Research ArticleCANCER

Intraoperative visualization of the tumor microenvironment and quantification of extracellular vesicles by label-free nonlinear imaging

See allHide authors and affiliations

Science Advances  19 Dec 2018:
Vol. 4, no. 12, eaau5603
DOI: 10.1126/sciadv.aau5603


Characterization of the tumor microenvironment, including extracellular vesicles (EVs), is important for understanding cancer progression. EV studies have traditionally been performed on dissociated cells, lacking spatial information. Since the distribution of EVs in the tumor microenvironment is associated with cellular function, there is a strong need for visualizing EVs in freshly resected tissues. We intraoperatively imaged untreated human breast tissues using a custom nonlinear imaging system. Label-free optical contrasts of the tissue, correlated with histological findings, enabled point-of-procedure characterization of the tumor microenvironment. EV densities from 29 patients with breast cancer were found to increase with higher histologic grade and shorter tumor-to-margin distance and were significantly higher than those from 7 cancer-free patients undergoing breast reduction surgery. Acquisition and interpretation of these intraoperative images not only provide real-time visualization of the tumor microenvironment but also offer the potential to use EVs as a label-free biomarker for cancer diagnosis and prognosis.


The tumor microenvironment, host to cancer-associated events (1) such as angiogenesis (2, 3), production of cancer-associated fibroblasts (CAFs) (4), and a reorganized extracellular matrix (ECM) (5), provides many potential biomarkers for cancer pathology. Tumor-associated extracellular vesicles (EVs), which play important roles in intercellular communication both inside and outside of the tumor microenvironment (6), have been found to promote tumor progression by directing cancer-associated events and changes (68), illustrating the clinical significance of EV detection. Various EV detection methods for cancer diagnosis have been proposed and investigated, such as flow cytometry performed on circulating exosomes (9), immuno-based detection (10), and fluorescence label-based approaches for visualization (11). However, detection and imaging of EVs have not been performed in a spatially-resolved and label-free way to study the unperturbed density and distribution of EVs in ex vivo human tumor tissues. Novel label-free multimodal multiphoton imaging technology has been implemented in a laboratory-based system to visualize the unperturbed EVs in an in vivo animal tumor model to evaluate its potential applicability to human breast cancer (12, 13). Few attempts, however, have been made to study human EVs at the point of procedure, such as during interventional surgical procedures with a portable system designed for intraoperative label-free nonlinear optical imaging of fresh untreated or unstained human tumor tissues.

Intraoperative optical imaging and spectroscopy have previously been demonstrated to address different diagnostic cancer needs with various imaging modalities, such as tumor margin assessment (1416), tumor type differentiation (17, 18), and lesion/malignancy determination (18, 19). Nonlinear optical imaging is particularly suitable for the intraoperative visualization of human tissue specimens because of the molecular and structural specificity it can provide (17, 18). Previous intraoperative nonlinear optical imaging approaches have used stimulated Raman scattering (18) and two-photon fluorescence (2PF) with exogenous labeling agents (17), as well as second-harmonic generation (SHG) (17). However, these approaches were mainly focused on tissue anatomy and macroscopic tissue features, attempting to generate images that replicate hematoxylin and eosin (H&E)–stained histology. Few efforts have been devoted to imaging and characterizing the tumor microenvironment intraoperatively, which hosts many potential biomarkers for tumor-specific cells and features, including EVs.

To visualize the unperturbed tumor microenvironment in real time, we designed and built a custom portable multimodal system for label-free nonlinear optical imaging. On the basis of our methodology for performing label-free imaging of EVs (12), the primary application of this intraoperative imaging system was to directly observe the spatial distribution of EVs within the human tumor microenvironment. Four nonlinear optical imaging modalities were integrated into this system and displayed in different colors in multimodal images: SHG was used for visualizing collagen fiber reorganization (2022), displayed in green. 2PF was used for visualizing elastin fibers and flavin adenine dinucleotide (FAD)–containing cell cytoplasm (20, 23), displayed in yellow. Third-harmonic generation (THG) was used for highlighting interface structures (20, 24) such as cell membranes, lipid boundaries, and EVs, displayed in magenta, and 3PF was used for mapping NADH (reduced form of nicotinamide adenine dinucleotide) in the lipids (20, 25), displayed in cyan. EVs were visualized, characterized (fig. S1), and spatially coregistered with other visualized tissue features in the tumor microenvironment. The results of this study uniquely contribute to a better understanding of the roles that EVs play in the tumor microenvironment, their substantial clinical potential as a diagnostic and prognostic biomarker for cancer aggressiveness and progression, and the potential for imaging and optically characterizing EVs in other biological samples such as liquid surgical waste, blood, and urine.


