Near-Infrared Spectral Tomography for Predicting Residual Cancer Burden during Early-Stage Neoadjuvant Chemotherapy for Breast Cancer

Neoadjuvant chemotherapy (NAC) has emerged as a crucial treatment approach for patients with locally advanced and inflammatory breast cancer, particularly those with HER2-positive and triple-negative subtypes. NAC aims to reduce tumor size before surgery, leading to improved surgical outcomes and the potential for breast conservation (1). Since the delay in definitive surgery that may occur from a complete course of NAC (that can last up to 8 months; ref. 2), prediction of tumor response to NAC during the early stages of treatment is essential for enhancing patient outcomes and long-term survival (3, 4). Response to NAC is predominantly assessed through physical examination, X-ray mammography, and ultrasound (5–7), in current clinical practice. However, these evaluations typically rely on changes in tumor volume that occur secondarily to physiologic variations, and their correspondence with final pathologic responses is only 19%, 26%, and 35%, respectively (8). Other imaging techniques, such as dynamic contrast-enhanced magnetic resonance (MR) imaging (MRI; refs. 6, 9), MR spectroscopy (10), blood oxygen level dependent (BOLD) MRI (11), positron emission tomography (PET; refs. 12–14), and CT (15), have been investigated for assessing response to treatment. However, these approaches involve costly imaging modalities that expose patients to radiation, and/or potential toxicities from contrast agents. Therefore, they are not well suited for longitudinal use during NAC, especially in point-of-care clinics where MRI, CT, or PET is rarely available, if at all.

In contrast, optical technologies like near-infrared spectroscopy (NIRS; refs. 14, 16–18) and photoacoustic CT (19) offer distinct advantages. They do not involve ionizing radiation or the administration of exogenous contrast agents, and do not necessitate expensive instrumentation or specialized facilities.

Optical imaging technologies enable efficient and effective longitudinal monitoring and early prediction of treatment response by capturing biophysical changes in tissue across different compartments, including the vascular, intra- and extracellular matrix compartments. Among these technologies, photoacoustic CT utilizes high-power short pulse laser sources. However, this imaging modality is expensive and raises significant safety concerns limiting its use in standard clinical practice. On the other hand, diffuse NIRS imaging provides a lower-cost alternative with lower safety risks, making it more suitable for routine clinical practice. NIRS systems exhibit higher sensitivity to subtle tissue changes, particularly in deep tissue regions by capturing signals from capillaries. In contrast, the signal obtained by photoacoustic tomography is more dominated from arterioles and venules (20).

In previous studies, NIRS has shown potential in detecting changes in tumor total hemoglobin concentration [HbT, the summation of deoxy-hemoglobin (Hb) and oxy-hemoglobin (HbO)], blood oxygen saturation (StO₂, percentage of HbO in HbT) and water content (H₂O) before morphologic alterations in tumor size, become evident through structural imaging. These NIRS measurements, taken before the start of the second cycle of NAC, have demonstrated the ability to differentiate between pathologic complete response (pCR) and incomplete response (pIR) with an AUC of 0.92 (14, 16–18). While pCR is a well-established indicator of long-term survival, it is limited in providing prediction information among patients with varying amounts of residual disease (21, 22). The concept of residual cancer burden (RCB) score/class complements the binary assessment of pCR by considering the range of pathologic responses that encompass different prognostic groups, from near-complete response to resistance (23, 24).

In this study, we aimed to assess the feasibility of utilizing a portable near-infrared spectral tomography (NIRST) for predicting RCB in the early stages of NAC based on imaging data obtained from 35 patients with breast cancer. HbT was derived from reconstructed NIRST data, and the normalized early change of HbT in the tumor region (ΔHbT%) from pretreatment to the end of the first cycle of NAC was utilized to predict RCB classes and scores determined after NAC completion.

