Preoperative dynamic breast magnetic resonance imaging kinetic features using computer-aided diagnosis: Association with survival outcome and tumor aggressiveness in patients with invasive breast cancer

Objectives To evaluate whether preoperative breast dynamic contrast-enhanced (DCE) magnetic resonance (MR) imaging kinetic features, assessed using computer-aided diagnosis (CAD), can predict survival outcome and tumor aggressiveness in patients with invasive breast cancer. Materials and methods Between March and December 2011, 301 women who underwent preoperative DCE MR imaging for invasive breast cancer, with CAD data, were identified. All MR images were retrospectively evaluated using a commercially available CAD system. The following kinetic parameters were prospectively recorded for each lesion: initial peak enhancement, the proportion of early phase medium and rapid enhancement, and the proportion of delayed phase persistent, plateau, and washout enhancement. The Cox proportional hazards model was used to determine the association between the kinetic features assessed by CAD and disease-free survival (DFS). The peak signal intensity and kinetic enhancement profiles were compared with the clinical-pathological variables. Results There were 32 recurrences during a mean follow-up time of 55.2 months (range, 5–72 months). Multivariate analysis revealed that a higher peak enhancement (DFS hazard ratio, 1.004 [95% confidence interval (CI): 1.001, 1.006]; P = .013) on DCE MR imaging and a triple-negative subtype (DFS hazard ratio, 21.060 [95% CI: 2.675, 165.780]; P = .004) were associated with a poorer DFS. Higher peak enhancement was significantly associated with a higher tumor stage, clinical stage, and histologic grade. Conclusions Patients with breast cancer who showed higher CAD-derived peak enhancement on breast MR imaging had worse DFS. Peak enhancement and volumetric analysis of kinetic patterns were useful for predicting tumor aggressiveness.

Introduction Dynamic contrast-enhanced (DCE) magnetic resonance (MR) imaging is the most sensitive modality for breast cancer detection, and is commonly used for the preoperative evaluation of newly diagnosed breast cancer cases [1][2][3]. Morphological features and patterns of contrast enhancement on MR imaging provide important information, not only of tumor histological features, such as vascularization, but also survival outcome prediction [4][5][6].
Regarding evaluating kinetic features, there are several methods to assess the kinetic enhancement patterns of breast tumor obtained from DCE MR imaging [7]. For quantitative analysis, pharmacokinetic models, such as Tofts and extended Tofts models, are standard for analyzing DCE MRI data. These models are suitable when the tissue under investigation is weakly vascularized or in the presence of a negligible intravascular concentration of contrast agent [8,9]. The Tofts model considers intravascular space as one compartment and extracellular extravascular space as the other compartment, excluding the functionality of intravascular space. Computation of pharmacokinetic parameters requires curve fitting to concentrationtime curves for multiple voxels. This is typically time-consuming and hinders online access of pharmacokinetic maps by radiologists [10]. Most commonly, breast kinetic enhancement patterns are measured semi-quantitatively using modest temporal resolution with at least two to three post-contrast T1-weighted acquisitions, with k-space centered at approximately 90-120 seconds after contrast injection for the first post-contrast images [7]. Using the data obtained at each of these time points, a time-signal intensity curve can be determined for a given lesion or region of interest. Despite its high utility in functional analysis, kinetic assessment using a manually drawn region of interest (ROI) is limited because it only reflects partial pixel information of the lesion. Moreover, ROI placement and interpretation are operator-dependent and timeconsuming.
Despite the controversy of whether it improves the diagnostic performance of MR imaging, computer-aided diagnosis (CAD) systems have the advantage of being easy to use, and are widely clinically available to automatically assess kinetic information and perform volumetric analysis for breast lesions [11][12][13][14]. CAD systems eliminate the need for labor-intensive tasks, such as manual drawing of ROI by a technician or radiologist. In addition, CAD has the potential to increase observer reproducibility in DCE MR imaging by performing automatic and quantitative analysis of the contrast material uptake [15][16][17].
The correlation between kinetic parameters on CAD and the histological characteristics of breast cancer have been evaluated [18]. CAD-derived kinetic information may be used to assess tumor aggressiveness, including hormone receptor status and histologic grade. In terms of survival outcomes, Yi et al. [19] reported that a smaller reduction in the washout component on DCE MR imaging assessed using CAD after neoadjuvant chemotherapy (NAC) in breast cancer patients was independently associated with a worse recurrence-free survival (RFS) and overall survival [19]. In a preoperative setting, Dietzel et al. [20] evaluated the potential of DCE MR imaging to predict disease-related death using CAD on 115 patients, finding that total tumor volume, initial peak enhancement, and time to peak enhancement showed a negative correlation with survival. The persistent curve type, however, showed a positive correlation with survival. Baltzer et al. [21] analyzed the enhancement kinetics of 59 patients using CAD and found that a washout component could be identified as a significant and independent predictor of distant metastasis occurrence. However, because their studies only included a small sample size, patients in NAC setting, or a qualitative assessment using in-house software, there is currently a lack of reports evaluating the correlation between preoperative MR imaging kinetic parameters assessed using a commercially available CAD with recurrence outcomes in patients with breast cancer.
Therefore, the purpose of our study was to evaluate whether the kinetic features of preoperative MRI, assessed using CAD, can predict survival outcomes in patients with newly diagnosed invasive breast cancer.

