Identifying an early treatment window for predicting breast cancer response to neoadjuvant chemotherapy using immunohistopathology and hemoglobin parameters

The study protocol was approved by institutional review boards and was HIPAA compliant. Written informed consent was obtained from all patients. From March 2014 to June 2016, 28 patients were recruited at three hospitals. All had been referred for NAC to the one of three medical oncologists (PD, ST, and KS) and agreed to participate in our study. Five patients did not complete the study because of a change in their treatment plan or a desire to withdraw from the study. One patient had technically problematic baseline imaging. Data from these six patients were not included in the analysis. The remaining 22 patients (mean age 55 years, range 34–74 years) were repeatedly imaged by NIR/US prior to initiation of NAC, at the end of the first three treatment cycles during chemotherapy, and prior to definitive surgery. Of the 22 patients (Table 1), 11 were HER2⁺ of which eight were also ER⁺; six were ER⁺/HER2^–, and five were TN. Of the 22 patients, 12 had T2 tumors, five had T1, and five had T3 tumors. Patients were treated with regimens based on their tumor biomarkers according to current clinical practice. For ER⁺/HER2^– tumors, patients were treated with dose-dense doxorubicin/cyclophosphamide every 2 weeks for four cycles followed by paclitaxel every 2 weeks for four cycles (ACT). The NIR/US cycle 1 to 3 measurements were performed at the end of the first three cycles before the paclitaxel started. For HER2⁺ tumors, all patients were treated with trastuzumab, pertuzumab, and docetaxel or paclitaxel with or without carboplatin (TPT) every 3 weeks for six cycles and one patient had two additional cycles of 5-flurouracil, epirubicin, and cyclophosphamide (FEC). The NIR/US cycle 1 to 3 measurements were performed at the end of the first three treatment cycles when TPT was given. For TN tumors, three were treated with ACT, the same as ER⁺/HER2^– patients, and two were treated with carboplatin and paclitaxel every 3 weeks for six cycles because of their BRCA1 gene mutation. One elderly ER⁺/HER2^– patient was treated with cyclophosphamide/docetaxel without doxorubicin (TC) every 3 weeks for four cycles.

The HER2⁺ cohort was monitored at an additional time point of 7 days after the first treatment, and one TN and four ER⁺/HER2^– patients were also monitored at an optional time point of 7 days after the first treatment. The median from 7 days was 0 with a range of 0 to 2 days. Moreover, four TN and five ER⁺/HER2^– patients were monitored at an additional time point at the end of cycle 5. Thus, a total of 16 patients had an additional time point at 7 days after the first treatment and nine patients had an additional time point at the end of cycle 5.

All 22 patients were studied after their diagnostic core biopsy with an average interval of 28 days (median 26 days, range 7–56 days). Among the 22 patients, one had pretreatment NIR measurements 7 days after biopsy and the remaining patients had the NIR measurements more than 10 days after biopsy. All patients received the first cycle of NAC after the initial NIR/US study (median 2 days, range 0–10 days). The average interval between post-treatment NIR/US and surgery was 19 days (median 15 days, range 2–67 days). During the treatment, the NIR/US scans were performed before their scheduled chemotherapy (median 0 days, range 0–5 days). Among the 22 patients, 18 patients had pretreatment MRI and 12 patients had post-treatment MRI. Two patients had pretreatment PET.

The histologic type of 19 patients was invasive ductal carcinoma; one patient had invasive lobular carcinoma, and the other two patients had invasive mammary carcinoma with mixed ductal and lobular features. One of the 22 patients had two distinct tumor masses in the same breast, adjacent to each other and with the same histologic characteristics. For this patient, one of the two masses was used for data analysis. Invasive carcinoma within the pretreatment core biopsies was graded using the Nottingham histologic score (NS). ER, progesterone receptor (PR), and HER-2/neu (c-erbB-2) immunohistochemistry was performed on formalin-fixed, paraffin-embedded core biopsy tissue. The ER and PR were scored by a modified San Antonio scoring system [34], where the total score ranges from 0 to 8 (scores 0–2 are negative, a score of 3 is equivocal and scores ≥ 4 are positive). Testing for the HER2 gene was performed by immunohistochemistry and by gene amplification utilizing the fluorescence in situ hybridization (FISH) technique, and the results were reported in accordance with 2014 ASCO/CAP guidelines [35]. Results were reported as equivocal HER2 if there was weak to moderate, incomplete membranous staining in > 10% of cells or if FISH showed a HER2/CEP17 ratio < 2, or if the HER2 copy number was ≥ 4 and < 6. HER2 results were negative if the immunohistochemistry or FISH assays fell below the thresholds for interpretation as equivocal. All assays were performed on pretreatment core biopsy samples.