Imaging was conducted using a custom developed 12-wavelegnth multispectral NIRST system, which has been extensively described in previous publications (25). Here, we provide a summary of the key aspects of the technology. The system utilized photomultiplier tube (PMT, H9305–3, Hamamatsu) and photodiode (PD, C10439–03, Hamamatsu) detection modules to simultaneously acquire frequency domain (FD) and continuous wave (CW) optical measurements. The FD source module consisted of six laser diodes (661 nm, 761 nm, 785 nm, 808 nm, 826 nm and 850 nm), which were modulated by high frequency (∼100 MHz) signals generated from a multi-channel RF synthesizer (HS2004, Holzworth Instruments). The CW source module consisted of six laser diodes (850 nm, 905 nm, 915 nm, 940 nm, 975 nm, and 1,064 nm), and was modulated by low-frequency sinusoidal signals generated directly from a laptop through a data acquisition board (USB 6255, National Instruments). The light from the FD and CW source modules was combined using two 6 × 1 fiber combiners, and then delivered to the breast surface at different positions through sixteen 2 × 1 bifurcated fiber bundles. The common ends of these fiber bundles were held on an adjustable elongated fiber-breast interface while the bifurcated ends were held in a rotating stage. During tomographic optical data acquisition, one of these 16 fiber bundles was coupled to a pair of FD and CW sources, while the transmittance/reflectance light from the breast was collected by the other 15 fiber bundles and transported to each of the photomultiplier/photodiode pairs to capture both FD and CW signals. This simultaneous measurement scheme allowed for efficient data acquisition, with a complete tomographic set of 240 (16 × 15) source-detector combinations at 12 wavelengths taking approximately 90 seconds.

The image reconstruction process was performed using NIRFAST, an open-source software platform (https://milab.host.dartmouth.edu/nirfast/; ref. 26). A fixed regularization parameter of λ = 0.5 was employed for image reconstruction, enabling the recovery of the hemoglobin, water, and scatter images.

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional research committee and with the Declaration of Helsinki. All subject imaging was conducted in accordance with a protocol approved by the Committee of the Protection of Human Subjects (CPHS) at Dartmouth-Hitchcock Medical Center. Written informed consent was obtained from each subject. Subjects were provided with a comprehensive explanation of the nature of procedure.

The study involved 35 subjects, ranging in age from 24 to 64 years. These subjects were enrolled either during 2007 to 2012 (16 subjects, cases 1–9 and 17–23 in Supplementary Table S1) or 2016 to 2022 (19 subjects, cases 10–16 and 24–35 in Supplementary Table S1). All subjects received the best practice NAC recommended by their medical oncologist at the time. Typically, for the subjects treated during 2007–2012, various taxane/anthracycline chemotherapy regimens were administered, following the guidance of CALGB 49909/NCCTG 9831 (27). For the subjects treated during 2016–2022, a Dose-Dense AC-T regimen was typically administered every two weeks for Her2-negative breast cancer. This regimen combines doxorubicin (A) and cyclophosphamide (C), followed by paclitaxel (T). In the case of Her2-positive cancer, the combination of docetaxel (T), carboplatin (C), trastuzumab (H), and pertuzumab (P) was typically administered every three weeks (28, 29).

Baseline NIRST imaging of both the ipsilateral and contralateral breasts was performed within a 10-day window prior to the initiation of therapy. During each NIRST breast imaging session, the subject was seated in a reclining chair positioned at an angle of approximately 100–160 degrees for comfort. The angle of the chair, the orientation of the subject's arm, the clock-face position and the distance from the nipple of the fiber-breast interface were documented in an exam data sheet to ensure consistency in tumor imaging positioning during subsequent imaging sessions. Additional NIRST imaging was conducted on the last day of first cycle of treatment (just before the start of the first infusion of second cycle), and also prior to the initiation of the first infusion of second part treatment (mid-point). Following the completion of all NAC cycles, NIRST images were acquired within a 10-day window before surgery. The position of the tumor was marked based on clinical radiologic reports obtained prior to the treatment, and the fiber array was adjusted accordingly. The separation between the two curved breast interface plates was measured and recorded to generate a patient-specific mesh for image reconstruction. Both breasts were imaged and the total exam time, including subject and breast positioning, typically did not exceed 10 minutes.