Patients
This retrospective study was approved by the Institutional Review Board of the Samsung Medical Center (SMC IRB 2017-03-057) and waived the need for informed patient consent.
Between March and December 2011, 823 consecutive women (mean age of 50 years, range 24-85 years) who had undergone surgery for invasive breast cancer were identified. The inclusion criteria for our study were as follows: (a) completion of preoperative DCE MR imaging with CAD at our institution, (b) initial unilateral breast biopsy-proven malignancy with a final pathologic diagnosis of invasive breast carcinoma, (c) a lesion manifesting as a mass on MR imaging. We identified 753 invasive breast carcinomas that met these criteria. We excluded cases where MR imaging was performed after diagnosis by vacuum-assisted or excisional biopsy (n = 97). Patients treated with NAC (n = 81), those who underwent breast MR imaging with a 3-T scanner (n = 264), and those who had a follow-up period of less than 3 months (n = 10) were also excluded. To exclude any magnetic strength effects during the MR examination, we only included patients who underwent 1.5-T MRI, ensuring that the MR images were obtained under homogeneous conditions. Finally, 301 women, aged 26-80 years (mean age 51 years) were included in our study. S1 Dataset provides detailed characteristics of patients (online). The mean interval between preoperative MR examination and surgery was 13 days (range 1-37 days).

MR imaging protocol
MR imaging was performed using a 1.5-T system (Achieva; Philips Medical Systems, Best, The Netherlands) with a dedicated bilateral phased-array breast coil, with the patient in a prone position. The MR imaging examination consisted of turbo spin-echo T1-and T2-weighted sequences, and a three-dimensional DCE sequence. The DCE MR imaging sequences were performed using the following parameters: repetition time (TR) = 6.5 ms, echo time (TE) = 2.5 ms, slice thickness = 1.5 mm, flip angle = 10˚, matrix size = 376 × 374 mm, field of view (FOV) = 32 × 32 cm. Axial DCE MR imaging was performed using 1 precontrast and 6 postcontrast dynamic series. Contrast-enhanced images were acquired 30, 90, 150, 210, 270, and 330 seconds after the contrast material injection. The length of each dynamic series was 60.05 seconds. A 0.1 mmol/kg bolus of gadobutrol (Gadovist; Bayer Healthcare Pharmaceutical, Berlin, Germany) was injected for dynamic contrast imaging, before a 20-mL saline flush.

