Features of MRI stromal enhancement with neoadjuvant chemotherapy: a subgroup analysis of the ACRIN 6657/I-SPY TRIAL

Abstract. Although the role of cancer-activated stroma in malignant progression has been well investigated, the influence of an activated stroma in therapy response is not well understood. Using retrospective pilot cohorts, we previously observed that MRI detected stromal contrast enhancement was associated with proximity to the tumor and was predictive for relapse-free survival in patients with breast cancer receiving neoadjuvant chemotherapy. Here, to evaluate the association of stromal contrast enhancement to therapy, we applied an advanced tissue mapping technique to evaluate stromal enhancement patterns within 71 patients enrolled in the I-SPY 1 neoadjuvant breast cancer trial. We correlated MR stromal measurements with stromal protein levels involved in tumor progression processes. We found that stromal percent enhancement values decrease with distance from the tumor edge with the estimated mean change ranging −0.48 to −0.17 (P≤0.001) for time points T2 through T4. While not statistically significant, we found a decreasing trend in global stromal signal enhancement ratio values with the use of chemotherapy. There were no statistically significant differences between MR enhancement measurements and stromal protein levels. Findings from this study indicate that stromal features characterized by MRI are impacted by chemotherapy and may have predictive value in a larger study.

Features of MRI stromal enhancement with neoadjuvant chemotherapy: a subgroup analysis of the ACRIN 6657/I-SPY TRIAL 1 Introduction Much like the wound healing process, cancer stroma comprises neovasculature, inflammatory cells, and extracellular matrix components that work cooperatively to support malignant progression. 1 Through active angiogenesis, stimulated cell proliferation, and the synthesis of connective-tissue, cancer stroma transforms into a highly cellular vascularized tissue. Consequently, cancer stroma has been commonly referred to as activated or reactive stroma. 2 While its similarity to wound healing is now well understood, the clinically relevant features of activated stroma in treatment and prognosis are unclear. To investigate properties of activated stroma that may play a role in cancer therapy or prognosis, better tools and assessment techniques are necessary. We previously showed that the stroma surrounding breast primary tumors was distinguishable by detectable enhancement patterns using contrast-enhanced breast magnetic resonance imaging (MRI). 3,4 Using a tumor proximity mapping technique, stromal enhancement could be related to distance from the tumor edge. 4 Our previous work suggested that increased peritumoral enhancement was in part attributable to neovasculature 5 and that chemotherapy-induced changes in global stromal enhancement were associated with disease outcome among a pilot cohort of patients receiving neoadjuvant chemotherapy (NAC). 3 In this study, we aimed to validate our approach to mapping stromal enhancement and to assess its prognostic value in a prospectively enrolled cohort in the I-SPY 1 TRIAL (investigation of serial studies to predict your therapeutic response with imaging and molecular analysis) with patients receiving NAC and serial MRI scans for breast cancer during treatment. To help identify biological differences among stromal enhancement patterns, we correlated stromal enhancement measured from MRI with stromal proteins that are actively involved in the tumor microenvironment measured from biopsy speciments using reverse-phase proteins microarrays (RPPAs). These proteins are collectively involved in angiogenesis, [6][7][8] wound healing, 6 inflammation, 9 and tissue remodeling. 10,11 In this study, we test the hypothesis of whether stromal enhancement has a radial dependence on the distance from the tumor edge, and the effect of chemotherapy on global stromal enhancement.

Study Design and Patient Selection
The I-SPY 1 TRIAL was a study of imaging and tissue-based biomarkers for prediction of treatment response and survival. The imaging component of this trial was conducted as a companion study to the American College of Radiology Imaging Network (ACRIN) 6657 and Cancer and Leukemia Group B (CALGB) 150007. Participation in both ACRIN 6657 and CALGB 150007 was required to enroll in the I-SPY 1 TRIAL. All patients gave a signed consent to participate. Eligible patients had histopathologically confirmed breast cancers without evidence of distant metastatic disease. The NAC included an initial four cycles of anthracycline-cyclophosphamide (AC) after which patients either underwent surgery or received four cycles of taxane-based treatment prior to surgery.