The assessment of pathologic RCB scores and classes was performed on the post-NAC surgical excision specimens. The Residual Cancer Burden Calculator developed by MD Anderson Cancer Center (http://www3.mdanderson.org/app/medcalc/index.cfm?pagename=jsconvert3) was utilized for this purpose. Pathologic features examined included the size of the primary tumor bed area, the overall cancer cellularity, the percentage of residual cancer that was invasive and in situ, the number of axillary lymph nodes containing metastatic carcinoma, and the diameter of the largest metastasis in an axillary lymph node (30). The RCB score is a continuous variable but specific cutoffs were used to classify cases into RCB classes. The cut-off values for the RCB classes were determined as follows: RCB-0 (RCB score = 0); RCB-I (0 < RCB score |$ \le $| 1.36); RCB-II (1.36 |$ \lt $| RCB score |$ \le \ $|3.28); RCB-III (RCB score |$ > $|3.28). RCB-0 is equivalent to a pCR and higher RCB scores are significantly associated with worse disease-free survival (21). The RCB assessment was performed by an author (K.E. Muller), who is a surgical pathologist with 7 years of expertise in breast pathology.

To define the tumor region of interest (ROI), the HbT distribution in the NIRST images was analyzed at each time point. The software MATLAB (MathWorks) was used to locate the position of maximum HbT in the suspicious area. The ROI was automatically calculated as the area with HbT values higher than 80% of the maximum value. The HbT values in ROI were normalized by the average of the entire contralateral breast imaged before the treatment. This approach helps minimize interpatient differences and allows for longitudinal data analysis over the course of therapy, as previous studies have demonstrated (31).

To calculate the mean and SD of the normalized changes in HbT within the ROIs (ΔHbT%), the reconstructed images obtained pretreatment and postcycle 1 treatment were compared. Given the limited sample size within each RCB subgroup, which may lead to insufficient statistical power for pairwise comparisons and interpretations, we used logistic regression of hemoglobin on RCB to predict the probability of different RCB class dichotomies based on ΔHbT%. Specifically, we predicted RCB-0 versus RCB-I/II/III, RCB-II versus RCB-0/I/III, and RCB-III versus RCB-0/I/II based on changes in HbT levels. To comprehensively evaluate and quantify the performance of the logistic regression model in classifying the RCB classes, we plotted receiver operating characteristic (ROC) curves to visualize the diagnostic performance. In addition, we calculated the area under the ROC curve (AUC) to assess the accuracy of the model's predictions. Furthermore, Pearson correlation coefficients were calculated to assess the correlation between ΔHbT% and RBC score or class.

The summary data generated in this study are available within the article and its Supplementary Data. All detailed imaging data are available upon reasonable request from the corresponding author.

Figure 1 shows representative images T2 MRI (post contrast, in axial, sagittal and coronal views) and NIRST from pretreatment and postcycle 1 for each RCB class. The HbT images were obtained from ipsilateral breasts before treatment and on the last day of the first cycle of NAC, as well as from contralateral breasts during the pretreatment imaging session. Table 1 summarizes the clinical information, RCB score/class, and HbT and ΔHbT% values for these four cases. All cases had invasive ductal carcinoma (IDC), with case 1 also associated with ductal carcinoma in situ (DCIS). Mastectomy was performed after completing NAC for these patients. In pretreatment images of these four cases, average HbTs in the cancer areas were 17.0 μmol/L, 38.8 μmol/L, 30.1 μmol/L and 50.3 μmol/L, respectively, while average values in the surrounding normal tissue areas were all in the 10–25 μmol/L range. In Fig. 1A, the pretreatment MRI and NIRST images of the ipsilateral breast clearly show a 3-cm mass. The ΔHbT% decreased by approximately 48.5% during the first cycle of NAC, and almost no contrast remained in the cancer area in the postcycle 1 image. Pathologic analysis confirmed a RCB-0 (pCR) result. For cases shown in Fig. 1B–D, HbT contrasts were visibly increased, and the corresponding ΔHbT% were 12.0%, 32.3%, and 75.3%, respectively. Surgical pathologic results confirmed RCB-I (0.78), RCB-II (2.86) and RCB-III (3.58) classes (with scores) for these cases, as indicated in Table 1.