Assessment of MR imaging and computer-aided diagnosis
The DCE MR images were transferred to a CAD system (CADstream, version 4.1.3, Merge Healthcare, Chicago, IL, USA) that analyzed the signal intensities within each voxel of the FOV obtained during the dynamic sequences. The American College of Radiology Breast Imaging Reporting and Data System (ACR BI-RADS) MR lexicon [22] defined a persistent kinetic pattern as a continuous increase in signal over time. A plateau is defined as a signal intensity that does not change over time, after an initial rise. When the signal intensity decreases by more than 10%, after its highest point following its initial rise, it is defined as a washout. Using CAD stream, a color overlay showing the changes in signal intensity over time was automatically generated in all slices using a predefined minimum threshold. We defined the minimum threshold as 50%, or a greater increase in pixel-by-pixel comparison between the precontrast and second post-contrast images. When the relative enhancement increase was more than 100% of the precontrast image, it was classified as "rapid uptake," and when it was 50-100% it was classified as "medium uptake." When the signal intensity continued to show an increase of more than 10% in the sixth postcontrast-enhanced images, compared to the second postcontrast-enhanced images, it was classified as "persistent." When the signal intensity showed a decrease of more than 10%, it was classified as a "washout," and when the signal intensity continued to show within a 10% range, it was classified as a "plateau." We maintained a single threshold value for early and delayed phase enhancements across all patients. Volumetric assessment of the kinetic components refers to the assessment of the percentage volume of each kinetic component found within the tumors at both early and delayed phases of enhancement. The CAD reports included the following kinetic features: the initial peak enhancement values, proportions of early phase medium and rapid enhancements, and the proportion of delayed phase persistent, plateau, and washout enhancements. The following three kinetic curve patterns presented by CAD were also included in the CAD reports: type 1, delayed persistent enhancement pattern; type 2, delayed plateau enhancement pattern; and type 3, delayed washout enhancement pattern. The CAD reports were stored prospectively by interpreting radiologists, and square ROI of the entire tumor was automatically segmented when an operator chose the image slice showing the largest diameter. For cases that had multiple cancers on the MRI, we used the values of the index cancer.
The MR imaging findings were retrospectively evaluated according to BI-RADS MR lexicon [22] by two board-certified radiologists (S.Y.N., E.S.K., with 6 and 10 years of experience in breast MR imaging, respectively, in consensus. In cases with discrepancy in interpretation between the two readers, a third radiologist (J.S.C. with 7 years of experience in breast MR imaging) reviewed the images to reach a consensus. The radiologists assessed the shape (oval, round, or irregular), margin (circumscribed, irregular, or spiculated), and internal enhancement characteristics (homogeneous, heterogeneous, rim, or dark internal septation) of each mass.

Histopathological features
The following parameters of the pathological reports of either breast-conserving surgery or mastectomy specimens were reviewed: pathologic diagnosis, histologic grade, surgical margin status, presence of an extensive intraductal component (EIC), lymphovascular invasion (LVI), estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), and Ki-67 expression status. Positivity for ER and PR was defined using a cut-off value of greater than 1% positively stained nuclei. Immunohistochemical HER2 scores of 3+ (strong homogeneous staining) were considered positive. In cases of HER2 scores of 2+, HER2 gene amplification was confirmed using silver in situ hybridization (SISH). For Ki-67 expression status, nuclear staining of at least 14% was considered to indicate a high level of expression. Breast cancers were divided into three molecular subtypes based on the immunohistochemical or SISH findings for ER, PR, HER2 as follows: luminal (hormone receptor positive and any of HER2 status), HER2-enriched (hormone receptor negative and HER2 positive), and triple-negative (hormonal receptor negative and HER2 negative) [23]. Tumors were divided into three histological groups: invasive ductal carcinoma, invasive lobular carcinoma, and others. The surgical margin status was classified as positive or close (<2 mm), or negative. Because frozen biopsy was routinely performed during surgery at our institution, immediate repeat excision was performed on the same day for patients showing a positive margin in the frozen biopsy. Only the final results of the surgical margin were analyzed. We reviewed the patients' medical records to identify whether they received adjuvant chemotherapy, adjuvant radiation therapy, or adjuvant endocrine therapy. We also reviewed the TNM stage after surgery and the follow-up information from the medical records including imaging studies for each patient.

Statistical analyses
Breast cancer recurrence was defined as local recurrence (limited to the ipsilateral breast or chest wall), regional recurrence (ipsilateral axillary, infraclavicular, or supraclavicular lymph nodes), contralateral breast, or distant metastasis to other parts of the body. DFS and diseasespecific survival (DSS) were calculated from the date of surgery to that of breast cancer recurrence, the date of death, the date last known to have no evidence of disease, or the date of the most recent follow-up. Patients without an event were censored at the date of the most recent follow-up, regardless of whether they were scheduled for future follow-up or they had been lost to follow-up.
The Cox proportional hazards model was used to determine the association between MR imaging kinetic features derived from CAD and DFS or DSS. The peak signal intensity and volumetric assessment of different kinetic components were compared to the clinical-pathological variables using a Student's t-test and analysis of variance (ANOVA).
Statistical significance was set at P < .05. Variables with P < .05 on the univariate analysis were entered as the input variables for a multivariate model. All statistical analyses were performed by a dedicated statistician using R statistical software (version 3.2.4; R Foundation for Statistical Computing, Vienna, Austria).

