Diffuse optical spectroscopic imaging reveals distinct early breast tumor hemodynamic responses to metronomic and maximum tolerated dose regimens

Specific details about the DOSI instrumentation have been well described elsewhere [4]. Briefly, DOSI uses near-infrared light (650–1000 nm) to measure deep tissue functional information with a handheld probe. For this study, fiber-coupled temporally modulated laser diodes (659, 689, 781, 829 nm or 658, 682, 785, 810, 830, 850 nm) and a broadband near-infrared light source (650–1000 nm) were used to illuminate the tissue from the skin surface. An avalanche photodiode was used to detect the remitted temporally modulated laser light and a spectrometer was used to measure the remitted broadband light. The lasers were frequency swept from 50 to 400 MHz; the amplitude and phase perturbations induced by the tissue were measured with a vector network analyzer or custom analog electronics [39]. The amplitude and phase at each wavelength and modulation frequency were fit to an analytical solution to the P1 diffusion approximation of the Boltzmann transport equation solved in the frequency-domain with semi-infinite boundary conditions [24]. This information was combined with broadband diffuse reflectance measurements to yield broadband optical absorption and reduced scattering properties [4]. Absolute concentrations of oxyhemoglobin (HbO₂), deoxyhemoglobin (HHb), water, and lipid were then determined by a least squares fitting procedure of the broadband absorption spectra to the known extinction coefficient spectra of these four chromophores according to Beer’s law. Additional composite metrics were also computed including total hemoglobin (THb = HbO₂ + HHb), oxygen saturation (StO₂ = HbO₂ /THb), and tissue optical index (TOI = HHb × water/lipid). All data analysis was conducted using custom processing codes developed in MATLAB R2014b (MathWorks Inc).

DOSI scans were conducted prior to chemotherapy administration (baseline) and as many days as possible within the first week of starting therapy dependent on patient availability and health status. A standardized DOSI measurement protocol was used to measure the patients [48]. Briefly, a sequential rectangular grid pattern with 1-cm spacing was transferred to the tissue using a transparency and surgical marker. The DOSI probe was placed against the breast and a measurement was taken at every point on the grid. Landmarks such as the nipple, areola, freckles, and moles were used to coregister longitudinal measurements. The size of the rectangular grid was chosen to fully include the tumor region with clear margins as determined by prior standard of care imaging and palpation. An example of a measurement grid is shown in Fig. 1 on the left breast using a 3D bust model. An interpolated heatmap of the TOI composite metric in shown over the right breast over a 3.7-cm tumor. The tumor and areola both show significant TOI contrast relative to the surrounding breast tissue, a highly conserved feature seen across many breast cancer subjects [7].

2D heatmaps of each chromophore (i.e., HbO₂, HHb, water, lipid) and composite metrics (i.e., THb, StO₂, TOI) were generated for each subject and at each time point. A tumor region of interest (ROI) was determined using a combination of peak TOI contrast, as described by previous studies [6, 7, 48], and tumor size was determined by ultrasound or MRI. The tumor ROI size remained constant throughout the longitudinal measurements but was allowed to laterally shift in cases of grid displacement. Tumor chromophore concentrations were calculated by taking the mean over the tumor ROI.

A total of 53 female breast cancer patients were measured at Boston Medical Center (BMC) and University of California, Irvine (UCI) from May 2004 to October 2017. Subject’s ages ranged from 26 to 71. One patient had bilateral breast cancer with both tumors measured for a total of 54 DOSI monitored tumors. Eligible subjects had a diagnosis of invasive breast cancer and were planned for neoadjuvant cytotoxic chemotherapy with definitive breast surgery following therapy. Each patient’s treatment regimen was determined by their oncologist; a subset of subjects was simultaneously enrolled in an investigational protocol testing MET dosing and scheduling (Clinicaltrials.​gov Identifier NCT00618657). The patients received a biopsy prior to treatment to confirm invasive cancer diagnosis, which provided the tumor receptor status (i.e., ER, PR, HER2). Potential subjects were excluded if they were pregnant or previously received treatment to the affected breast. Data from a portion of these 54 subjects had been used in prior published works, including the initial observation of oxyhemoglobin flare [45], which analyzed 24 of the current 54 subjects throughout week 1, including day 1. The remaining 30 subjects’ data at day 1 have not been published before, although 14 of the current 54 subjects were also enrolled in the ACRIN 6691 study [10, 48], which included baseline, early (days 5–10), midpoint, and pre-surgical timepoint measurements. All patients provided written informed consent. This project was approved by the Institution Review Board at Boston University, Boston Medical Center and UCI.