The number of patients confirmed pathologically as RCB-0, RCB-I, RCB-II and RCB-III were 15, 1, 9, and 10, respectively. Figure 2 presents box plots of ΔHbT% of each RCB class. Negative and positive ΔHbT% values indicate a decrease and increase in HbT in the cancer area after the first cycle compared with the pretreatment estimate for each subject. The results in Fig. 2 demonstrated that ΔHbT% is an excellent biomarker for differentiating between each pair of RCB classes. The P values for differentiating RCB-0 from RCB-II or III were <0.0001, while the P value for differentiating RCB-II from RCB-III was <0.04. The P values for distinguishing RCB-I from the other RCB classes could not be obtained given only a single RBC-I case occurred.

The average cancer HbT in pretreatment and postcycle 1 images, HbT in the entire pretreatment contralateral breast, and normalized ΔHbT% (post cycle 1) in each RCB class are summarized in Table 2.

Figure 3 shows box plots and ROC curves depicting the performance of ΔHbT% in distinguishing RCB-0, or RCB-II or RCB-III from all cases in the other classes. As shown in Fig. 3A and B, P values < 0.001 along with AUC and accuracy values of 0.97 and 0.94, 0.90, and 0.83, respectively indicate that ΔHbT% can effectively and accurately differentiate RCB-0 or RCB-II from all cases in the other classes. Supplementary Table S2 provides information on the average cancer HbT in the pretreatment and postcycle 1, HbT in the entire pretreatment contralateral breast, and normalized ΔHbT% post cycle 1 in each RCB class, as compared with the corresponding values in cases from other classes.

Table 3 displays the Pearson correlation coefficients between RCB scores or RCB classes of the 35 patients and several variables including patients’ age, BMI, estrogen receptor (ER), progesterone receptor (PR), Her 2, as well as ΔHbT%. The Pearson correlation coefficients provide insights into the degree of correlation between these variables and RCB scores or RCB classes. On the basis of Table 3, the correlation coefficient for ΔHbT% stands out as the only variable that exhibits moderate correlation with RCB score or RCB class. In contrast, the other variables listed in Table 3 show no significant correlation with RCB score or RCB class.

Compared with other breast cancer imaging modalities and other NIR imaging tools that require contrast injection (13, 32) and/or can only differentiate pCR from pIR to NAC in 1–3 months (5, 16, 33), the results from this study involving 35 patients demonstrate statistically significant mean differences in ΔHbT% for RCB-0, RCB-II, and RCB-III cases. Because RCB is significantly associated with event-free survival within each breast cancer subtype (21, 30, 34), the identification of ΔHbT% as a prognostic and predictive biomarker of RCB could potentially lead to chemotherapy deescalation trials. This approach aims to minimize chemotherapy toxicities in specific breast cancer subtypes that exhibit a strong response to the first cycle of chemotherapy.

Another notable advantage of utilizing the NIRST system for monitoring and predicting breast tumor response to NAC is its portability and fast data acquisition. The imaging process can be conveniently performed while the patient is comfortably seated in a reclining chair and waiting for an infusion to start. Moreover, the data acquisition itself is relatively quick, taking approximately 90 seconds to complete (25, 35).

