 (0.21–0.38) for high CPE. The median HR (interquartile range) for high CPE was 0.08 (0.08–0.08) in the group of patients with a Nottingham Prognostic Index above 3.4, and 0.22 (0.12–0.36) in those indicated for chemotherapy according to the Dutch clinical guidelines. In patients with a high-risk 70-gene signature the median hazard ratio (interquartile range) for high CPE was 0.20 (0.19–0.21), and 0.10 (0.09–0.11) in those with an intermediate/high-risk 21-gene recurrence score (Supplementary Fig. S1).

In 415 patients with unilateral invasive ER⁺/HER2⁻ breast cancer, contralateral parenchymal enhancement (CPE) on dynamic contrast-enhanced MRI appears to complement existing prognostic models and molecular assays based on the tumor. In patients considered to be at high risk according to these models and assays, pretreatment CPE was able identify subgroups of patients at relatively low risk. Patients with high CPE had a significantly superior survival compared to patients with low CPE, up to 30% after 10 years.

CPE was significantly associated with overall survival (OS) and DDFS. After adjustment for patient and tumor characteristics as well as systemic treatment, CPE was still significantly associated with survival. In the patients with a Nottingham Prognostic Index above 3.4 or indicated for chemotherapy by the Dutch guidelines, CPE was significantly associated with both OS and DDFS survival. In the patients at high or intermediate risk according to the 70-gene signature and 21-gene recurrence score, CPE was only significantly associated with OS. After adjustment for patient and tumor characteristics as well as systemic treatment, CPE was not able to find a patient population with a significantly better OS in the patients with an intermediate or high-risk 21-gene recurrence score. In this group, the HR of patients in the high CPE group with respect to patients in the low CPE group only slightly increased (from 0.14 to 0.16), but the 95% confidence interval increased (from 0.00–0.84 to 0.00–1.10). The slightly higher P values in the models with the most stringent corrections are not likely because of the additional adjustment for systemic treatment, because this did not differ between low, intermediate, and high CPE groups (P = 0.784). Therefore, this suggests a lack of power in the analysis of the group with an intermediate- and high-risk 21-gene recurrence score as underlying cause.

A recent study in patients without a BRCA mutation showed that age at diagnosis was inversely related to background parenchymal enhancement (BPE), and was the only significant risk factor for development of breast cancer after multivariable analysis (34). Even though we found a comparable negative correlation between CPE and age at diagnosis, CPE was associated with survival after adjustment for age at diagnosis.

Recent results from the MINDACT and TAILORx trials suggest that chemotherapy may be omitted in clinically high-risk patients with a low-risk molecular assay. These results, indicating a potential reduction in overtreatment up to 22% (5, 6), are likely to change clinical practice in the near future. Patients indicated for chemotherapy according to conventional clinical models and a high-risk Mammaprint or an intermediate or high-risk Oncotype DX are likely to receive chemotherapy. Our results indicate that in the subgroups of patients with a high-risk 70-gene signature or an intermediate or high-risk 21-gene recurrence score, CPE might be a promising biomarker for even more patient-specific treatment. We did not evaluate the ability of CPE to identify patients with unfavorable outcomes (i.e., those with low CPE) even if they had low-risk genomic signatures. The reason for this is that recent randomized trials have shown that patients with low-risk genomic signatures have excellent outcome, even without administration of chemotherapy (5, 6). Moreover, as a result of this, the number of events in this group in our dataset is too small to allow meaningful statistical analysis.

Competing risk analysis was performed to explore whether the association of CPE with overall survival primarily occurred in breast cancer–related deaths or if it also exists in non-breast cancer–related deaths. We did not find significant evidence that CPE is related to non-breast cancer–related death. We performed this analysis in all included patients to maximize power; however, the limited number of events makes it difficult to draw hard conclusions. Therefore, competing risk analysis should be performed in future large validation studies.

The relation between CPE and survival was shown to be robust; HRs did not deviate substantially when the percentile threshold in the CPE calculation was altered from the original top-10% threshold. We chose to split CPE in tertiles. This provided the opportunity to assess the effect of incremental CPE on survival (in contrast to e.g., dichotomization), while retaining the highest number of patients per CPE group.

The biological reason why CPE is associated with long-term survival is not yet fully understood. CPE is a measure of the perfusion of the healthy unaffected stroma. A possible explanation for the relation between CPE and survival may be that tumors in patients with high CPE have a higher hormone sensitivity and thus may be more receptive to hormonal therapy. This theory is supported by the previously reported findings that CPE is more strongly associated with survival in patients receiving hormonal therapy, and that CPE does not stratify survival in patients with a triple-negative tumor (17). Hormone sensitivity cannot be the only explanation of our findings, because ER and progesterone receptor-positivity do not show such strong association with survival as CPE. Enhancement of stroma may also be indicative of effective drug transport or an effective immune response. To test these hypotheses, future research will address several aspects such as the correlation of the level of immune infiltration in the primary tumor with CPE.

