Multispectral photoacoustic imaging of breast cancer tissue with histopathology validation

Abstract

Intraoperative multispectral photoacoustic pathology assessment presents a promising approach to guide biopsy resection. In this study, we developed and validated a novel photoacoustic technique to differentiate between healthy and cancerous tissues. Our method consisted of photoacoustic contrast calculations as a function of wavelength, followed by projections of the resulting spectra from training data into a two-dimensional space using principal component analysis to create representative spectra, then calculation of the average cosine similarity between the spectrum of each pixel in test data and the representative spectra. The test healthy tissue region had a 0.967 mean correlation with the representative healthy tissue spectrum and a lower mean correlation (0.801) with the cancer tissue spectrum. The test cancer tissue region had a 0.954 mean correlation with the cancer tissue spectrum and a lower mean correlation (0.762) with the healthy tissue spectrum. Our method was further validated through qualitative comparison with high-resolution hematoxylin and eosin histopathology scans. Healthy tissue was primarily correlated with the optical absorption of blood (i.e., deoxyhemoglobin), while invasive ductal carcinoma breast cancer tissue was primarily correlated with the optical absorption of lipids. Our label-free histopathology approach utilizing multispectral photoacoustic imaging has the potential to enable real-time tumor margin determination during biopsy or surgery.

1. Introduction

Breast cancer remains one of the most prevalent diseases affecting women worldwide [1]. Early detection and accurate diagnosis are critical to improving survival rates and patient outcomes [2]. While conventional imaging modalities such as ultrasound, mammography, and magnetic resonance imaging are important for breast cancer screening and diagnosis, they have limitations in terms of sensitivity, specificity, or acquisition time [3–9]. Photoacoustic imaging has recently emerged as a promising non-invasive technique for breast imaging that combines the high contrast of optical imaging with the deep tissue penetration and spatial resolution of ultrasound [10–12]. In addition, multispectral photoacoustic imaging utilizes multiple wavelengths of light to provide functional and molecular information about tissues [8,13,14]. By exploiting the unique spectral absorption properties of various tissue components, multispectral photoacoustic imaging enables simultaneous visualization of physiological parameters with molecular sensitivity such as oxygen saturation and tissue metabolism [15].

Conventionally, molecular-level tissue characterization is achieved through hematoxylin and eosin (H&E) histological analysis, which requires time-consuming steps such as formalin fixation, tissue sectioning, and staining, typically requiring more than 24 hours to complete [16,17], as illustrated in Fig. 1(a). Consequently, surgeons are unable to make real-time decisions regarding tumor margins during surgical resection procedures or immediately after a biopsy procedure. In contrast to H&E staining, multispectral photoacoustic breast imaging has the potential to provide a rapid, label-free, non-invasive alternative that could significantly expedite the diagnostic process and facilitate immediate clinical decision-making for breast cancer, considering related demonstrations for other diseases (e.g., thyroid [18], prostate [19], skin [20,21], and ovarian [22] cancers), as illustrated in Fig. 1(b).

Fig. 1.

(a) Conventional histological procedures include tissue sectioning, freezing, preparing, and staining to produce tissue fixed on slides for staining and analysis. (b) As an alternative to conventional procedures, we propose a multispectral photoacoustic imaging system that will image unprocessed tissues to more rapidly obtain important histopathology information with virtual stains.

Multispectral photoacoustic imaging within the near-infrared (NIR) range, extending up to 2000 nm, enables differentiation between healthy and cancer tissues through the analysis of spectral signatures from various tissue components, including hemoglobin, lipids, and water, based on their distinct optical absorption profiles [23]. In addition, this imaging technique offers critical insights into breast tumor vasculature, oxygenation status, and metabolism, enabling accurate tumor staging, treatment planning [24], and surgical guidance for tumor resection.

This paper presents results from the first known application of multispectral photoacoustic imaging to differentiate between healthy and cancerous breast tissues using wavelengths up to 2000 nm [25]. A novel spectral analysis is introduced to identify differences in photoacoustic signals originating from healthy versus cancer tissues. Our approach is then qualitatively validated using corresponding H&E histopathology scans.