In addition to the imaging techniques mentioned previously, Jiang and colleagues utilized BOLD MRI to image patients undergoing NAC while applying a 6-minute oxygen breathing challenge (11). In their study, the authors observed a significantly higher BOLD response to oxygen breathing in patients who achieved a pCR based on the BOLD MRI obtained before starting NAC. However, other experiences in utilizing BOLD MRI with oxygen breathing for breast imaging (36, 37), as well as a pilot study on assessing dynamic vascular changes in breast cancer with NIRST using targeted manipulations of inspired end-tidal partial pressure of oxygen and carbon dioxide (38), highlight several difficulties with using an oxygen challenge. The first challenge is ensuring complete closure of the mask. Inadequate mask closure due to variations in subjects’ facial shapes leads to air leakage and inconsistent administration of oxygen dosage. Another difficulty lies in the administration of oxygen, itself, which can cause hyperventilation, a decrease in arterial blood partial pressure of carbon dioxide, and vasoconstriction. These effects counteract the expected increases in BOLD response in tissue. Furthermore, each subject may exhibit varied responses to an increase in oxygen breathing. These difficulties limit the accurate/reliable determination of tumor BOLD response due to an oxygen breathing challenge (37–40).

Several studies have been conducted to observe changes in tumor vasculature during chemotherapy infusion (41) and to explore the presence of a flare response, which refers to a significant alteration in the magnitude and spatial extent of hemoglobin in responders within 1 day of NAC initiation (42). The observed changes in tumor vasculature within 24 hours of treatment initiation are likely occurring during the infusion, when the dominant chemotherapeutic bolus is circulating (42, 43). These changes are believed to originate from a rapid decrease in cellular metabolism caused by the cytotoxic NAC-induced apoptotic activity (i.e., cell damage and subsequent blood oxygenation change). In addition, tissue perfusion changes may result from an acute inflammatory response triggered by cell damage and death. The NIRST system employed in this study presents an opportunity to investigate the vascular reactivity during infusion and provide valuable insights into these biological processes.

In Fig. 1A, it can be observed that the tumor in the NIRST image appears to be shifted toward the center of the breast compared with its position in the MRI image. This shift is likely attributed to the modest compression applied during the NIRST imaging process to ensure optimal contact between the fiber bundles and the skin. In future work and instrument design, it would be beneficial to consider this effect and develop methods to minimize or account for it.

In this study, a comparison between RCB-0/I and RCB-II/III groups was not conducted because only RCB-I outcome occurred, which was identified as triple-negative breast cancer (TNBC). Previous studies have reported significant differences in clinical outcomes between RCB-0 and RCB-I groups in TNBC cases (21). Furthermore, the presence of residual disease is recognized as a risk factor for recurrence in TNBC or HER2-positive subtypes (28, 29).

As depicted in the boxplot of Fig. 3, statistically differentiating RBC-III data from the other groups using ΔHbT% as an indicator during the first cycle was challenging. This difficulty arose because of larger variations within the RCB-III group in terms of ΔHbT%. In addition, the incomplete assessment of ΔHbT% may have resulted from the limited tumor coverage achieved by using only one plane of fiber bundles, even though every effort was made to place the array on the same part of the breast during subsequent imaging exams.

By utilizing amplitude and phase data at 12 wavelengths obtained by our NIRST imaging system, we derived various functional parameters, including vascular oxygen saturation (the ratio of oxy-hemoglobin to HbT), water concentration, scattering properties, as well as HbT. However, among these parameters, only HbT showed potential as a biomarker to predict the breast cancer response to NAC. This observation may be attributed to the broader impact of these properties across the tissue, where the changes induced by NAC were smaller relative to those observed in HbT. Consequently, the current single-plane imaging technique may not be sufficiently accurate to capture these subtle changes.

