We are grateful to all of the patients who were included in this study and to the Translational Breast Cancer Research Consortium investigators and data managers for their efforts. We are very appreciative of funding support from the Translational Breast Cancer Research Consortium from the AVON Foundation, the Breast Cancer Research Foundation, and Susan G. Komen for the Cure. We gratefully acknowledge the American College of Radiology Imaging Network (ACRIN) for granting permission to include patients who were treated on the ACRIN 6657/I-SPY Trial (supported by National Cancer Institute grants U01 CA079778 and U01 CA080098).
Increased pathologic complete response (pCR) rates observed with neoadjuvant chemotherapy (NCT) for some subsets of patients with invasive breast cancer have prompted interest in whether patients who achieved a pCR can be identified preoperatively and potentially spared the morbidity of surgery. The objective of this multicenter, retrospective study was to estimate the accuracy of preoperative magnetic resonance imaging (MRI) in predicting a pCR in the breast.
MRI studies at baseline and after the completion of NCT plus data regarding pathologic response were collected retrospectively from 746 women who received treatment at 8 institutions between 2002 and 2011. Tumors were characterized by immunohistochemical phenotype into 4 categories based on receptor expression: hormone (estrogen and progesterone) receptor (HR)-positive/human epidermal growth factor receptor 2 (HER2)-negative (n = 327), HR-positive/HER2-positive, (n = 148), HR-negative/HER2-positive, (n = 101), and triple-negative (HR-negative/HER2 negative; n = 155). In all, 194 of 249 patients (78%) with HER2-positive tumors received trastuzumab. Univariate and multivariate analyses of factors associated with radiographic complete response (rCR) and pCR were performed.
For the total group, the rCR and pCR rates were 182 of 746 patients (24%) and 179 of 746 patients (24%), respectively, and the highest pCR rate was observed for the triple-negative subtype (57 of 155 patients; 37%) and the HER2-positive subtype (38 of 101 patients; 38%). The overall accuracy of MRI for predicting pCR was 74%. The variables sensitivity, negative predictive value, positive predictive value, and accuracy differed significantly among tumor subtypes, and the greatest negative predictive value was observed in the triple-negative (60%) and HER2-positive (62%) subtypes.
Neoadjuvant chemotherapy (NCT) is used increasingly for the treatment of invasive, high-risk breast cancer. The use of effective first-line chemotherapeutic agents as well as targeted therapies, such as trastuzumab for human epidermal growth factor receptor 2 (HER2)-positive disease in the neoadjuvant setting have resulted in a high rate of pathologic complete response (pCR) ranging from 40% to 67%, depending on the study population.1-4 It remains unknown whether excision of the tumor bed in the setting of pCR improves the risk of locoregional recurrence; thus, there has been keen interest in determining whether negative imaging after systemic therapy might identify a patient subset that could be treated safely with radiation alone without surgery. Prior studies evaluating radiotherapy as the definitive modality for treating the breast in patients who have a clinical complete response to NCT have resulted in unacceptably high locoregional failure rates.5-7 This may be attributed largely to poor patient selection, because it is known that clinical examination to detect residual disease or response is limited.8-10 A more sensitive tool for evaluating in-breast response could more effectively identify potential candidates for radiotherapy alone after a complete response to NCT.
One such tool is magnetic resonance imaging (MRI), which has been used with increasing frequency in recent years because of its high sensitivity for detecting breast cancers compared with mammography or ultrasound.11-14 The ability of a radiologic complete response (rCR) by MRI to predict a pCR has been the subject of active investigation. Several groups have evaluated the predictive accuracy of MRI for assessing response to NCT and reported limited correlation between an rCR after the completion of NCT and a pCR.13, 15-17 Although the performance of MRI appeared better in some tumor phenotypes compared with others, previous small, retrospective studies lacked the power to detect a significant difference between clinically distinct subsets. In the current study, we sought to determine the performance of MRI after NCT in a larger multicenter data set to better define the accuracy of post-treatment breast MRI in the prediction of pCR. Moreover, we wanted to identify which tumor-related variables were associated with the highest correlation between an rCR and a pCR to identify a patient population that may be most amenable to treatment with whole breast radiation without surgery based on demonstrating an rCR.