MRI Acquisition
The I-SPY 1 TRIAL patients underwent contrast-enhanced MRI at visits before the start of chemotherapy (T1), after one cycle of AC (T2), between AC and taxane-based (T3) regimen, and at the completion of chemotherapy prior to surgery (T4). Images were acquired as previously described. 12,13 Briefly, MRI was performed on the tumor bearing breast (ipsilateral) only. Images were acquired on a 1.5-T scanner using a dedicated breast radiofrequency coil. Patients were imaged in the prone position. The MR imaging protocol included a localization acquisition and a T2-weighted sequence, followed by a dynamic contrastenhanced series. For T2-weighted imaging, a fast spin-echo sequence with fat suppression was used (two-dimensional spin-echo; field of view, 16 to 20 cm; section thickness, 3 mm; fat saturation; echo train length, 8 to 16; one echo; effective echo time, 80 to 140 ms; repetition time, 4000 to 6000 ms). Gadopentetatedimeglumine (Magnevist, Bayer HealthCare) was used as a contrast agent and was injected at a dose of 0.1 mmol∕kg of body weight (1.2 mL∕s) followed by a 10-mL saline flush. The contrast-enhanced series consisted of a high-resolution (≤1 mm in-plane spatial resolution), threedimensional (3-D) fast gradient-recalled echo sequence (repetition time ms/echo time ms, ≤ 20∕4.5; flip angle, ≤45 deg; field of view, 16 to 18 cm; minimum matrix, 256 × 192; 64 sections; section thickness, ≤2.5 mm; voxel size, 29.3 mm 3 ). The entire breast was scanned across the sagittal plane with 64 slices for each 2.5 mm thickness over a total scanning time. Imaging time for the T1-weighted sequence was required to be between 4.5 and 5 min, with the data set acquired before injection of the contrast agent (t 0 ) and repeated at least two times in the early (t 1 ) and late phases (t 2 ) after contrast injection. The resulting temporal sampling of the center of k-space for the first contrastenhanced phase was between 2 min 15 s and 2 min 30 s, providing image contrast most representative of this time point. An interphase delay between the first and the second contrastenhanced phases was used as needed to result in temporal sampling of the second contrast-enhanced phase between 7 min 15 s and 7 min 45 s.
Despite the long scan duration, the first postcontrast time sample occurred at 2.5 min using the standard k-space sampling, which was close to the effective sampling of 3 min or less that was recommended by the American College of Radiology guideline for breast MRI. 14

Percent Enhancement and Signal Enhancement Ratio Analysis
For breast MRI, contrast enhancement kinetics, as captured by the signal intensity-time curves are used to distinguish among malignant ( Fig. 1 red curve: rapid rise and signal washout), normal, and benign tissues ( Fig. 1 blue and green curves: slow rise, little or no washout). In this signal intensity-time curve, S 0 , S 1 , and S 2 are the signal intensity values in the precontrast (t 0 ), early postcontrast (t 1 ), and late postcontrast phases (t 2 ), respectively.
] is a method developed to measure contrast enhancement kinetics from high spatial resolution, low 5, contrast-enhanced MR images commonly used for clinical breast MRI. 15 High SER values consistently identify malignant tissues with a strong signal washout characteristic. 16

Proximity Mapping of Breast Stroma
In this study, contrast-enhanced MRI was used to assess lowlevel contrast enhancement patterns in the fibroglandular breast tissue outside of the primary tumor. As previously described, 4 tumor regions on MR images were identified using an enhancement criteria of 70% based on visual agreement with radiological assessments in clinical practice 17 and were applied to the first postcontrast image. 18 Normal-appearing stroma surrounding the tumor was defined as fibrogladular tissues and was segmented from adipose tissue using a fuzzy C-means clustering method. 19 Maps of PE and SER were generated using a customized software program. 18 The tumor proximity map for normalappearing breast stroma was generated as shown In Fig. 2 using the 3-D tumor and stromal tissue masks. For each pixel in the stroma, the proximity was defined as the minimum 3-D distance to any tumor pixel. The resulting map is shown as the green overlay in Fig. 2(c). Proximity analysis regions were defined from the proximity map as 5-mm-thick 3-D distance bands from 0 to 40 mm outside of the tumor mask (0 to 5 mm, 5 and 10 mm, etc.), and were applied to the PE and SER maps to calculate the mean PE and SER values within Journal of Medical Imaging 011014-2 Jan-Mar 2018 • Vol. 5 (1) each distance band. Regions closest to the tumor boundary from 0 to 5 mm were considered as tumor periphery. 20 Global PE and SER values were measured as the average of the mean PE and SER values over 5 to 40 mm, where the tumor periphery was excluded to minimize effects of the tumor masking PE threshold. All subsequent calculations of stromal effects on radial distance, global PE, and SER were focused at regions from 5 to 40 mm from the tumor edge.