Each patient’s resected tumor was evaluated by the local institution’s pathologist to generate a pathology report. After treatment, the patient was assigned a tertiary response status according to the National Surgical Adjuvant Breast and Bowel Project Protocol [44]: pathologic complete response (pCR), partial response (PR), and no response (NR). pCR was defined as no residual tumor burden in resected tumor. PR was defined as a decrease in the largest tumor dimension by > 50% from diagnosis to resection while NR was defined as a decrease of < 50%. A binary classification of response was also utilized in which responders, defined as subjects achieving either pCR or PR, were compared against NR subjects.

Generalized estimating equations (GEE) were used to longitudinally model the DOSI chromophores throughout the first week after chemotherapy utilizing SAS (SAS Institute) [31]. The GEE accounted for the correlation between multiple measurements on a single patient and allowed for an unbalanced dataset with subjects considered as clusters, an exchangeable correlation structure, and a normal model with an identity link function. Separate models were run on each of four chromophores: HbO₂, HHb, water, and lipid. In addition, for each outcome variable, three separate models were run on population stratifications of treatment schedule and pathologic response to isolate the effects of specific covariates and interaction terms: (1) MTD responders vs MTD non-responders, (2) MET responders vs MET non-responders, and (3) MTD responders vs MET responders. Additional covariates included in these models included as follows: institution (Boston Medical Center vs UC Irvine), age, hormone receptor status (estrogen receptor or progesterone receptor), and HER2 status. Significance for model parameters was determined at a level of 0.05 and when adjusted for multiple comparisons with Bonferroni correction at a level of 0.0125. Additional covariates relating to treatment such as chemotherapeutic agent or drug mechanism were not included as these parameters were highly correlated with treatment schedule (e.g., all patients administered adriamycin + cyclophosphamide received an MTD schedule). Post hoc contrasts between outcome means adjusted for covariates in the statistical models were performed at each day post chemotherapy between strata of interest: (1) MTD: responders vs non-responders and (2) responders: MTD vs MET. Significance for post hoc contrasts, when adjusted for multiple comparisons with Bonferroni correction, was determined at a level of 0.0036.

Linear discriminant analysis (LDA) was performed using MATLAB to assess the prognostic accuracy of percent change in oxyhemoglobin on day 1 among MTD and MET population. This analysis assumed multivariate normal densities and equal covariance for each group. Fivefold cross-validation was used to train and test the dataset and limit overfitting. We note that there was approximately twice as many responders as non-responders in the MTD cohort and no efforts were made to account for this imbalance in the classification analysis. Posterior probabilities were calculated for each subject from the linear classification of responders from non-responders. Receiver operating characteristic (ROC) curves were generated by iterating through all posterior probability thresholds. The area under the curve (AUC) of the ROC curve was utilized to evaluate the performance of the model along with the optimal sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).

The characteristics of all analyzed subjects are shown in Table 1, which shows the overall population (n = 54) statistics as well as characteristics stratified by treatment schedule: maximum tolerated dose (MTD, n = 35) and metronomic (MET, n = 19). The MTD patients received treatment every 2 or 3 weeks at the highest dose without unacceptable side effects per the standard of care. In contrast, MET patients received treatment every week at a smaller dose than MTD. The average overall subject age was 49.9 ± 11.1 years with a slightly higher age among the MTD subset, 51.3 ± 11.1 years and slightly lower in MET subset, 47.4 ± 10.8 years. The average tumor size was 3.5 ± 1.7 cm for all subjects, 3.8 ± 1.9 cm for MTD subjects, and with 3.0 ± 1.3 cm MET subjects. Age and tumor size were not significantly different (p > 0.05) between the MTD and MET cohort through the Wilcoxon rank sum test using MATLAB, utilizing a nonparametric test to avoid any assumptions of the underlying distributions.