MATERIALS AND METHODS
Patients who received neoadjuvant therapy before and after breast MRI were retrospectively identified at 8 National Cancer Institute (NCI)-designated comprehensive cancer centers. Between January 2002 and February 2011, 770 women fulfilled study criteria. The identification of patients across institutions was stipulated by cross-referencing billing codes of patients who had a breast cancer diagnosis, received treatment with chemotherapy, and underwent imaging studies with breast MRI. Among this group were 165 patients who were treated on the I-SPY trials, which are prospective multicenter trials of women who receive NCT, including radiographic and pathologic endpoints. All participating institutions were members of the Translational Breast Cancer Research Consortium (TBCRC), which sponsored the study, and included the University of Alabama at Birmingham, the University of Pittsburgh Medical Center, Dana-Farber Cancer Institute, The University of Texas M. D. Anderson Cancer Center, Duke University, the University of Chicago, the University of North Carolina Chapel Hill, and the University of California-San Francisco. Institutional review board approval for the study was obtained at each institution. In addition to pre-NCT and post-NCT MRI studies, eligible patients were required to have undergone definitive surgery with pathology available for review. Patient, tumor, and treatment-related variables were entered into a secure, password-protected online database. Documentation of baseline and post-treatment images with mammography and ultrasound of the breast and lymph nodes also were recorded.
Histologic tumor types were recorded as follows: invasive ductal carcinoma, pure invasive lobular carcinoma, invasive mammary carcinoma with ductal and lobular features, and invasive mammary carcinoma not otherwise specified. Estrogen receptor (ER) and progesterone receptor (PR) status (positive or negative, with positive defined as ≥1% positive cells) according to percentage of positive cells and HER2 status (positive or negative) also were collected. HER2 status was determined by local testing according to the 2007 American Society of Clinical Oncology/College of American Pathologists guidelines.18 All histopathology and biomarker assessments were performed at the individual sites. A pathologic response in the breast was categorized as no residual invasive disease or ductal carcinoma in situ (DCIS); no residual invasive cancer with DCIS present; and residual invasive disease, including microscopic residual invasive disease.19 To assess the primary endpoints of sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of MRI, a pCR in the breast was defined as the resolution of both invasive disease and DCIS.
All patients were required to have both baseline and post-treatment MRI studies to be eligible for the study. At the time of presentation, all patients in the cohort had an enhancing lesion on MRI corresponding to the known biopsy-proven cancer. The baseline size of the lesion defined as the maximal diameter in a single dimension on pretreatment MRI, mammogram, and ultrasound, was recorded. Radiographic tumor (T) classification was assigned based on the largest imaging size by any modality. Patients who had unknown primary lesions were excluded. Specific parameters for dynamic contrast-enhanced breast imaging were not defined for eligibility; however, the institutions included on this study have high levels of expertise in breast MRI. A central review of MR images was not performed. A complete MRI response in the breast was defined as the resolution of all areas of abnormal enhancement, mass, or distortion.
All patients in the study received preoperative systemic therapy. The number of cycles of chemotherapy before post-treatment MRI was recorded. It was also noted whether patients received trastuzumab, bevacizumab, or neoadjuvant hormone therapy.
The sensitivity, specificity, PPV, NPV, and accuracy of MRI for detecting residual disease in the breast were estimated from the data with pCR defined, as discussed above, as the complete resolution of both invasive cancer and DCIS. True-negative (TN) was defined as negative on both MRI and pathology; false-negative (FN) was defined as negative on MRI but positive on pathology; true-positive (TP) was defined as positive on both MRI and pathology; and false-positive (FP) was defined as positive on pathology but negative on MRI. Accuracy was defined as the percentage of test results correctly identified by the test, ie, (TP + TN)/total test results = (TP + TN)/(TP + TN + FP + FN). The accuracy of mammography and MRI was defined separately for each individual modality (Fig. 1). The combined accuracy of both modalities was scored as accurate if both modalities predicted a pCR.
Comparisons of patient subsets are based on the chi-square test for contingency tables. Multivariable logistic regression was used to examine the simultaneous effects of multiple factors. The SAS statistical software package was used for all analyses (version 9.2; SAS Institute, Inc., Cary, NC).
Data from 770 women who received treatment at 8 tertiary NCI comprehensive cancer centers were collected. After excluding patients who had missing data (MRI response after NCT or final pathology), 746 patients remained evaluable for the primary endpoint. The median age of the cohort was 49 years (range, 20-86 years). Histologic subtype rendered on initial core biopsy was generally representative of newly diagnosed cancers: 61% were ER-positive, 54% were PR-positive, and 34% were HER2-positive. A summary of patient and tumor characteristics is provided in Table 1.
Table 1. Patient Characteristics and In-Breast Radiographic and Pathologic Response
The most common chemotherapy regimen delivered included doxorubicin, cyclophosphamide, and a taxane (51%); other regimens (35%); taxane and carboplatin (9%); and doxorubicin and cyclophosphamide (5%). Five percent of patients received bevacizumab, and 18 patients received neoadjuvant hormone therapy only. Of 248 patients with HER2-positive breast tumors, 194 (78%) received trastuzumab. The remaining 54 patients with HER2-positive breast tumors were treated before 2005 and did not receive trastuzumab.
