Pretreatment prediction of pathologic complete response to neoadjuvant chemotherapy in breast cancer: Perfusion metrics of dynamic contrast enhanced MRI

The purpose of this study was to investigate imaging parameters predicting pathologic complete response (pCR) in pretreatment dynamic contrast enhanced-magnetic resonance imaging (DCE-MRI) in breast cancer patients who were treated with neoadjuvant chemotherapy (NAC). A total of 74 patients who received NAC followed by surgery were retrospectively reviewed. All patients underwent breast MRI before NAC. Perfusion parameters including Ktrans, Kep and Ve of tumor were measured three-dimensionally. These perfusion parameters of background parenchyma of contralateral breasts were analyzed two-dimensionally. Receiver-operating characteristic (ROC) analysis and multivariable logistic regression analysis were performed to compare the ability of perfusion parameters to predict pCR. Of 74 patients, 13 achieved pCR in final pathology. The fiftieth percentile and skewness of each perfusion parameter – Ktrans, Kep, and Ve of tumor were associated with pCR. Perfusion parameters of contralateral breast parenchyma in 2D analysis also showed predictive ability for pCR. The model combining perfusion parameters of contralateral breast background parenchyma and those of the tumor had higher predictive value than each single parameter. Thus, perfusion parameters of tumor, background parenchyma of contralateral breast and their combinations in pretreatment breast MRI allow early prediction for pCR of breast cancer.

Most previous studies have focused on changes in imaging parameters before and after treatment. Although it is important to know such changes, identifying the factor that can predict pCR with pretreatment images will truly allow early prediction of pCR. Early and accurate prediction of tumor response would enable proper choice of treatment regimen and proper operative planning by the surgeon.
We investigated imaging parameters for early prediction of pCR with pretreatment DCE-MRI, focusing on tumor perfusion parameters and background parenchymal enhancement of contralateral breast parenchyma.

Patients.
A total of 74 breast cancer patients who underwent pretreatment DCE-MRI with surgery after NAC were included. After completion of NAC, 22 patients underwent breast conservative surgery (BCS) and 52 patients underwent mastectomy. Immunohistochemical staining of final surgical specimens revealed luminal type in 47 cases, HER-enriched type in 14 cases, and triple-negative type in 13 cases.
Thirteen (17.5%) of these 74 patients achieved pCR based on specimens obtained from surgery after NAC. Of these 13 patients, seven were confirmed to be luminal type while six as HER2+ type. There was no triple negative type. Cancer staging was classified according to AJCC 7th. Stage 0 was the most common (nine of 13 patients). The remaining four patients had lymphovascular invasion in final pathology. They were classified as stage I (n = 1), II (n = 2), and III (n = 1) ( Table 1).

Predictive ability of perfusion parameters. Tumor and background parenchyma of contralateral
breast. Each perfusion parameter of DCE-MRI in tumor and background parenchyma of contralateral breast (BP CL ) did not show high predictive ability for pCR based on receiver operating characteristic (ROC) curve analysis (AUC < 0.7, 0.449 to 0.683) ( Table 2). However, mean Ktrans value of BP CL (AUC = 0.683) had relatively better predictive ability than other single perfusion parameters.
Combination of parameters from tumor and background parenchyma of contralateral breast. Parameters with relatively high AUC values in 3D histogram analysis were extracted, including 50th percentile values and skewness of Ktrans, Kep, and Ve values. These parameters were combined with perfusion parameters of BP CL .
The model combining perfusion parameters of BP CL and those of tumor showed higher predictive value than each single parameter. The combination of Ve of BP CL with 50th percentile and the skewness of Ve in tumor had the highest predictive power for pCR (AUC = 0.807, p = 0.002). The AUC value of a combination of Ktrans of BP CL and skewness of Ktrans in tumor was 0.760 (p = 0.003), and that of a combination of all three parameters -Ktrans of BP CL , 50th percentile Ktrans in tumor, and skewness of Ktrans in tumor -was 0.757 (p = 0.004). Those showed higher predictive ability for pCR than each single parameter. Other combinations of parameters also showed high predictive power with AUC > 0.7, although they were not statistically higher than each single parameters. These included a combination of Ktrans of BP CL and 50th percentile Ktrans of tumor (AUC = 0.731, p = 0.092), and a combination of Ve of BP CL and skewness of Ve in (AUC = 0.718, p = 0.061) ( Table 3).
Univariate and multivariable logistic regression analyses were performed to investigate the correlation between MRI perfusion parameters and pCR. All individual perfusion parameters were cateogrized as 0 or 1, based on optimal cutoff values obtained through ROC analysis. Combination models were scored by the sum of categorized single perfusion parameters (Table 4). In multivariable logistic regression analysis, odds ratio (OR) of each single perfusion parameter was not associated with pCR, except Ktrans of BP CL (OR = 0.01, 95% CI: < 0.001-0.55, p = 0.023) ( Table 5). However, combination models showed much higher association with pCR than single parameters, with higher OR and statistical significance.  with a combination of all three parameters of Ve also showed a high OR (5.26, 95% CI: 1.89-14.64, p = 0.002) ( Table 6).

