Neoadjuvant chemotherapy (NAC) is frequently used in the standard of care (SoC) to treat patients with locally advanced breast cancer to downstage the disease and facilitate breast-conserving surgery. In addition, the setting of NAC systemic treatment affords the opportunity to assess the pathologic response to NAC. Studies have shown that pathologic complete response (pCR) or minimal residual disease following NAC is prognostic for survival (1–6). However, since assessment of pathologic response can only be ascertained from surgical tumor specimens after NAC has already been completed, the treating clinician’s options to tailor therapy regimens during NAC to improve pathologic response outcome and consequently survival are limited. Therefore, minimally invasive methods that can provide an accurate prediction of response in the early stages of NAC are urgently needed. With many innovative treatment regimens using targeted therapies and/or immunotherapies being tested in clinical trials for breast cancer treatment, improved capability of accurate and early prediction of pathologic response to NAC may allow rapid individualized regimen de-escalation/alteration for responding/non-responding breast cancer patients in the future, facilitating precision medicine and leading to improved treatment outcomes.

In current SoC and clinical trials, the measurement of imaging tumor size change according to the Response Evaluation Criteria in Solid Tumors (RECIST) guidelines (7, 8) is the standard approach to assess tumor response to therapy. However, many studies have shown that tumor size changes in response to therapy, especially targeted therapies, often lag well behind changes in the underlying tumor biological functions (9–11), such as perfusion/permeability, cellularity, and metabolism. Therefore, the anatomic imaging approach of tumor size measurement is generally less effective for the early prediction of therapeutic response compared to functional imaging methods. As a noninvasive method for evaluation of microvascular perfusion and permeability, dynamic contrast-enhanced (DCE) MRI has been increasingly used in research settings, including early phase clinical trials, to evaluate breast cancer responses to NAC. There are usually three approaches in the analysis of DCE-MRI time-course data: qualitative curve shape description, calculation of semi-quantitative metrics such as contrast agent (CA) uptake and wash-out slopes, and quantitative pharmacokinetic (PK) modeling to extract parameters, such as K^trans (CA volume transfer rate constant) and v_e (extravascular, extracellular volume fraction), which are more directly reflective of the underlying biological functions and are in principle independent of scanner vendor platforms and data acquisition details. Thus, compared with qualitative and semi-quantitative DCE-MRI, quantitative DCE-MRI is hypothetically a more desirable approach for the assessment of cancer therapy response. Using summary receiver operating characteristic (SROC) analysis, a recent meta-analysis (12) of 14 published studies including 739 patients showed that the area under the curve (AUC), sensitivity, and specificity of K^trans for early discrimination of pCR and non-pCR after one to two NAC cycles were 0.90, 84%, and 83%, respectively. However, the promise of K^trans as an imaging biomarker for the prediction of breast cancer response to NAC has been demonstrated mostly by single-site and single-vendor platform DCE-MRI studies (13) that often sacrificed spatial resolution and/or coverage for the high temporal resolution necessary for PK modeling of the time-course data. There remain significant technical challenges in implementing quantitative DCE-MRI in multi-center (MC) and multi-vendor platform (MP) settings to assess breast cancer response to NAC, as many technical aspects in data acquisition and analysis, from temporal resolution to selection of the PK model and software tool, can affect the accuracy and precision of the derived PK parameters (14–17). Therefore, it is of paramount importance to standardize data acquisition and analysis strategies in MC and MP study settings (14, 15).

In this study, through preliminary analysis of data from an ongoing project, we report our initial experience in implementing quantitative DCE-MRI with simultaneous high spatial and temporal resolutions in an MC and MP setting to predict breast cancer response to NAC, compare predictive performances among quantitative and semi-quantitative DCE-MRI parameters, and tumor size measurement, and make initial best-practice recommendations with regard to PK analysis of breast DCE-MRI data acquired from different sites with different vendor platforms.

To the best of our knowledge, this is the first MC and MP study using quantitative DCE-MRI to predict breast cancer response to NAC, where GRE-based product sequences of k-space undersampling during acquisition and view-sharing during reconstruction from Siemens, Philips, and GE platforms were used for high spatial and temporal resolution breast DCE-MRI. The data acquisition scheme used for DCE-MRI in this MC and MP setting allows for bilateral full breast coverage with adequate spatial resolution for accurate morphological evaluation, as well as sufficient temporal resolution for quantitative PK analysis of the breast DCE time-course data (14, 15). In today’s SoC breast DCE-MRI protocols, owing to trade-offs between spatial and temporal resolution in conventional sequences, the necessity for accurate tumor morphology assessment with high spatial resolution and spatial coverage results in low temporal resolutions between 60 s and 120 s, which precludes meaningful PK modeling of time-course data with acceptable accuracy (15, 16). With potentially faster dynamic imaging methods available from vendors through combinations of accelerated data acquisition approaches and advanced reconstruction algorithms (15), the ability to acquire breast DCE-MRI with simultaneous high spatial and temporal resolutions using vendor product sequences may facilitate the translation of quantitative DCE-MRI into clinical workflow.