Research on stromal enhancement includes background parenchymal enhancement (BPE) and signal enhancement ratio (SER) and focuses on diagnosis, prognostication, and response monitoring. BPE refers to the volume and intensity that normal stroma enhances on MRI after injection of contrast agent and is typically scored manually by radiologists in four incremental categories (35). SER is a quantitative measure defined as the signal ratio between early and late subtraction scans (36). Increased BPE has been associated with decreased diagnostic performance (37, 38). BPE has also been studied in relation to the risk of developing breast cancer (8, 9, 34, 39–43). In general, a higher risk of breast cancer is reported with increased BPE (8, 9, 39, 40, 43). Patients with increased BPE have an increased risk of tumors with positive resection margins (44). More pronounced stromal enhancement surrounding the tumor is related to extensive ductal carcinoma in situ (45). Increased SER around tumor leads to a higher local recurrence rate (14, 46).

BPE generally decreases after neoadjuvant chemotherapy (47–49) and is related to tumor response (12). A high BPE before neoadjuvant chemotherapy has been related to a pathologic complete response (13) as well as to an inferior survival (15). Increased enhancement surrounding the tumor prior to treatment (11) and after one cycle of chemotherapy (10) have been associated with longer disease-free survival. The current study did not asses CPE in neoadjuvant therapy setting.

This study has some limitations. First, the study uses the same patient population in which the CPE biomarker was developed. In the current study, new data were added by analyzing the fresh-frozen tumor tissue and generating the molecular assays. Furthermore, clinicopathologic markers were modeled using two common prognostic models. CPE has shown consistency by adding value to these models and assays, yielding additional evidence to justify validation of CPE in an independent large follow-up validation study.

Second, we reproduced the molecular assays from the RNA of the tumors; they are not the official Mammaprint and Oncotype DX tests. A benefit of this approach is that we used one and the same tissue sample for the creation of both assays. Therefore, both gene signatures represent exactly the same part of the tumor, preventing potential bias due to tumor heterogeneity (50). This approach also provided us with the continuous 70-gene signature and 21-gene recurrence score values, allowing detailed correlation with CPE without loss of power due to dichotomization. A potential disadvantage of reproducing the assays ourselves is the risk that the assays might not reflect the same score when ordered. However, the reproduced assays behave as expected in relation to survival. A lower 70-gene signature is associated with poorer patient survival, similar to the findings described by van't Veer and colleagues (3). Likewise, a higher 21-gene recurrence score is related to poorer patient survival, as described in the publication of Paik and colleagues (4).

Finally, MR imaging was not timed on the menstrual window. Delaying the MRI to time for the menstrual window would lead to an undesired delay of surgery. This limitation was, however, present in patients who had an event and in those who had not. Therefore we considered it unlikely that this limitation biased our results.

In conclusion, in patients at high risk according to conventional prognostic models or molecular assays that are based on the tumor, subgroups of patients at low risk were identified using pretreatment enhancement of the normal stroma in the contralateral breast. Patients with high enhancement had little absolute risk at 10 years, with survival benefits between 18% and 30%. This shows contralateral parenchymal enhancement to be a promising biomarker towards even more patient-specific treatment.

No potential conflicts of interest were disclosed.

Conception and design: B.H.M. van der Velden, M.A. Viergever, K.G.A. Gilhuijs

Development of methodology: B.H.M. van der Velden

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): T. Bismeijer, C.E. Loo, K.G.A. Gilhuijs

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): B.H.M. van der Velden, S.G. Elias, T. Bismeijer, L.F.A. Wessels, K.G.A. Gilhuijs

Writing, review, and/or revision of the manuscript: B.H.M. van der Velden, S.G. Elias, C.E. Loo, M.A. Viergever, L.F.A. Wessels, K.G.A. Gilhuijs

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): B.H.M. van der Velden

Study supervision: B.H.M. van der Velden, S.G. Elias, M.A. Viergever, K.G.A. Gilhuijs

We acknowledge J. Wesseling, E. Lips, F. Nieboer, L. Mulder, the NKI-AVL Core facility Molecular pathology & biobanking (CFMPB) as well as the Genomics Core Facility for tissue processing and RNAseq.

This research is part of the STW Perspectief Population Imaging Genetics (ImaGene) program and supported by the Dutch Technology Foundation STW, which is part of the Netherlands Organisation for Scientific Research (NWO), and partly funded by Ministry of Economic Affairs.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.