Stromal RPPA Analysis
For the analysis of stromal protein levels, breast cancer frozen biopsy specimens in this patient cohort were collected at T2 (N ¼ 68) and were subjected to laser capture microdissection (LCM) to enrich for stroma. Stromal populations were isolated (>95% purity) from 8 μm cryosections using an Arcturus Pixcell IIe LCM system (Arcturus, Mountain View, California) as described previously. 21 Microdissected material was lysed in extraction buffer [tissue protein extraction reagent (ThermoFisher), 2× SDS-PAGE sample buffer (ThermoFisher) mixed 1:1 and 2.5% beta-mercaptoethanol] at a ratio of ∼175 laser pulses∕μL of extraction buffer. Samples were heated at 100°C for 5 min, brought to room temperature, briefly centrifuged, and stored at −20°C until printing. Cell lysates were printed in triplicate spots (∼10 nL per spot) onto nitrocellulose-coated slides (Grace Biolabs, Bend, Oregon) using an Aushon 2470 Arrayer (AushonBiosystems, Billerica, Massachusetts). Standard curves of control cell lysates were also included for quality assurance purposes. 22 Total protein levels were assessed in each sample by staining with Sypro Ruby Protein Blot Stain (Invitrogen, Carlsbad, California) according to manufacturer's instructions. Arrays were immunostained with 59 antibodies specific for various stroma-related proteins that were selected based on their involvement in mediating angiogenesis, immune response, inflammation as well as more general pathways such as cell proliferation, motility, and survival (Table 1). All antibodies were validated before use by immunoblotting. 23 Immunostaining was performed as previously described. 24 Briefly, each slide was probed with one primary antibody targeting the protein of interest (Table 1). Biotinylated goat anti-rabbit (1:7500, Vector Laboratories Inc., Burlingame, California) and rabbit anti-mouse (1:10, DakoCytomation, Carpinteria, California) IgG were used as secondary antibodies. Signal amplification was performed using a tyramide-based avidin/biotin amplification system (DakoCytomation, Carpinteria, California) followed by streptavidin-conjugated IRDye 680 (LI-COR, Lincoln, Nebraska) for visualization. Images were acquired using a TecanPowerScanner (Tecan, Mannedorf, Switzerland). Antibody staining intensities were quantified using the MicroVigene v3.5.0.0 software package (Vigenetech, Carlisle, Massachusetts). The final results represent negative control-subtracted and total protein-normalized relative intensity values for each endpoint within a given patient sample. 25

Statistical Analysis
Patient characteristics were compared between I-SPY 1 TRIAL subjects with and without MR stromal analysis using Wilcoxon rank-sum test or Fisher's exact test as appropriate. Similar comparisons were made between subjects who had 3-year  All statistical analyses were performed using the R statistical software language. 27 3 Results

Patient and Tumor Characteristics
As shown in the consort diagram (Fig. 3), there were 221 evaluable patients enrolled in the I-SPY 1 TRIAL. Seventyone patients had stromal MR evaluation and 144 did not. Characteristics of these patients (total I-SPY 1 versus stromal MRI evaluated) are shown in Table 2. The two groups were similar in age, tumor size, nodal status, subtype, and regimen received. However, notable differences were found in were also significantly associated with RFS. Prediction of RFS by other known prognostic factors including age and chemoresponse were not statistically significant: details of the estimated relationships between these patient tumor characteristics and RFS in terms of hazard ratios, confidence intervals, and p-values can be found in Table 2.

Stromal Enhancement Patterns
As described above, we developed the tumor proximity technique to map the stromal enhancement at incremental distances from the primary tumor edge as visualized on MRI. Using retrospective pilot cohorts, we previously showed that MR PE and SER had higher stromal contrast enhancement patterns in patients with invasive breast cancer. 3,4 Here, we validated our previous findings using the ISPY 1 TRIAL cohort and investigated the association of stromal enhancement patterns with treatment.