All patients received a diagnosis of invasive carcinoma with the majority of patients (91%) receiving a diagnosis of invasive ductal carcinoma (IDC). In total, 65% of subjects had hormone receptor-positive tumors. Fifteen subjects were diagnosed with HER2-positive tumors; of those, seven received trastuzumab (Tr) as part of their first neoadjuvant treatment and the other eight received Tr at a timepoint which was not in the scope of this study. Within the MTD treatment schedule, one patient received a combination of HER2-targeted therapies with both Tr and pertuzumab (Pzb) along with the cytotoxic agent docetaxel (DTX). One MTD patient received a combination of carboplatin (Cb), DTX, and Tr. Three MTD patients received adriamycin (A) + cyclophosphamide (C) + DTX, while 30 MTD patients received A + C, which is the current standard of care therapy. Among the MET patients, 14 patients received carboplatin (Cb) + paclitaxel (nPTX) + bevacizumab (Bev) and 5 patients received Cb + nPTX + Tr.

At the time of surgery, 35 patients (65%) were determined to be responders while 19 patients (35%) were non-responders. A total of 71% of MTD patients were responders (n = 25) and 53% of MET patients were responders (n = 10). All proportions (location, histology, receptor status, and pathologic response) were not significantly different between MTD and MET cohorts (p > 0.05) using the Z-Test for proportions implemented in MATLAB.

The percent change from baseline was examined to normalize for varying baseline tumor chromophore concentrations among patients, with a focus on the primary aim of day 1 postchemotherapy changes. The observed changes across week 1 are shown in Fig. 2. For subjects receiving MTD treatment, the largest difference between responders and non-responders occurred on day 1. Responders on day 1 demonstrated a mean 40% increase in HbO₂ at day 1 compared to a 13% decrease in non-responders. All MET subjects displayed much smaller changes on day 1: responders with a 3% increase and non-responders with a 1% decrease. These differences are visualized in Fig. 3, which shows representative 2D DOSI heatmaps of HbO₂ concentrations at both baseline and day 1 for two different pCR subjects, one of whom received MTD treatment while the other received MET therapy. The MTD patient had a large increase of 50% in HbO₂ from baseline to day 1 contrasted with the MET patient, which only increased by 0.3%. In addition, there were relatively small changes in both MET responders and non-responders across the entire week 1 as compared to the MTD patients.

Separate GEE models were fit for the MTD and MET subject populations to isolate the effect of treatment and evaluate the effects of age, institution, hormone receptor status, HER2 status, days post chemotherapy, response, and the corresponding interaction terms. Within the MTD cohort, the interaction term of response and day 1 was a significant predictor of HbO₂ (p < 0.0001) while the MET cohort failed to demonstrate a statistically significant predictor. In order to isolate the effect of pathologic response, a separate GEE model was run on the responder population to determine the effects of age, institution, hormone receptor status, HER2 status, days post chemotherapy, treatment schedule, and the corresponding interaction terms. The interaction term of treatment schedule and day 1 was a significant predictor of HbO₂ (p = 0.0008) in responders. Post hoc analysis of outcome adjusted estimates to compare the effect of treatment demonstrated a significant difference in day 1 HbO₂ changes between MTD responders and MET responders (39.45 ± 11.98%, p = 0.0010) seen in Additional file 1: Table S1A. Similar analysis on the efficacy of HbO₂ as a prognostic biomarker at day 1 for patients receiving MTD demonstrated a difference of 48.77 ± 9.51% (p < 0.0001) between responders and non-responders seen in Additional file 1: Table S1B. None of the additional covariates (age, institution, hormone receptor status, HER2 receptor status) contributed significantly to the model.

Separate LDAs were run on the MTD and MET subject populations to evaluate HbO₂ performance on day 1 as a prognostic predictor when isolating the effect of treatment, results shown in Additional file 2: Figure S1. The MTD cohort had an AUC of 0.89, sensitivity of 0.91, specificity of 0.89, PPV of 0.95, and NPV of 0.80. In stark contrast, the MET cohort had an AUC of 0.50, sensitivity of 0.50, specificity of 0.75, PPV of 0.71, and NPV of 0.55.

HHb, water, and lipid did not exhibit significant prognostic or regimen-dependent changes in either MTD or MET patients. GEE analysis was run on these chromophores for each of the three population stratifications and there were no significant interactions of treatment or response at any day. The most notable interaction term was response and day 1 post chemotherapy (p = 0.0136) for water in MTD patients, which came close to significance (p < 0.0125 for Bonferroni correction). At this timepoint, non-responding MTD patients had a moderate decrease in water concentration compared to responders (Additional file 3: Figure S2).