2. Methods

2.1. Tissue acquisition

Formalin-fixed paraffin-embedded blocks of both healthy (i.e., normal) and cancerous (i.e., invasive ductal carcinoma, grade 3, stage 2) human breast tissue were purchased from PrecisionMed, Inc. (Carlsbad, CA, USA), and three unprocessed tissue sections per block were extracted and placed on microscope slides for imaging. To obtain a baseline for morphological comparison, H&E histological samples of healthy and cancerous breast tissue were taken from sections adjacent to those being imaged.

2.2. Multispectral volumetric imaging system

Our photoacoustic imaging system consisted of a Vevo F2 LAZR-X laser (FUJIFILM VisualSonics, Toronto, Canada), connected to a 256-element UHF57x ultrasound probe (25-57 MHz bandwidth), and a probe-specific fiber jacket that coupled the ultrasound probe to optical fibers. The laser wavelengths ranged 680-970 nm (5 nm increments, 59 wavelengths total) and 1200-2000 nm (5 nm increments, 161 wavelengths total), and the mean laser energy was 37.7 mJ. The parameters of the imaging system are listed in Table 1.

Table 1. Multispectral volumetric photoacoustic imaging system parameters.

Parameters	Value
Number of Elements	256
Center Frequency	41 MHz
Bandwidth	78%
Average Laser Energy	37.7 mJ
Laser Wavelength	680 - 970 nm & 1200 - 2000 nm
Field of View	$x$ mm x 13 mm x 14 mm
Resolution	0.200 mm x 0.016 mm x 0.055 mm

We designed and built a custom volumetric imaging platform to acquire data, consisting of a microscope slide holder, mounted to a 3D translation stage (Thorlabs Inc., Newton, NJ, USA) with 25 mm travel range and 10 $\mu$ m minimum step resolution (see Fig. 2). The microscope slide was positioned horizontally within a water tank. The ultrasound probe was suspended above the slide with its axial dimension oriented perpendicular to the surface of the slide. This setup enabled accurate translation of the slide along the lateral, axial, and elevational dimensions relative to the ultrasound transducer.

Fig. 2.

(a) Illustrations of our multispectral photoacoustic imaging system. The custom volumetric imaging platform acquires 3D spectral images by linear scanning. (b) Photograph of the experimental setup to image breast tissues on microscope slides. (c) Schematic diagram of a tissue sample on a slide relative to the image plane and scanning direction employed to achieve volumetric images.

Prior to imaging, the laser was calibrated to ensure that multiwavelength data were normalized to the same energy by characterizing the transmission energy of the LAZR-X optical fiber across the chosen wavelength range. To achieve optimal photoacoustic signals, the slide in the imaging platform was translated into the spatially calibrated laser-illuminated area by adjusting the translation module in the axial dimension.

Multispectral data were first acquired from two healthy and two cancerous sections, to identify their representative photoacoustic spectra. The two remaining slides were imaged to test the spectrum representations. In each case, the volumetric data spanned an $x$ mm $\times$ 13 mm $\times$ 14 mm field of view (FOV) with 0.200 mm $\times$ 0.016 mm $\times$ 0.055 mm voxel size in the elevation, axial, and lateral dimensions, respectively. The elevation distance $x$ mm was determined based on the size of each tissue sample to ensure full coverage of the entire sample.

2.3. Spectral analysis

From a theoretical perspective, the optically induced initial pressure distribution at wavelength $\lambda$ and position $\mathbf{\text{x}}$ is given by [26]:

$$p_0(\mathbf{\text{x}},\lambda )=\mathrm{\Gamma }{\mu }_a(\mathbf{\text{x}},\lambda )\mathrm{\Phi }(\mathbf{\text{x}},\lambda )$$ (1)

where $\mathrm{\Gamma }$ is the Grüneisen parameter, ${\mu }_a$ is the absorption coefficient, and $\mathrm{\Phi }$ is the optical fluence. Contrast, $C$ , is a measurement of the ratio of photoacoustic signal amplitudes, which is expected to be proportional to the acoustic pressure, as follows:

$$\begin{array}{rl}C(\mathbf{\text{x}},\lambda )& \propto p_0(\lambda )\end{array}$$ (2)

$$\begin{array}{rl}& \propto \mathrm{\Gamma }{\mu }_a(\mathbf{\text{x}},\lambda )\mathrm{\Phi }(\mathbf{\text{x}},\lambda )\end{array}$$ (3)