Radial dependence and interaction with RFS
As shown in Fig. 4, PE values in the I-SPY 1 dataset decreased continuously with distance from the tumor edge: from proximal tissue (within 5 to 10 mm) to distal tissue (35 to 40 mm). Notably, the decreases in PE with distance from the tumor edge remained consistent through all the study time points, T1 to T4 (Fig. 4).
Next, 3-year RFS, recurrent (N ¼ 17) versus nonrecurrent (N ¼ 51) was added to the mixed model. The same decreasing radial trend was found in both groups as shown in Table 4 Patients who missed an MR evaluation (N=18) Fig. 3 Consort diagram for the study. Patients were accrued to ISPY 1 (n ¼ 237). Of these, 221 were available for analysis, and 71 patients had MRIs that were assessable for stromal enhancement. The study protocol was to obtain four MRIs for each patient at V1 (prior to chemo); V2 (after first chemocycle); V3 (between AC and T chemotherapy); V4 (prior to surgery).
Journal of Medical Imaging 011014-6 Jan-Mar 2018 • Vol. 5(1)  For the recurrent group, significance change was found only in T2 (estimated mean change permm ¼ −0.39, 95% CI: −0.75, −0.02; P ¼ 0.04), but the lack of significance in the recurrent group may be due to the small sample size. The estimated mean difference in distance effects between nonrecurrent and recurrent patients from time points T1 to T4 were −0.01, 0.03, −0.28, and −0.01, suggesting that the decreasing rates of PE from the tumor edge in these groups were similar with the exception of T3. However, limited numbers preclude making a more definitive statement. We also examined changes in PE between adjacent times (T1 versus T2, T2 versus T3, T3 versus T4 in the recurrent and nonrecurrent groups) at the tumor periphery within the nonrecurrent and recurrent groups. In general, the PE signal was lower in the later time point, with the only exception being between T1 and T2 in the nonrecurrent group. The difference was significant only between T2 and T3 in the recurrent group. Here the estimated (pseudo-) median difference was −3.23 per mm (95% CI −6.17, −0.80; Wilcoxon signed-rank test P ¼ 0.01).

SER pattern from the tumor edge and interaction with RFS
Using the same mixed effects model as for PE, we observed an increasing trend, rather than a decrease, in mean SER as a function of distance from the tumor edge at the pretreatment time point (T1). Mean SER values increased significantly at T1 only. Here the estimated mean increase per mm was 0.004 (95% CI: 0.001, 0.006; P ¼ 0.003) (Fig. 5). In MRIs taken at time points during and after chemotherapy (T2 to T4), Table 3 Change of stromal PE values per mm from the tumor edge at time points before the start of chemotherapy (T1), after one cycle of AC (T2), between AC and taxane-based (T3) regimen, and at the completion of chemotherapy prior to surgery (T4).
Estimated mean change (per mm) 95% CI P-value changes in SER as a function of distance were not significantly different from zero, Fig. 5. Similar to PE, we added the recurrence status as a group variable to the mixed effects model with distance to assess differences in SER enhancement patterns in recurrent and nonrecurrent patients. We focused on T1, since that was the only significant time point. The estimated mean increase (95% CI) in the nonrecurrent group was 0.0046 (95% CI: 0.0017, 0.0074; P ¼ 0.002) and in the recurrent group was 0.0012 (95% CI: −0.0040, 0.0063, P ¼ 0.66).

Global stromal PE and SER and chemotherapy effect
To determine whether the intrinsic stromal enhancement averaged throughout the breast stroma could be associated with  outcomes, we measured global PE at each time point, T1 to T4. Although global PE was not significantly associated with RFS at any time point, we observed a decrease in global PE among the cohort from T1 to T4 (Fig. 6). As shown in the scatter plots, when we examined the PE values at the pretreatment time point (T1), there was no change in PE after the first dose of chemotherapy at T2. However, PE significantly decreased at the interregimen (T3) and pre-surgery (T4) time points relative to pretreatment levels (P < 0.001 for each comparison), suggesting that ongoing changes in global PE occurred with increasing exposure to chemotherapy. Of note, we had previously found that global SER had predictive values among a pilot cohort of patients treated with NAC. 3,4 In this paper, a similar analysis using global SER values did not reveal significant differences between T1 and T4, or significant associations between SER and RFS.

Global Stromal PE and SER and Stromal
Protein Levels Using unsupervised clustering of stromal protein and protein signaling architecture from the RPPA data obtained at T2, two major subgroups were identified (Fig. 7).