With inherent differences in hardware and software among the three major MRI vendor platforms, standardizations in VFA and DCE acquisition parameters were implemented across the three sites to minimize differences in the acquired data and consequently, variations in results from data analysis (14, 15, 17). These included the use of the same three FAs in VFA acquisition for R_1,0 mapping, and the same center and peripheral portions of the k-space in the three vendor-specific k-space undersampling and view-sharing product sequences for DCE-MRI acquisition. In addition, the same FA (10°), minimal TE (0.9 ms–2.9 ms), and similar TR in the range of 5.0 ms–6.2 ms were used across the platforms to ensure DCE data sensitivity to the water exchange effects (38), supporting the use of the SSM for PK analysis. Despite efforts in data acquisition standardization and centralized data analysis using a single software tool for both VFA R₁ fitting and DCE-MRI PK modeling, VFA measurements of phantom R₁ and patient tumor R_1,0 on one vendor platform resulted in substantial biases compared to ground truth values in the phantom and measurements on the other two platforms, as well as literature reported breast tumor R_1,0 values (25–27). The ≥100% overestimation of tumor R_1,0 on this platform also caused patient data fitting failures when using either the TM or SSM. Therefore, PK modeling of patient data from this vendor platform was performed using fR_1,0 only, excluding the use of mR_1,0 from the analysis. The sources of errors in VFA R₁ mapping on this platform are still under investigation. VFA fitting is inherently challenging owing to the need to fit multiple variables, which requires numerous individual images of unique FAs. Because this is not practical owing to the scan time, several assumptions are typically made to simplify the fitting, including the linearity and spatial uniformity of B₁ over multiple transmitted power settings, as well as the direct correspondence between the delivered FA of the independent B₁ mapping sequence and the VFA sequence. These assumptions rely on the use of non-clinically validated research tools from vendor platforms. However, these assumptions may not always hold true on specific vendor implementations. Furthermore, each vendor uses a different method to map the FA. There are no standards for the reporting of FA or B₁ maps, including units, presenting challenges when assimilating data across multivendor platforms. One valuable lesson learned here for implementing MC and MP quantitative DCE-MRI is the importance of QA/QC scans of phantoms with ground truth R₁ values to determine whether all vendor platforms provide reliable R₁ measurements. If this is not the case, either error sources should be identified and corrective actions taken, or an alternative solution, such as the use of fR_1,0 for PK analysis, should be found.

In this study, using K^trans as the reference imaging biomarker for the prediction of breast cancer response to NAC, we found that voxel-based analysis using a literature-reported fR_1,0 value was the optimal approach for PK analysis of DCE-MRI data collected from the MC and MP settings, whether the TM or SSM was used for data modeling. This approach allowed the inclusion of data from one platform, where the mR_1,0 values calculated from VFA measurements were unreliable. The statistically insignificant differences in K^trans between the use of fR_1,0 and mR_1,0 ( Figure 1 ) suggest that it is reasonable and practical to use fR_1,0 for PK modeling (44), which can mitigate random and systematic errors from R_1,0 measurement across vendor platforms and potentially eliminate the need for B₁ mapping and VFA acquisition in MC and MP trials. Although the use of fR_1,0 instead of mR_1,0 in PK analysis is expected to cause systematic errors in the estimated PK parameters, unlike random errors, the impact of systematic errors is lessened when percent changes in PK parameters, such as V21% and V31%, are used in a longitudinal study to predict breast cancer response to NAC. Compared with voxel-based analysis, ROI-based analysis dilutes tumor heterogeneity in perfusion and permeability, resulting in significantly smaller K^trans values and a narrower range of K^trans changes in response to therapy. The latter may potentially reduce the predictive performance of K^trans for NAC response. Another disadvantage of ROI-based analysis is the inability to assess changes in PK parameter heterogeneity in response to treatment.

The preliminary results from this MC and MP study show that after only one NAC cycle, semi-quantitative and quantitative DCE-MRI metrics outperformed tumor size measurement in the early prediction of breast cancer response to NAC. The quantitative parameter K^trans consistently provided a more accurate prediction of NAC response than both size measurement (LD) and SER after the first NAC cycle and at the NAC midpoint. These findings agree with many similar studies that used DCE-MRI to assess the breast cancer response to NAC (12, 13). The larger decreases in K^trans (iqr) and k_ep (iqr) in pCRs compared to non-pCRs at V2 and especially V3 indicate greater decreases in tumor perfusion/permeability heterogeneity in patients responding to NAC regimens.