The wavelength-dependent optical absorption coefficient, ${\mu }_a$ , is assumed to be the primary contributor to the photoacoustic signal amplitude. Therefore, the photoacoustic spectrum is considered sufficiently proportional to optical absorption across different wavelengths. Although noise, thermal differences, and fluence variations caused by scattering may impact this proportionality, these contributing factors are considered negligible, and the photoacoustic spectra are often considered sufficient to provide qualitative trends for tissue identification [8].

Using the rationale provided above, Fig. 2 illustrates the data acquisition steps and associated experimental setup for spectral analysis. After data acquisition, contrast was calculated as a function of laser wavelength $\lambda$ for each pixel within the acquired 2D photoacoustic spectral images, using the equation:

$$C(\lambda )=\frac{S_i(\lambda )}{S_o(\lambda )}$$ (4)

where $S_i$ and $S_o$ are the mean photoacoustic signal amplitudes within regions of interest (ROIs) inside and outside the laser-excited region, respectively. The outside ROI was located 2 mm proximal to the laser-excited ROI and was not placed at the same depth as the laser-excited ROI to avoid local reflection artifacts. Note that the ratio in Eq. (4) places each $S_i$ measurement from different positions and acquisitions on a similar scale, regardless of pulse-to-pulse energy fluctuations. Each resulting photoacoustic spectrum was normalized by its L2 norm, thus scaling the L2 norm of each spectrum to 1, which preserved the general trends and features of the original spectrum by preserving ratios between values within the spectrum.

Because the prior information of photoacoustic endmembers was unknown, unmixing algorithms were not implemented. Instead, each photoacoustic spectrum was projected into a two-dimensional space using principal component analysis [27]. Representative spectra for healthy and cancerous tissues were then identified through careful analysis of the data distribution in the feature space. To validate these representative spectra, data from the test tissue slides noted in Section 2.2 were analyzed by calculating the average cosine similarity between the photoacoustic spectrum of each pixel and each of the discriminating spectra previously identified, as follows:

$$\text{correlation}=\sum _{\lambda }\overline{C}(\mathbf{\text{x}},\lambda )\ast {\overline{C}}^{\ast }(\lambda )$$ (5)

where $\overline{C}(\mathbf{\text{x}},\lambda )$ is the normalized photoacoustic spectrum value of test tissues at location $\mathbf{\text{x}}$ and wavelength $\lambda$ , and ${\overline{C}}^{\ast }(\lambda )$ is the normalized representative spectrum of either the healthy or cancer tissue.

2.4. Fluence approximation

Due to the strong optical absorption of water relative to other biological chromophores at laser wavelengths ranging 1380-2000 nm [28], we performed a semi-quantitative approximation of the effective fluence spectrum, which we define as the fluence seen by the tissue sample after optical attenuation based on the Beer-Lambert law [29]. The estimated effective fluence, ${\mathrm{\Phi }}_{\text{eff}}$ , on the tissue at wavelength, $\lambda$ , can be expressed as:

$${\mathrm{\Phi }}_{\text{eff}}=\frac{E}{A}{e}^{-d{\mu }_{\text{water}}(\lambda )}$$ (6)

where $E$ =37.7 mJ is the constant laser energy used in our setup, $d$ =2 cm is the water path length, $A$ =0.28 ${\mathrm{cm}}^2$ is the laser illumination area, defined based on the angle of the optical fibers coupled to the ultrasound probe (which directs laser light to illuminate a 0.2 cm depth x 1.4 cm width region of interest across the image plane), and ${\mu }_{\text{water}}$ is the optical absorption spectra of water [28].