Discussion
In locally advanced breast cancer, NAC has been suggested as a means to improve prognosis. Patients who acquired pCR after NAC can expect better DFS or OS compared to patients with non-complete response 14 . Many studies have investigated imaging parameters of MRI with the aim to predict pCR in breast cancer patients who undergo NAC. Using conventional breast MRI, initial tumor size, tumor size reduction, and ADC value have been suggested as predictive factors of pCR in pre-and post-treatment MRI, after four cycles of NAC 15 .   However, conventional breast MRI cannot reflect neoangiogenesis of the tumor, which is one of the most important prognostic factors of breast cancer. Tumor neoangiogenesis reflects dissemination of malignant cells as well as the level of contrast enhancement. Imaging parameters capable of revealing vascular kinetics of tumors are important to monitor effects of chemotherapy and predict tumor prognosis 16,17 . DCE-MRI allows quantitative measurements of kinetic parameters related to perfusion and permeability of tumor. Studies using DCE-MRI have investigated imaging parameters with the goal to predict tumor response and prognosis. Martin D. Pickles et al. have suggested that patients with high perfusion and vascular permeability in pretreatment DCE-MRI are significantly less likely to have long-term DFS and OS as a result of neovascularization 18 . In a subsequent study, the authors also suggested that quantification of perfusion parameters could provide information to distinguish responders from non-responders before finishing the treatment, by demonstrating that early treatment change was possible for non-responders 19 . However, these studies focused on changes of MRI imaging parameters before and after NAC as a means to predict pCR. Post-treatment MRI must be performed for prediction, which is expensive. In addition, the predictive factors presented in these studies could not be considered as true "predictive" factors because they were determined as predictors after at least one cycle of chemotherapy.
Our data demonstrate that multiple perfusion parameters obtained from pretreatment DCE-MRI can predict pCR, especially when single parameters are combined. Perfusion parameter of BP CL was explored using 2D analysis and tumor perfusion parameters were assessed by 3D analysis. Perfusion parameter Ktrans of BP CL displayed the highest AUC (0.683) in 2D analysis, while the skewness of Ktrans displayed the highest AUC value (0.647) in 3D analysis of tumor. The AUC value of these combined perfusion parameters had a higher predictive ablity for pCR than each single parameter by ROC analysis. Multivariable logistic regression analysis of perfusion parameters that were combined after categorization revealed much higher association with pCR than each single perfusion parameter showing increased OR and statistical significance.
The present results indicate that perfusion parameters of BP CL could be a predictive factor for pCR. The relationship between BPE and tumor response has already been studied with a focus on qualitative or quantitative BPE reduction before and after NAC 20,21 . Higher BP CL is associated with worse prognosis, such as early tumor recurrence 12,22 . However, since these studies involved qualitative categorization of BPE rather than quantitative measurement, there might be a limit to use of BPE as an objective index for predicting tumor response. We  used perfusion parameters derived from pretreatment DCE-MRI to investigate the relationship between tumor response and BP CL . Each single perfusion parameter of BP CL showed fair ability to predict pCR, with a higher predictive ability when tumor perfusion parameters are combined. This study has some limitations. First, it was a retrospective study with a small number of patients. The number of patients with pCR was much smaller than that of patients with non-pCR. Therefore, our results might not be generalizable. Second, patients received pretreatment MRI regardless of their menstrual cycle, because our study was a retrospective study. This would not be a limitation in postmenopausal patients. However it could affect perfusion parameters of BPE in contralateral breast parenchyma in premenopausal patients. Third, we excluded bilateral breast cancer, because it would be difficult to measure perfusion parameters of the contralateral breast parenchyma.
We investigated the correlation between pretreatment MRI parameters and postoperative final pathology without long-term follow-up after surgery. Thus, DFS and OS, including early or late tumor recurrence which could be correlated with pCR after NAC, were not investigated. Additional study will be needed to investigate imaging parameters in pretreatment MRI to predict DFS and OS after NAC.
In conclusion, early predictive factors of pCR in breast cancer patients using pretreatment MRI only, were explored. Each single perfusion parameters of tumor and BP CL showed fair predictive power for pCR of breast cancer. However, the combination of both perfusion parameters of tumor and BP CL showed higher predictive power for pCR of breast cancer.