Discussion
The complex changes associated with cancer-activated stroma reflect dynamic processes of tissue remodeling inherent to tumor progression. Indeed, several features of the cancerous stroma have been found to have prognostic value and have relevance for therapy resistance. 29 In part, to develop and translate specific stromal properties as predictors of outcomes, more sophisticated tools to assess stromal phenotypes need to be developed. Our previous work was the first to describe lowlevel quantifiable enhancement characteristics of breast cancer stroma. [3][4][5] As the technical tools have been refined, algorithms to quantify stromal enhancement, to a large degree, have become automated. Using refined techniques, we found that stromal enhancement decreased with exposure to chemotherapy in the 71 patient cohort from I-SPY 1. In addition, we validated previous findings that PE maintained a decreasing pattern with radial distance. These data further validate our previous findings in this cohort from the I-SPY 1 TRIAL. PE measures the signal enhancement in tissues reflecting the vascular permeability. We previously observed that microvessel density was associated with increased PE in the periphery of tumors, consistent with increased delivery of gadolinium contrast to the extravascular space 5 and decreased radially with distance from the tumor edge to the surrounding stroma. In the present cohort, we again found a radial dependence of PE from the tumor edge; interestingly, this dependence was maintained through all study time points (T1 to T4). This finding suggests that some degree of vasculature is maintained in tissues proximal to the tumor despite therapy. In the present cohort, due to the limited sample size, we were not able to correlate residual disease or RCB 28 class with stromal enhancement characteristics. The question of whether stromal enhancement characteristics resolve in patients who achieve pathologic complete response remains interesting and relevant for future investigations.
Given the previously established correlation between microvessel density and increased PE, it was surprising that no correlations in this study were identified between MR enhancement measurements and the protein/phosphoproteins measured. This result may be explained by: (1) the small sample size in which data for both MR and protein measurements were available; (2) differences in MR enhancement attributed to protein/ phosphoprotein levels that were not measured in this study; and (3) the average numeric PE and SER descriptors may be limited in describing the regional variation of stromal behaviors. Recent advances in radiomics-based analysis have demonstrated the power of transforming imaging data into multidimensional mineable radiologic features 30,31 that are relatable to tissue molecular characteristics and disease prognosis. [32][33][34] Future studies will assess biological differences between stromal specimens with radiomic-based analsysis of MR enhancement patterns in larger cohorts. Radial dependence of protein/ phosphoprotein levels will also be investigated.
In addition to measuring radial dependence of PE, we also evaluated the average PE of the entire stroma or global stromal PE at each time point. In the present cohort, global stromal PE decreased with the increasing exposure tochemotherapy from T3 to T4. These findings suggest that the average stromal vascular permeability decreased with chemotherapy, suggesting there are global effects on stromal vasculature that are of unknown significance.
We also measured SER values, which in contrast to PE, reflect washout kinetics of the gadolinium contrast. In our previous studies, we found an association with higher global SER values at posttreatment time points with improved RFS. We did not find this association in the present cohort. However, we observed a small increasing trend in mean SER with distance from the tumor edge at T1. We observed similar associations in our pilot cohort. 4 While both studies were limited by numbers and events, findings were consistent.
We recognize that the current study population was limited to a cohort of patients enrolled in the I-SPY 1 neoadjuvant trial. Due to relatively small numbers and events, limited conclusions could be drawn from this cohort. However, the findings in this study are consistent with observations made in previous cohorts, namely the radial dependence of PE and SER in the stroma. A new finding is the persistence of the radial dependence of PE despite the treatment with chemotherapy, and the decrease in global PE with the increasing exposure to chemotherapy from T3 to T4.
Enhancement threshold plays a significant role in defining tumor regions and subsequent stromal analyses. In the past, we found that functional tumor volume defined by PE threshold set at 70% has the strongest prediction of RFS 17,35 and was used to define the tumor region and subsequent segmentation of the stoma in this study. However, this threshold may be unique to each breast cancer subtype and is a subject of our ongoing research. 36,37 We recognize that the current study relied on measurements of average numeric global PE and SER descriptors that may over simplify the behaviors of the tumor microenvironment. However, the current proximity mapping technique uniquely accounts for tissue vascular permeability in all voxels by their location relative to the closest tumor voxel. Ongoing efforts to refine the proximity analysis and integrate methods with Journal of Medical Imaging 011014-11 Jan-Mar 2018 • Vol. 5 (1) radiomic-based analysis may improve the predictive performance for survival. While PE and SER will remain the primary imaging parameters for stromal analysis in DCE-MRI, the proximity mapping methodology can also be extended to diffusionweighted imaging. Other descriptors such as apparent diffusion coefficient 38 or fractional anisotropy (FA) 39 measurements that depict tumor cellularity and water mobility may provide additional important information for breast stromal characterization.

Disclosures
No conflicts of interest, financial or otherwise, are declared by the authors.