3. Results

Figure 3 shows an example co-registered pair of volumetric ultrasound and photoacoustic images at 750 nm optical laser wavelength, as well as representative healthy and cancer spectra compared with deoxyhemoglobin (Hb) [23] and lipid [30] spectra. In the lower wavelength range, both the healthy and cancer tissue spectra are similar to that of Hb, indicating the strong influence of Hb on the photoacoustic effect in the wavelength range of 680-970 nm. In addition, the representative cancer tissue spectrum has local absorption peaks at wavelengths of 935, 1210, and 1720 nm, which are more pronounced than those in the representative healthy tissue spectrum, although the 1720 nm peak is least pronounced. In particular, at 1720 nm wavelength, the normalized amplitude of the cancer spectrum deviates by 4.63 standard deviations ( $\sigma$ ) from the mean amplitude measured between 1380-2000 nm. In contrast, the healthy spectrum shows a smaller deviation of 2.76 $\sigma$ from its mean amplitude over the same wavelength range. These differences can be explained by the effective fluence differences described in the next paragraph.

Fig. 3.

(a) Co-registered volumetric photoacoustic and ultrasound images of the test healthy tissue at 750 nm laser wavelength. (b) Representative healthy and cancer tissue spectra shown with the optical absorption spectra of lipid [30] and deoxyhemoglobin (Hb) [23].

Figure 4 shows the effective fluence on a tissue sample after laser energy attenuation by water, approximated using Eq. (6), with the optical absorption spectra of water [28] and lipid [30] shown for comparison. At 1210 nm, the estimated effective fluence is 10.3 ${\mathrm{mJ/cm}}^2$ . However, beyond 1385 nm the effective fluence drops below 0.1 $\mu$ J/ ${\mathrm{cm}}^2$ , except in the range from 1620 nm to 1750 nm where the fluence exceeds 0.1 $\mu$ J/ ${\mathrm{cm}}^2$ . At the lipid absorption peak wavelength of 1720 nm, the fluence reaches 0.5 $\mu$ J/ ${\mathrm{cm}}^2$ . Lipids exhibit stronger absorption than water at both 1210 nm and 1720 nm wavelengths. At 1210 nm, the absorption coefficient of lipid peaks at approximately 1.73 ${\mathrm{cm}}^{-1}$ , exceeding that of water (i.e., 1.26 ${\mathrm{cm}}^{-1}$ ). Similarly, at 1720 nm, lipids show a higher absorption coefficient peak of approximately 9.47 ${\mathrm{cm}}^{-1}$ relative to that of water (which is approximately 6.16 ${\mathrm{cm}}^{-1})$ . The outcome of the effective fluence estimation relative to the absorption spectra of lipids and water supports the observations achieved with the three peak wavelengths of 935 nm, 1210 nm, and 1720 nm in Fig. 3(b). At these wavelengths, lipids have higher absorption peaks compared to water, while the water absorption remains moderate (e.g., less than 10 ${\mathrm{cm}}^{-1}$ ), with a lower effective fluence achieved at 1720 nm wavelength relative to that achieved at 935 nm and 1210 nm.

Fig. 4.

Effective fluence after attenuation through water, approximated based on Eq. (6), shown for wavelength ranges of (a) 680 nm to 2000 nm and (b) 1370 nm to 2000 nm, with the optical absorption spectra of water [28] and lipid [30] additionally shown for comparison. The purple box in (a) indicates the wavelength range in (b).

Figure 5 shows the results of cross-correlating the chosen photoacoustic spectra with each voxel in the volumetric test images. The test healthy tissue region had 0.967 mean correlation with the healthy tissue representative spectrum and lower mean correlation (i.e., 0.801) with the cancer tissue representative spectrum. The test cancer tissue region had 0.954 mean correlation with the cancer tissue representative spectrum and lower mean correlation (i.e., 0.762) with the healthy tissue representative spectrum. The tissue boundaries are visible with a threshold ranging 0.95-1.0 in the correlation images.

Fig. 5.

Volumetric images of correlations.