Material and Methods
Patients. This study was approved by the Institutional Review Board of Seoul St. Mary's Hospital. It was performed in accordanace with procedures complied with HIPAA guidelines. The requirement for informed consent was waived due to its retrospective nature. A total of 294 breast cancer patients who underwent DCE-MRI between February 2014 and May 2016 were included. Of them, 214 patients who were not candidates of NAC before surgery were excluded. Two patients were also excluded from the remaining 80 patients because they had not received breast surgery. Of the remaining 78 patients, DCE-MRI raw data for postprocessing analysis were lost for four patients. Finally, 74 patients were enrolled (Fig. 1).

NAC.
All patients underwent pretreatment MRI before the first cycle of NAC. All 74 patients received anthracycline-based chemotherapy. Twenty-nine patients were treated with a combination of anthracycline and cyclophosphamide, of which seven underwent additional Herceptin use. Forty-five patients were treated with the combination of anthracycline and taxane, of which seven underwent additional Herceptin treatment. After completion of NAC, all patients underwent surgery including breast conserving surgery or total mastectomy. Pathology. Pathologic complete response (pCR) was defined as the absence of any invasive component, or residual ductal carcinoma in situ (DCIS) component in the specimen obtained from surgery 23,24 . A non-pCR was defined as the presence of microscopic invasive tumor in the final pathology 23 . All pathological information including immunohistochemical staining was obtained from surgically obtained specimens. However, immunohistochemical staining results were obtained from biopsy samples for pCR patients who did not even have a residual DCIS component in surgical specimen.     Imaging analysis. To evaluate perfusion parameters from DCE-MRI, a standard Tofts model was used.
Parameters such as volume transfer rate (Ktrans [min −1 ]), volume of extravascular extracellular space (EES) per unit volume of tissue (Ve), and flux rate constant between EES and plasma (Kep, [min −1 ]) 25 were calculated using postprocessing software (Olea Sphere, version 3.0, Olea Medical, La Ciotat, France). In Olea Sphere, it was chosen to derive the arterial input function (AIF) from perfusion weighted image pixels selected in ascending aorta, and the AIF was calculated automatically. Analysis of perfusion parameter was performed for the tumor itself, and the background parenchyma of the breast on the contralateral side without tumor involvement.
Tumor. Perfusion parameters of the tumor were obtained by 3D histogram analysis. In the analysis, after the whole outline of tumor was drawn manually, magic wand, a semi-automated region-growing segmentation tool was used to find and analyze the enhancing portion of the tumor (Fig. 2a,b). The 25th, 50th, 75th percentile, mean value, and skewness and kurtosis of each Ktrans, Kep, and Ve were then obtained.
Background parenchyma of contralateral breast. Perfusion parameters of BP CL were also obtained by twodimensional (2D) analysis. Mean values of Ktrans, Kep, and Ve were derived. These parameters were measured by drawing a eliptical region of interest (ROI) at the level showing the largest fibroglandular tissue without enhancing lesions, except for the skin and fat layer (Fig. 3).

Statistical analyses.
To test the predictive ability of each single perfusion parameter for pCR based on DCE-MRI, ROC curve analysis was performed. Optimal cut-off values, sensitivities, and specificities of each parameter were obtained from the Youden index on the ROC curve.
A multivariate logistic regression model combining perfusion parameters of BP CL and those of the tumor were fitted and their performance was evaluated by area under the curve (AUC). Univariate and multivariate logistic regression models were used to calculate OR of converting perfusion parameters with optimal cutoff. Additionally, univariate logistic regression was perfomed to gauge the effect of combined perfusion parameters (Table 4). A Bonferroni correction was applied for multiple comparisons, with a correction factor derived from the number of perfusion parameters. For comparison of total 30 parameters -21 parameters of each single perfusion parameters ( Table 2) and 9 combining perfusion parameters (Table 3), critical p value was calculated as 0.0017 from 0.05/30. All statistical analyses were performed using SAS ver. 8.4 software (SAS Institute Inc., Cary, NC, USA). A p value < 0.05 was considered statistically significant.