The median interval between post-treatment scans and surgery was 20 days. Definitive surgery was mastectomy with or without reconstruction in 54% of women and breast conservation in 46%. Nineteen percent underwent sentinel lymph node biopsy before starting systemic treatment, 42% underwent sentinel lymph node biopsy as part of definitive surgery, and 72% underwent level I or II axillary lymph node dissection.
Performance of Breast Magnetic Resonance Imaging
Overall, post-treatment MRI detected residual disease in the breast with 83% sensitivity, 47% specificity, a PPV of 47%, an NPV of 83%, and 74% accuracy. Table 2 also illustrates the performance of MRI of the breast stratified according to immunohistochemical phenotypes. There were significant differences in sensitivity, NPV, PPV, and accuracy among subtypes. The NPV was highest for patients who had HR-negative/HER2-positive and triple-negative breast cancers. HER2-positive breast tumors differed according to receipt of trastuzumab (Table 3). The addition of trastuzumab in HER2-positive breast tumors significantly increased the rate of rCR compared with no trastuzumab (32.5% vs 15%, respectively; P < .0001) and was reflected in the increased rate of pCR (33.5% vs 18.5%, respectively; P = .01). However, this did not translate into a difference in the accuracy of MRI for predicting a pCR among patients.
Table 2. Comparison of Post-Treatment Breast Magnetic Resonance Imaging Performance Between Tumor Subtypes
P values were calculated by comparing the 4 molecular subtypes for each performance measure (sensitivity, specificity, NPV, PPV, and accuracy) with magnetic resonance breast imaging using the chi-square test.
Table 3. Comparison of Post-Treatment Breast Magnetic Resonance Imaging Performance for Pathologic Complete Response in Patients With Human Epidermal Growth Factor Receptor-Positive Tumors Treated With and Without Trastuzumaba
A pathologic complete response was defined as resolution of invasive disease and ductal carcinoma in situ.
Receipt of trastuzumab was unknown in 1 patient who had HR+/HER2+ breast cancer.
As a secondary aim, we also assessed whether a complete response on post-NCT mammogram or ultrasound studies (when available) in combination with MRI increased the accuracy for predicting a pCR. The accuracy of mammogram and MRI, separately and in combination, for predicting a pCR is illustrated in Figure 1. A complete response on both mammography and MRI did not significantly increase the ability to detect patients who would achieve a pCR over MRI alone; although, as expected, an rCR by mammography (P < .0001) or ultrasound (P < .0001) also was associated significantly with an rCR on MRI (data not shown). The small number of patients who underwent imaging after treatment with all 3 modalities (MRI, mammography, and ultrasound) precluded a formal evaluation of trimodality performance.
Covariates Associated With Radiologic and Pathologic Complete Responses
In the total group, 182 of 746 patients (24%) achieved an rCR, and 179 of 746 patients (24%) achieved a pCR. Tumor characteristics that were significantly associated with both an rCR and a pCR on MRI included lower radiographic baseline T classification, high grade, negative ER and/or PR status, and HR-negative/HER2-positive or triple-negative immunohistochemical phenotype (Table 1). On univariate analysis, race was associated with pCR but not with rCR. HER2-positive breast tumors had a trend toward an increased rCR rate (P = .05). Treatment regimens that were associated with the highest rates of rCR included taxane (either paclitaxel or docetaxel)/cyclophosphamide chemotherapy (P < .0001) and the receipt of trastuzumab-containing regimens among women with HER2-positive breast tumors (P = .0001). Tumor size on pretreatment MRI was also evaluated as a continuous variable to determine whether a significant cutoff for rCR could be determined. Patients who had tumors ≤5 cm in greatest dimension on a y pretreatment MRI had a greater chance of obtaining an rCR (P = .09) and for achieving a pCR in those who achieved an rCR (P = .06), although these findings did not reach statistical significance (Fig. 2).
Multivariate logistic regression analysis for variables that significantly influenced rCR and pCR was performed with a final model that included baseline T classification, age, race, and tumor subtype. Variables that were associated independently with rCR included triple-negative or HER2-positive tumor subtype (P = .03) and lower T classification at presentation (P = .002). Variables that were associated independently with the likelihood of attaining a pCR were triple-negative or HER2-positive tumor subtype (P < .0001), lower T classification (P = .01), and African American race (P = .0003).