Pathologic response was assessed by applying the Miller-Payne grading criteria to definitive surgery specimens in comparison with initial core biopsies (Table 1). Two breast pathologists (AR and PH) individually evaluated cases from their respective hospitals and additional cases from the third hospital. The Miller-Payne system [36] divides pathologic response into five grades based on a comparison of tumor cellularity between the pretreatment core biopsy and the definitive surgical specimen. Grade 1 indicates no change or some minor alteration in individual malignant cells but no reduction in overall cellularity; this is a pathological nonresponse (pNR). Grade 2 indicates a minor loss of tumor cells (up to 30%) but with overall cellularity still high; this is a partial pathologic response (pPR). Grade 3 indicates an estimated 30–90% reduction in tumor cells (pPR). Grade 4 indicates a marked disappearance of tumor cells (> 90%), with only small clusters or widely dispersed individual cells remaining (almost pCR). Grade 5 indicates that no malignant cells are identifiable in sections from the tumor bed (pCR). Grade 5 may show that necrosis, granulation tissue, histiocytes, and vascular fibroelastotic stroma remains, often containing macrophages. Residual ductal carcinoma in situ (DCIS) is acceptable for Miller-Payne grade 5.

Ultrasound examinations were performed using a commercial ultrasound system (Phillips IU22 or GE Logiq 5) at the corresponding hospital. Three NIR systems with identical designs were used at the three hospitals, and the details have been given previously [24]. Briefly, the NIR/US probe consists of the commercial US transducer located centrally, with source and detector light guides (optical fibers) distributed around the periphery of the NIR/US probe. Four laser diodes of 740 nm, 780 nm, 808 nm, and 830 nm optical wavelengths were sequentially switched to nine positions on the probe, while the reflected light was coupled by the light guides to 14 parallel detectors. The entire NIR data acquisition interval was less than 5 s. For each patient, US images and optical measurements were acquired simultaneously in the cancer region and a normal region within the contralateral breast in the same quadrant as the cancer. At each cancer and normal region, multiple datasets were acquired. The optical data acquired from the normal contralateral breast was used as a reference for calculating the background optical absorption and reduced scattering coefficients that were used in the image reconstruction of the lesions.

Details of the optical imaging reconstruction algorithm with experimental validation have been described elsewhere [37]. Briefly, the NIR reconstruction takes advantages of the ultrasound localization of lesions to segment the imaging volume into a region of interest (ROI) and background nonlesion regions. Since the spatial resolution of diffused light is poorer than that of US, the ROI is chosen to be at least two to three times larger than the tumor size measured by coregistered US in x to y dimensions. In addition, because the depth localization of diffused light is very poor, a tighter ROI in the depth dimension is set by using coregistered US. For each patient, the same size of ROI as that obtained from the pretreatment US is used for processing all datasets obtained at different treatment cycles. Therefore, the changes in tumor size seen by US during treatment have no major effect on the NIR image reconstruction. Among the 22 patients, two patients had initially palpable tumors with ill-defined and heterogeneous pretreatment ultrasound images. For these patients, tumor sizes estimated from pretreatment MRI measurements were used to assist in determining the US ROI in the x to y dimensions. The ROI in the depth dimension was typically set from the top border of the ill-defined tissue pattern to the chest wall as seen by US.

The optical absorption distribution at each wavelength was reconstructed and the tHb concentration, oxygenated hemoglobin (oxyHb) concentration, and deoxygenated hemoglobin (deoxyHb) concentration maps were computed from absorption maps at the four wavelengths [38]. The maximum values of tHb, oxyHb, and deoxyHb were measured for each set of hemoglobin maps. For each patient imaged at each time point, an average maximum that was obtained from 5 to 10 quality NIR images at the tumor location was used to characterize the tumor. Data with patient motion as evaluated by using two coregistered US images before and after each NIR measurement were excluded from averaging. To assess the response of each patient, the tHb obtained before treatment was taken as the baseline and the percentage normalized to the baseline (%tHb) was used to quantitatively evaluate the remaining tumor vascular fraction during chemotherapy.

Nineteen patients had well-defined tumors visible by US. Tumor sizes were measured by US technologists under direct supervision of attending radiologists. The percentage ratio (%US) of the largest dimension of each post-treatment measurement over the largest dimension of pretreatment measurement was used to evaluate the morphological change during NAC. One patient had a much larger tumor size than the US transducer, and the baseline US measurement was not accurate. The initial tumor mass of this patient measured by MRI was 6.9 cm. US scans performed from both medial lateral and cranial-caudal directions using the 5-cm US transducer were used to estimate the approximate mass center. Then the combined probe of 10-cm diameter was placed at the estimated mass center for NIR data acquisition. This procedure had minimal effect on NIR reconstruction. For the two patients with initially palpable but ill-defined and heterogeneous US images, the tumor location at each measurement was tracked using previous US images as references. The tumor clock position, distance of the tumor from the nipple, and depth below the skin were documented for each case. Additionally, the tumor posterior shadowing and surrounding tissue structures, as well as the metal clip position, were also reviewed and used to help identify the tumor for each subsequent measurement. If an MRI was ordered for clinical reasons, the MRI measurements were obtained from the medical records of the patients.