Correlation between MR kinetic parameters and clinical-pathological variables
The correlation between CAD-derived MR kinetic features with clinicopathological variables are presented in Table 4. Higher peak enhancement was significantly associated with a higher T stage, higher clinical stage, and a higher histologic grade. Pathologic diagnosis of invasive ductal carcinoma, higher histologic grade, and triple-negative subtype were significantly associated with a higher washout component on the delayed enhancement phase on DCE MR imaging. Higher histologic grade, ER negativity, PR negativity, and p53 positivity were associated with a higher mean percentage volume of rapid enhancement components on early phase enhancement. On the contrary, lower histologic grade, ER positivity, PR positivity, and p53 negativity were associated with a higher mean percentage volume of medium enhancement components on early phase enhancement. Persistent enhancement on delayed phase was seen in tumors with a lower histologic grade.

Discussion
Multivariate survival analysis revealed that patients with breast cancer who showed a higher peak enhancement (DFS hazard ratio, 1.004 [95% CI: 1.001, 1.006]; P = .013) on DCE breast MR imaging and triple-negative subtype (DFS hazard ratio, 21.060 [95% CI: 2.675, 165.780]; P = .004) had worse DFS. Higher peak enhancement was significantly associated with a higher T stage, higher clinical stage, and higher histological grade. Pathological diagnosis of invasive ductal carcinoma, higher histologic grade, and triple-negative subtype were significantly associated with a higher washout component on the delayed enhancement phase on DCE MR imaging. Our study showed that higher peak enhancement was associated with poorer DFS as  [18] found that a higher peak signal intensity was seen in ER-negative, PR-negative, and triple-negative tumors, compared to ER-positive, PRpositive, and non-triple-negative tumors. Kinetic features on DCE MR imaging are known to be dependent on the tumor perfusion flow, microvessel density, vascular permeability, and volume of extracellular-extravascular space composition [25][26][27]. Among CAD-derived kinetic parameters, peak enhancement is considered to reflect the concentration of the contrast agent in both intravascular and extravascular interstitial spaces [8]. In previous reports, high microvessel density and increased vascular permeability were associated with a higher nuclear grade [28], axillary lymph node metastasis, and distant metastasis [4,21,29]. Other studies have shown that peak enhancement was significantly greater among patients with lymph node metastasis and lymph node extracapsular extension [30,31]. We believe that these results might explain the relationship between the peak enhancement on DCE MR imaging and poor prognosis.
Regarding delayed phase kinetic features, a washout kinetic pattern has been reported to correlate with a higher histologic grade, higher Ki-67 expression, increased vascular permeability, and HER2-enriched subtype [18,32,33]. In addition, a higher vascular permeability due to increased expression of vascular endothelial growth factor typically found in fast-growing tumors, may cause a higher washout component [8,34]. Based on these prior results, we expected that a higher washout component on delayed phase enhancement might be associated with a poorer prognosis. In our study, a higher washout component on delayed phase enhancement was associated with the pathological diagnosis of invasive ductal carcinoma,  [35] evaluated the association between CAD-generated kinetic features and DFS of patients with primary operable breast cancer, reporting that a higher peak enhancement and washout component were associated with poorer DFS. When comparing with their study, profiles of CAD-generated kinetic features were very different. For example, the mean value of washout component in patients with recurrence was much lower than that of ours (10.38% vs. 39.19%). We speculate that these differences might be caused by using different machines or different protocols. Therefore, using same threshold of CAD-generated features obtained from other institution might be not feasible. Larger studies may be necessary to provide more comprehensive information.
In our study, none of the included morphological findings (shape, margin, internal enhancement) were significantly associated with DFS. Tumors with an irregular shape often occur in high grade breast cancer [36]. However, there was no significant difference in the current study.