The increasingly high rate of pCR in some subsets of patients who receive NCT for invasive breast cancer has raised the question of whether patients who achieve a pCR require surgical resection of the tumor bed or whether the breast can be appropriately managed with radiation therapy alone.17 However, the safe omission of surgery in patients who receive neoadjuvant therapy and achieve an rCR depends on the ability to accurately estimate pCR preoperatively. MRI has an increased sensitivity for detecting residual disease in the breast compared with either mammography or ultrasound, making it a potentially useful tool in this setting.20-25 The performance of MRI in predicting a pCR in women with breast cancer who receive NCT has been the topic of several publications over the past several years.12, 14, 15, 17, 26-36 Those trials mostly included small single-institution studies, which reported that post-treatment MRI correlated more closely with final pathology than with physical examination, mammography, or ultrasound. However, a broad range of predictive accuracy (58%-93%) was reported. One of the best characterized studies is the I-SPY 1 trial,29-31, 37 in which the investigators observed that, in 216 evaluable patients, MRI findings were a better indicator of a pathologic response to NCT than clinical assessment.30 However, the I-SPY 1 trial was underpowered to detect significant differences between subsets.
The objective of the current analysis was to inform the feasibility of using MRI in-breast response in a prospective trial, omitting breast surgery and using radiotherapy alone, in patients who achieved an rCR after NCT. Because the omission of surgery in patients who have an rCR would not be considered feasible if residual DCIS remained in the breast, we defined a pCR as the resolution of both invasive disease and DCIS. This definition of pCR has also been associated with an improvement in disease-free survival compared with patients who have resolution of invasive disease but with residual DCIS in the breast after NCT.19 We also wanted to determine whether there were differences in the accuracy of MRI between tumor subtypes.
We observed a statistically significant difference in MRI sensitivity, NPV, PPV, and accuracy for detecting a pCR among different immunohistochemical phenotypes when pCR was defined according to either the standard definition of resolution of invasive disease only (data not shown) or more rigorously as the resolution of both invasive disease and DCIS. Previous trials have suggested differences in the performance of MRI among different immunohistochemical phenotypes.15, 31, 33, 38 We confirmed that, among patients who achieved an rCR, positive HR status and low tumor grade were most commonly associated with residual disease at surgery, suggesting that an rCR on preoperative MRI in these patient populations should be interpreted with caution. However, it is important to note that differences in NPV, PPV, and accuracy among phenotype subsets are impacted by variations in the prevalence of pCR among subtypes. Similarly, comparisons indicating significant differences among variables for rCR are reflective of disease sensitivity to NCT rather than differing MR assessment quality across these subtypes. Although it is retrospective, to our knowledge, the current study represents the single largest analysis to date addressing this question.
There are several important limitations to note. Our study did not include either a central radiology review or a central pathology review. However, data collection was performed across 8 NCI comprehensive cancer centers, each with substantial expertise in MRI techniques and interpretation as well as significant expertise in surgical breast pathology. Our goal was to create a data set that was generalizable across different situations and institutions, and the endpoints chosen were objective and unambiguous, minimizing potential for subjective bias. In addition, we did not collect data on markers of proliferation, such as Ki-67, that would allow further characterization of ER-positive tumors into luminal A and B subtypes and would facilitate analyses of differences between these subsets in imaging accuracy. Finally, the specificities of the MR coil strength, contrast type, and infusion protocol for every institution were not collected in this retrospective study; and, indeed, they changed over time at each institution. Although these limitations should be acknowledged, they nevertheless allow for greater generalizability of our findings across many different clinical settings.
In this study, the addition of mammographic rCR to MRI in a subset of patients did not improve the ability to detect residual microscopic disease. The role of ultrasound could not be evaluated, because few patients in the cohort underwent post-treatment ultrasound in addition to MRI. Although our findings emphasize that an rCR based on MRI alone is not sufficient to reliably rule out residual disease, it nevertheless remains a useful and sensitive imaging modality in this setting, particularly among certain breast cancer subtypes.
In conclusion, the accuracy of breast MRI was 74% for predicting a pCR in patients who received preoperative chemotherapy for invasive breast cancer. Sensitivity, NPV, PPV, and accuracy of MRI for predicting a pCR differed significantly among molecular breast cancer subtypes, and the highest NPV was among the HR-negative/HER2-positive (62%) and triple-negative (60%) phenotypes. The observed NPV of MRI overall after neoadjuvant systemic therapy does not support using MRI alone to accurately identify patients who may be candidates for a study of radiation alone in the context of a clinical trial. However, the high sensitivity of MRI, particularly in HR-negative and HER2-positive phenotypes, supports its potential role in a prospective trial to evaluate the omission of surgery in women who achieve an rCR after NCT, possibly in combination with other tests, such as tumor bed biopsy.
We are grateful to the TBCRC investigators and research coordinators for their efforts. We are very appreciative of funding support to the TBCRC from The AVON Foundation, The Breast Cancer Research Foundation, and Susan G. Komen for the Cure. The American College of Radiology Imaging Network (ACRIN) granted permission to include patients who were treated on the ACRIN 6657/I-SPY Trial, which was supported by National Cancer Institute grants U01 CA079778 and U01 CA080098.