Consistent with previous studies (36, 45), SSM K^trans was substantially greater than TM K^trans in this cohort of malignant breast tumors when DCE-MRI acquisition was sensitive to the water-exchange effect (38). Since there is no significant difference between SSM and TM K^trans in benign breast lesions or normal tissue (36, 45), SSM K^trans potentially has better predictive performance than TM K^trans due to the former’s greater dynamic ranges of change in response to therapy, assuming that microvascular properties of a responding tumor shift towards those of a benign lesion or normal tissue. This is manifested by the fact that, after the first NAC cycle, while V21% of mean SSM K^trans was a good predictor of NAC response with an ROC AUC value of 0.83, V21% of mean TM K^trans was only a fair predictor with ROC AUC = 0.70 (not shown in Table 4 ). Although at the NAC midpoint, both SSM and TM V31% K^trans (and k_ep) are excellent predictors of NAC response, the percent changes of SSM K^trans and k_ep were larger than those of the TM counterparts, and the P-values from comparing the two response groups were generally smaller for the SSM parameters.

Our initial experience shows that when the water exchange effect is explicitly modeled in the SSM analysis of DCE-MRI data, the k_io parameter may provide complementary information to the more commonly modeled K^trans parameter, which only focuses on CA kinetics. Recent studies have shown that k_io is an imaging biomarker of metabolic activity (46). In fact, V31% of the tumor mean k_io was a fair early predictor of NAC response with an ROC AUC value of 0.71 (not shown in Table 4 ). Combining K^trans and k_io in a multivariate predictive model may further improve the predictive accuracy for NAC response. NAC regimens often reduce tumor permeability/vascularity (21), resulting in reduced interstitium [CA] during a DCE study. This, in turn, makes k_io quantification less reliable. In other words, reduced CA extravasation results in a smaller R₁ difference between extracellular and intracellular spaces. This decreases the DCE-MRI sensitivity to the water exchange effect and, consequently, negatively impacts the accuracy and precision of SSM quantification of k_io. Therefore, caution should be exercised when evaluating the estimated k_io values from the SSM analysis. Unreliable voxel k_io values should be filtered out. The smaller fraction of filtered k_io at V4 in pCRs compared to non-pCRs quantitatively reflected lower CA extravasation in the former, consistent with the K^trans and k_ep results. The right side of each panel in Figure 7 represents a simple method for quantitatively estimating the extent of CA extravasation using Σ[CA_o]. As expected, the pCR tumor at V3 showed low Σ[CA_o] owing to reduced perfusion and permeability in response to NAC treatment, resulting in large areas of unreliable k_io being filtered out. A more accurate measure of voxel DCE data sensitivity to water exchange should include an estimation of the extent and duration of |R_1i − R_1o| absolute difference between intracellular and extracellular R₁ exceeding (or at least close to) the exchange kinetics defined by k, the transmembrane water molecule exchange process (42). However, this may add complications in translational studies, such as this MC and MP study.

Our model-based approach for estimating BAT eliminated the need to manually align the AIF with voxel-based tissue DCE curves in PK analysis, which improves automation in the entire data processing workflow. Manual inspection of the model-selected BATs showed robust and qualitatively good performance. Future work will include quantifying the performance using a digital phantom for comparison with manual alignment.

There are several limitations to this preliminary MC and MP study. The main limitation is the small sample size of 15 patients in total, which caused large 95% CI ranges of the ROC AUC values for the prediction of NAC response and may artificially inflate the predictive performances of the quantitative PK parameters. The small sample size at each site also renders cross-vendor platform comparison of results unreliable, and therefore, was not performed. Second, for correlation analysis between MRI metrics and pathologic response outcomes, the small sample size precluded meaningful analysis stratified by breast cancer subtypes, as shown in Table 1 . Therefore, the results presented here are more reflective of those from the general breast cancer population treated with SoC NAC regimens. Third, we did not perform a multivariate analysis by combining clinicopathological features with individual MRI metrics with high predictive performance, which may further enhance the capability of early prediction of NAC response, especially after the first NAC cycle. Fourth, we only used the iqr of voxel-based parameters to characterize changes in tumor heterogeneity in response to NAC, without performing a more comprehensive radiomics analysis (47), which may provide better predictive performance. Lastly, only three FAs within a relatively narrow range were used in VFA measurements of phantom R₁ and breast tumor R_1,0. In clinical practice, where breast tumor R_1,0 values are unknown and could vary greatly, using more FAs over a larger range may result in more accurate R_1,0 mapping, and consequently, more accurately estimated PK parameters. However, the usual time constraint for a clinical MRI protocol makes it difficult to add more FAs to the VFA acquisition. Furthermore, the small number of FAs used in this study is unlikely to be the reason why one vendor platform returned substantially biased R₁ and R_1,0 values.

In conclusion, the initial results from this MC and MP study validate findings from many single-site studies that quantitative DCE-MRI is superior to tumor size measurement for the prediction of breast cancer response to NAC. Both SSM and TM K^trans showed better predictive performance than the semi-quantitative SER metric. Furthermore, K^trans and k_ep derived from the SSM, using DCE-MRI data acquired with sensitivity to the water exchange effect, generally performed better than the TM counterparts in the prediction of NAC response, especially after only one cycle of NAC. Due to potential large variations in the accuracy of VFA-measured R_1,0 on different vendor platforms, SSM PK analysis using a fixed, literature-reported breast tumor R_1,0 could be a best-practice approach in quantitative DCE-MRI prediction of breast cancer response to NAC in an MC and MP setting.