We have previously developed a logistic regression model to predict the NAC response of a patient using pretreatment tumor clinicopathologic variables, tumor subtype, and baseline tHb values [21]. Briefly, logistic regression is a statistical modeling approach that can be used to describe the relationship of several predictor variables X_1, X₂… X_k to a dichotomous response variable Y, where Y is coded 1 (responder) or 0 (nonresponder) for its two possible categories [39]. The model can be written in the form of the conditional probability of the occurrence of one of the two possible outcomes of Y, as follows:

The estimated outputs (probability) for each set of predictor variables range from 0 to 1. Given the data on Y, X₁, … X_k, the unknown parameters βn, n = 0, 1, …, k_, n = 0, 1, …k can be estimated using the maximum likelihood method.

In this study, we used the data from 32 patients obtained from an earlier study as a training set [20] to estimate a total of four groups of logistic models, and validated these models using the new dataset reported in this study as a testing dataset. The earlier data obtained from 2008 to 2011 were acquired from almost identical NIR systems with the same data processing and image reconstruction procedures as reported in this study.

To validate that the early data and new data are generated from the same population, we have introduced a dummy variable X_k + 1 in Eq. 1 which is coded as 0 for early data and 1 as new data [40, 41]. We have estimated the model with this dummy variable along with the eight predictor variables using the combined datasets. The estimate on β_k + 1 (P = 0.214) is statistically insignificant. Therefore, we can assume both datasets come from the same patient population.

The Matlab (version 2016a) logistic regression function glmfit was used to estimate the coefficients β _n βn, n = 0, 1, …, k, and glmval was used to calculate the receiver operating characteristic (ROC) with these coefficients for the training set. The same coefficients obtained from the training set were used to predict the response for the testing set.

We also assessed the overall performance of the prediction models through the ROC curves and the area under the curves (AUCs) for each pair of training and testing sets of each prediction model. The early data used for training had 20 ER⁺/HER2^– patients, six TN patients, five HER2⁺ patients, and one ER^–/PR⁺/HER2^– patient [20]. Similar to the new patient cohort, ER⁺/HER^– (n = 14) and TN (n = 6) tumors were treated with ACT every 2 weeks for eight cycles. Six ER⁺/HER2^– tumors and one ER^–/PR⁺/HER2^– tumor were treated with ACT (n = 3) or TC (n = 4) every 3 weeks for six cycles. HER2⁺ tumors (n = 5) were treated with trastuzumab and docetaxel or paclitaxel with or without carboplatin (TPT) every 3 weeks for six cycles.

Since the early data did not contain any patients treated with dual HER2 blockade, we have randomly selected six HER2⁺ patients treated with this regimen from the current study and added these six datasets to the training data. For each random selection, the total of patients in training was 38 (32 from the early data and 6 from this study) and testing was 16 from this study. A total of 6 out of 11 random selections result in 462 combinations of paired training and testing datasets for each group of predictors, and the mean of 462 AUC values was used to evaluate the training and testing results of each prediction model. Each pair of training and testing ROCs was generated using a threshold of 0.5. This was used to separate responders (> 0.5) and nonresponders (≤ 0.5) for each prediction model output. We also used the 462 AUC values to construct the 95% confidence interval (CI) for the mean AUC for each model using a binomial formula. These confidence intervals can provide summary information on comparisons of the different models in terms of their AUC values. For example, if model I has a higher mean AUC than model II, and if their corresponding confidence intervals do not overlap, then this is an indication that model I may have a higher prediction power than the model II in terms of the AUC criterion. However, this interpretation should be understood with the caveat that the 462 values are not true random samples.

To select the independent predictors, Spearman’s rho was evaluated between each predictor and Miller-Payne grade and between each pair of predictors. Spearman’s rho is more appropriate for assessing the relationship for both continuous and discrete variables. Note that both training and testing data were combined to assess the predictors and the Spearman’s rhos reported in this section are from the entire cohort of both earlier and new data (Table 2). To compute rho, the tumor HER2, ER, and TN status were coded as: 1 for TN and 0 for otherwise; 1 for HER2⁺ and 0 for HER2^–; and 0 for ER⁺ and 1 for ER^–. Note that 1 presents increased probability of achieving pCR and 0 otherwise.

Both HER2 and ER are highly correlated with Miller-Payne grade (rho = 0.45, P < 0.001; rho = 0.29, P = 0.035) and are independent of each other (rho = 0.01, P = 0.928); thus, they were selected as predictors. TN tumors did not have a significant correlation with Miller-Payne grade (rho = 0.124, P = 0.362). However, TN was selected as a predictor because it is used clinically to characterize this group of patients. Among tumor pathological parameters, NS is a traditional pathological variable used by oncologists to predict response. NS is highly correlated with mitotic counts (rho = 0.82, P < 0.001). Thus, only NS was selected as an independent pathological predictor and used in each HER2, ER, and TN subtypes to predict response.Baseline tHb and %tHb changes measured during first three treatment cycles are highly correlated with Miller-Payne grade ( see Table 2) and were selected to assess the optimal time window to predict response.

A two-sample two-sided t test was used to calculate the statistical significance for comparisons between responder groups, and a difference with a P value of 0.05 was considered significant. The t test was also used to test the difference between AUCs of different prediction models when their 95% CIs overlapped. Minitab 17 software (Minitab, State College, PA) was used for statistical calculations.