Fig 1. Magnetic resonance (MR) images with computer-aided detection (CAD) Color Overlay Map in a 66-yearold Woman with grade 2 invasive ductal carcinoma in the left breast. (A) Axial contrast-enhanced T1-weighted MR
image shows a 26-mm irregular mass (arrows). (B) Axial maximum-intensity-projection MR image shows CAD color overlay over the breast mass. L1 denotes the first lesion of the left breast. Areas in red, yellow, and blue indicate a rapid washout-type delayed enhancement, plateau-type delayed enhancement, and persistent-type delayed enhancement pattern, respectively. (C) Kinetic curve graph showing rapid initial enhancement and rapid washout-type curve. The initial peak enhancement value was 559%. With respect to the delayed phase enhancement, 9% of the mass showed washout, 50% of the mass showed a persistent-type curve, and 41% showed a plateau-type curve. Regional recurrence was diagnosed in the ipsilateral axillary lymph node 26 months after surgery. https://doi.org/10.1371/journal.pone.0195756.g001 In addition to MR imaging, the evaluation of dynamic patterns on breast imaging has also been applied to other modalities. Contrast-enhanced ultrasound (CEUS) imaging has been introduced for the evaluation of tumor kinetics and it enables determinations of the diffusion pattern and real-time quantification of the contrast agent [37][38][39]. Several quantitative parameters could be automatically calculated on CEUS from the time-intensity curve as follows: peak intensity; time to peak intensity; mean transit time; slope; rise time; and area under the timeintensity curve. CEUS facilitates the functional evaluation of microcirculation and hemodynamic characteristics without extravascular enhancement, which is seen in contrast-enhanced CT and MRI. However, there are no standardized criteria for differentiating between malignant and benign tumors because of different US machines and contrast agents. In addition, there is no general consensus about the methods of imaging acquisition, protocol of contrast agent injection, and ROI setting for time-intensity curve analysis. CEUS is not routinely used in clinical practice because it is difficult to scan an entire breast within a few minutes and time-consuming post-imaging analysis in real-time examinations [40]. Contrast enhanced spectral mammography (CESM) is currently increasing in use, which combines the relative Axial maximum-intensity-projection MR image shows a CAD color overlay over the breast mass. R1 denotes the first lesion of the right breast. Areas in red, yellow, and blue indicate rapid washout-type delayed enhancement, plateau-type delayed enhancement, and persistent-type delayed enhancement patterns, respectively. (C) Graph of the kinetic curve shows rapid initial enhancement and rapid washout-type curve. The initial peak enhancement value was 119%. With respect to the delayed phase enhancement, 39% of the mass showed washout, 4% of the mass showed a persistent-type curve, and 58% showed a plateau-type curve. There was no recurrence during the 65-month follow-up period.
https://doi.org/10.1371/journal.pone.0195756.g002 ease, low cost, and practicality of mammography with the high sensitivity of MRI [41][42][43]. CESM uses intravenous contrast media to obtain the images of enhancing structures. However, dynamic enhancement pattern could not be shown because CESM only shows the existence of enhancement. In the present study, we evaluated the value of CAD-derived kinetic features and our results showed that they might have the potential to be used to characterize tumor aggressiveness and predict the patients' outcome.
We only included patients who underwent 1.5-T MR imaging because Jansen et al [44] demonstrated that kinetic curves might not be completely consistent across three different MR imaging systems with the same field strength (1.5-T). In another study, qualitative measures of curve shape were not consistent across different field strengths, even when the acquisition parameters were standardized; the maximum percent signal enhancement was significantly higher at 3 T than at 1.5 T [45,46]. Further studies are required to investigate differences between various MR imagers.
Our study had some limitations. First, this study had a relatively short follow-up period (mean, 55.2 months). It is well known that women with ER-positive tumors remain at risk for late recurrences after primary treatment, which might partly account for the small number of patients with recurrence, making it difficult to observe a robust survival outcome [47]. Second, our results might be difficult to generalize because they were based on a single center, mass lesions, a single imaging protocol, and single CAD system. Third, we could not evaluate inter-or intra-observer agreement of MR kinetic values because the CAD reports were stored prospectively. However, the same results are likely between observers given that we used a standardized ROI policy for choosing image slices that revealed the largest diameter and only mass lesions (non-mass enhancements were excluded). In our experience, significant segmentation differences are markedly decreased in mass lesions compared with nonmass enhancement lesions. In addition, the purpose of our study was not to evaluate the diagnostic accuracy of CAD, but observational survival analysis. Fourth, although these factors might have affected CAD-derived values, we did not consider the effect of the patient's motion or field inhomogeneity. Finally, of the 32 recurrences, 2 recurrences were detected within the first 6 months from the initial diagnosis, and these could have been due to residual disease.
In conclusion, our findings suggested that preoperative MR imaging kinetic features assessed using a commercially available CAD have the potential to predict survival outcomes. Patients with breast cancer who showed a higher peak enhancement on breast MR imaging may exhibit a worse DFS. In addition, peak enhancement and volumetric analysis of the early and delayed phase enhancements of breast cancers showed correlation with tumor stage, pathological findings, and histological grade, which are indicators of tumor aggressiveness.