Figure 6 shows the maximum correlation projections along the axial dimension, representing the maximum correlations between the representative photoacoustic spectra and each test tissue in this dimension. Each maximum correlation projection is compared with the corresponding camera image and H&E-stained image. The tissue contours of the maximum correlation projection images agree with the corresponding physical images and H&E scans. Observed differences between the maximum correlation projections, physical images, and H&E stained images can be caused by the H&E stained sections originating from adjacent tissue sections of the same block (to avoid histology analyses of water-submerged sections), reliance on manual translations to acquire volumetric images, subtle misalignment of the slide with either the orthogonal dimensions of the imaging transducer or the camera lens, and resolution differences between the images.

Fig. 6.

Maximum correlation projections and comparison with slide images and H&E stain images. The slide images are photographs of the actual tissue imaged during the photoacoustic image acquisition process.

Figure 7 shows zoomed-in H&E images of the healthy and cancer tissues from the training set. The size of the 200 $\mu$ m x 55 $\mu$ m photoacoustic image resolution cell is indicated as black boxes, based on the 500 $\mu$ m scale bar shown in Fig. 7. Therefore, for each photoacoustic image voxel in our 3D image reconstruction (or for each pixel in a 2D photoacoustic image), the corresponding photoacoustic spectrum is determined by the combination of the individual chromophore spectra contained within a photoacoustic image resolution cell (e.g., the lipid and Hb spectra shown in Fig. 3), weighted by their respective densities. We observe a few key histopathological features [31]. First, in the zoomed-in H&E image of the healthy tissue, the most prominent components are the pink or purple collagen fibers that form the stromal network, providing structural support to the tissue. Throughout the stroma, there are sparingly scattered cells with dark blue nuclei stained by hematoxylin. Small blood vessels are visible within the stromal tissue, containing red blood cells that appear red in color. In the zoomed-in H&E image of the cancer tissue section, dense clusters of atypical cells with enlarged, dark-blue stained nuclei are shown, which is typical of cancer cells. These cells are arranged in irregular patterns and show increased nuclear-to-cytoplasmic ratios, indicating malignant development. White elliptical spaces representing lipids (e.g., adipose tissues) remain unstained due to their neutral properties. Therefore, the higher concentration of these empty elliptical droplets in the cancer tissue section indicates a denser distribution of lipids relative to the healthy tissue. Relative to the photoacoustic image resolution cell, most of the photoacoustic image pixels are expected to contain a significant proportion of lipids, as supported by the 0.954 mean correlation with the cancer spectrum in Fig. 3(b) containing spectral peaks that are consistent with the lipid optical absorption spectrum.

Fig. 7.

Zoomed-in H&E stain images of healthy and cancer tissues, with arrows pointing to chromophore features of interest and boxes that denote the relative size of a photoacoustic image resolution cell.

4. Discussion

The development of a 3D multispectral photoacoustic-ultrasound imaging system facilitates potential real-time pathological examination of unprocessed breast cancer tissues in visible and NIR spectral ranges, targeting hemoglobin and lipids. Conventional histopathological analysis relies on lengthy H&E staining procedures to examine cell morphology and content distribution. In contrast, our system employs fast spectral correlation between the tested tissues and selected discriminating tissue spectra for diagnosis. Our system setup and spectral analysis approach appears to both predict cancer presence from scalar correlations and reconstruct 3D spectroscopic images that map the localized optical absorption distribution within the tissue. These 3D images can then be used to infer tissue histopathology. The dual promise of diagnostic imaging and real-time guidance has implications for both cancer detection and treatment monitoring.

The three peak wavelengths in the cancer spectrum shown in Fig. 3 (i.e., 935 nm, 1210 nm, and 1720 nm) are strongly associated with lipid absorption [23,30], suggesting that lipid concentration serves as a potential discriminator between healthy and cancerous breast tissues in the NIR range. Considering that the predominant cell type in breast tissues is the adipocyte, or fat cell, specialized for lipid storage and synthesis, lipid content may be altered during cancer development as a result of abnormal cell proliferation. This finding is further validated in the comparison between the H&E histology images of healthy and cancer tissues (Fig. 7).