Comparing the earlier taxane/anthracycline regimens to the later dose-dense AC-T regimen, which involved more frequent administration of chemotherapy, the latter is expected to disrupt tumor cell growth more effectively and increase the likelihood of achieving a pCR. However, the study results showed that only 6 of 19 subjects (31%) in the later treatment group (2016–2022) achieved pCR, while 9 of 16 subjects (56%) in the earlier treatment group (2007–2012) achieved pCR (Supplementary Table S1). These observations may be influenced by the limited sample size within each breast cancer subtype group, and thus, further investigation in future large-scale clinical studies is warranted.

Like any investigation of a new imaging system, our study has certain limitations that should be acknowledged. First, the sample size in each RCB subgroup was limited. Second, ensuring consistent positional accuracy in localizing the fiber bundle array on the same part of the breast during sequential imaging sessions posed a challenge. Despite efforts to maintain arm position, chair angle, and fiber bundle holder orientation, variations in fiber bundle position relative to the tumor were observed in each session. Third, potential spatial heterogeneity in tumor response to NAC was not assessed, as we used average HbT over the entire cancer ROI in the imaging plane. This approach may not capture localized variations within the tumor. Although several studies have shown that RCB-III has the lowest number of presenting patients (14.7%; ref. 21) amongst all RCB classes, our cohort of 35 patients included only one RCB-I case. While MRI was used to identify tumor location, and precautions were taken, variations in fiber bundle positions were unavoidable. However, given the centimeter-scale spatial resolution of NIRST, these variations are unlikely to significantly impact the observed trends. To mitigate the sensitivity of fiber bundle position, we have developed a new breast interface utilizing flexible circuit board technology. This innovative interface incorporates embedded photodetectors and integrated optical source fibers, providing improved coverage of the breast. This advance allows for more reliable and whole breast assessment of ΔHbT% in predicting and monitoring tumor response to NAC. The utilization of the technology will be implemented in the next phase of our NAC imaging study, addressing some of the limitations identified in this study.

In conclusion, utilization of a portable (NIRST) system has proven valuable in imaging changes in breast tumor vasculature to monitor and predict tumor response to NAC in the early stage of treatment. The statistical analysis involved 35 patients with breast cancer undergoing NAC, and demonstrated that changes in total hemoglobin (ΔHbT%) after the first cycle of the treatment, as compared to pretreatment values, serve as an excellent prognostic factor for differentiating RCB classes with high AUC and accuracy values. The correlation analysis revealed a fair association between ΔHbT% and RCB class, further supporting the use of ΔHbT% as a reliable biomarker. This advancement will facilitate the translation of NIRST into larger clinical trials with enhanced ease of use. By integrating the tool into the clinical workflow of infusion suites, early assessment of prognostic biomarkers that predict RCB class could be achieved, to guide treatment decisions and optimize chemotherapy regimens that can improve patient management.

R.M. DiFlorio-Alexander reports grants from NIH during the conduct of the study. K.D. Paulsen reports personal fees from CairnSurgical and from InSight Surgical Technologies outside the submitted work; in addition, K.D. Paulsen has a patent for Sensor Strips for Intrinsic Near Infrared Spectroscopy Imaging pending. No disclosures were reported by the other authors.

X. Cao: Data curation, formal analysis, visualization, methodology, writing–original draft. K.E. Muller: Resources, methodology, writing–review and editing. M.D. Chamberlin: Resources, supervision, writing–review and editing. J. Gui: Formal analysis. P.A. Kaufman: Conceptualization, formal analysis, supervision. G.N. Schwartz: Resources, supervision. R.M. diFlorio-Alexander: Resources, formal analysis. B.W. Pogue: Conceptualization, formal analysis, supervision, methodology, writing–review and editing. K.D. Paulsen: Conceptualization, formal analysis, supervision, methodology, writing–review and editing. S. Jiang: Conceptualization, resources, visualization, methodology, project administration.

This study was supported by NIH research grants R01 CA176086 (to S. Jiang) and R01 EB027098 (to K.D. Paulsen/S.Jiang).

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.