The altered lipid profile in breast cancer tissues (Fig. 7) may have implications for diagnosis, prognosis, and potential therapeutic targets if detected and localized by multispectral photoacoustic imaging. In contrast, the healthy H&E image reveals a higher concentration of hemoglobin and a lower lipid content in most photoacoustic resolution cells, hence the associated photoacoustic spectra are expected to more closely resemble hemoglobin, with less pronounced characteristic lipid peaks. These expectations for cancer and healthy tissue are supported by the photoacoustic spectra presented in Fig. 3(b). In addition, the similarities between hemoglobin and healthy tissue in Fig. 3(b), particularly in the 680-970 nm wavelength range, suggest that oxygen saturation is another potential criteria to distinguish healthy tissue margins during intraoperative breast tumor margin assessments [32].

Based on the outcome of the effective fluence measurements (Fig. 4), the absorption peak at 1720 nm is likely less pronounced in Fig. 3 due to insufficient effective fluence reaching the tissue after a 2-cm light path in water. This concern can potentially be addressed by implementing a closer light delivery mechanism to enhance the contrast between the peak and baseline signals. For example, with the optical absorption of water being approximately 6.2 ${\mathrm{cm}}^{-1}$ at 1720 nm wavelength, reducing the distance between the laser and tissue from 2 cm to 0.5 cm would increase the effective fluence by four orders of magnitude, from sub- $\mu$ J/ ${\mathrm{cm}}^2$ to mJ/ ${\mathrm{cm}}^2$ , based on Eq. (6).

Each volumetric image comprises approximately 20 2D slices with an inter-slice spacing of 0.2 mm. The total acquisition time for a single image is 20 minutes, excluding data processing and manual control of the translation stage. While the translation stage has a resolution of 0.01 mm, allowing for potentially finer slice spacing and therefore more balanced resolution in three dimensions of the volumetric image, implementing finer increments would significantly increase the acquisition time. Therefore, future iterations of our validation approach could leverage the existing motor control port of our imaging system to offer a fully automated scanning approach, potentially reducing acquisition times. In addition, the implementation of sparse sampling techniques could potentially accelerate the imaging process while preserving essential spectral information, making the system more suitable for intraoperative applications.

Compared to traditional spectral unmixing approaches that provide chromophore distribution maps [18,20–22,33], our method does not require fluence compensation for medium attenuation nor assume constant fluence across the entire wavelength range. Instead, the healthy and cancer tissue spectra from the test set were mapped directly to their respective pathology using photoacoustic spectra from the training set, which were acquired under identical conditions to the test set. This data analysis approach is beneficial because it does not rely on prior information (e.g., absorption spectra of individual chromophores, fluence attenuation) that is otherwise required for traditional spectral unmixing approaches.

One limitation of the proposed approach is the heavy reliance on qualitative manual inspection, which introduces inherent subjectivity and potential observer bias in the assessment process, although our visual inspection approach provides promising confirmation of our findings. Co-registration between photoacoustic images and H&E stained histological sections will provide additional ground truth labels for each voxel, offering a more objective method in future studies. To further enhance diagnostic capabilities, the incorporation of cell nuclei density—a more direct examination criterion—into the classification algorithm can also be developed in the future by integrating laser sources with 266 nm wavelength (i.e., a wavelength at which cell nuclei have higher absorption coefficients than those of cytoplasm and extracellular matrix [34]). Although this wavelength is not available with our current imaging system, this addition has the potential to provide more comprehensive and accurate real-time pathological examinations, potentially bridging the gap between in vivo imaging and traditional histopathology.

5. Conclusion

We present the first known label-free multispectral photoacoustic-ultrasound imaging as a promising method for imaging breast tissues using wavelengths up to 2000 nm without staining to acquire pathological information. Healthy tissue was primarily correlated with the optical absorption of blood, while invasive ductal carcinoma breast cancer tissue was primarily correlated with the optical absorption of lipid. These preliminary results indicate possible rapid determination of breast tumor margins. The co-registered ultrasound images provide complementary anatomical context to the pathological information derived from multispectral photoacoustic images.

Funding

Chan Zuckerberg Initiative10.13039/100014989 (2022-30951); National Institutes of Health10.13039/100000002 (R01 EB032960, P30 CA006973).

Disclosures

The authors declare no conflicts of interest.

Data availability

Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.