Intraoperative imaging of the unperturbed tumor microenvironment

The unperturbed human breast tumor microenvironment in ex vivo–resected specimens was imaged intraoperatively and without the use of any exogenous stains or labels using the portable multimodal nonlinear optical imaging system within 30 min after surgical excision of the tissue. Abundant optical signals associated with tissue features were identified, including molecular signatures from autofluorescent NADH and FAD that are associated with metabolic activity within individual cells in the tumor microenvironment. These real-time image-based results reveal the substantial advantage of this label-free intraoperative system over conventional H&E-stained or immunohistochemically stained slides that often require several days to prepare. The multimodal images acquired from different sites (Fig. 1) demonstrate the capability of visualizing heterogeneous human breast tissue. The digital images of colocated H&E-stained histology slides (Fig. 1) were acquired for confirming the nonlinear optical signatures of various tissue structures. In the multimodal nonlinear optical images, EVs were found near the breast cancer region (Fig. 2), suggesting their origin and their tumor-related functions (26).

Fig. 1 Multimodal intraoperative label-free nonlinear optical images of human breast tissues and the corresponding histology.

Multimodal label-free nonlinear optical image (left) and the colocated histology (right) of (A) invasive ductal carcinoma (IDC) with an overall orientation of collagen alignment and tumor cell infiltration (red dashed arrows), (B) adipocytes (red dashed arrows) and blood vessel (red solid arrows), (C) adipocytes (red dashed arrows) and mammary duct (red solid arrows), and (D) healthy breast tissue from breast reduction surgery. Scale bars, 100 μm.

Fig. 2 EV enrichment in the tumor microenvironment.

(A) Multimodal label-free nonlinear optical image of a tissue site with ductal carcinoma in situ (DCIS) (boundary marked by red dashed line). (B) Colocated H&E histology. (C) THG-contrast image visualizing the DCIS boundary and EVs. (D) Binary image of EVs segmented from (C). Scale bars, 100 μm.

The most common subtype of invasive carcinoma in breast tissue, invasive ductal carcinoma (IDC), is characterized by thick rows of large groups of tumor cells that construct a nest-like structure (Fig. 1A), with an overall orientation marked by red arrows. These cells were identified by their yellow or magenta cytoplasm, which represents the 2PF signal primarily generated from FAD (27), and the THG signal generated from membrane-bound structures such as EVs (12) and ribosomes, respectively. In addition, the collagen fibers (green channel) were aligned in the same direction in response to the tumor cell infiltration, as verified by the colocated histology images.

Mammary ducts and blood vessels in the tumor microenvironment are associated with DCIS and angiogenesis, respectively. These structures were visualized in our study and found to be residing with adipocytes inside the ECM (Fig. 1, B and C). Epithelial cells of mammary ducts and the interfaces of adipocytes are highlighted in the THG channel. In addition, the strong 3PF signal was mainly generated from NADH associated with lipids and localized inside the adipocytes. The yellow-colored endothelial cells of a blood vessel (Fig. 1B) were visualized by 2PF imaging. Last, the ECM visualized by SHG supported the tissue structures discussed above. These images demonstrate the heterogeneity of the tumor microenvironment and the capability of our imaging system to faithfully visualize its major structural and molecular components. For comparison, we also imaged healthy breast tissues (Fig. 1D) from cancer-free subjects undergoing breast reduction surgery. The healthy breast tissues can be readily distinguished from the cancerous breast tissues by the prevalence of tissue areas with highly organized collagen fibers and a lack of cancer cells and angiogenesis.

In addition to macroscopic features (Fig. 1), micro- to nanoscale features such as enrichment of EVs near a site of DCIS were also identified (Fig. 2). A multimodal label-free nonlinear optical image (Fig. 2A) includes multiple nonlinear optical contrasts, respectively, from SHG-visible ECM, 2PF/3PF-visible tumor cells within the region of DCIS, and THG-visible EVs. Most tissue features were confirmed by the colocated histology image (Fig. 2B), except for the EVs, which are lost during histological preparation. In the nonlinear optical image, EVs were primarily visualized by THG imaging because of the strong phase-matching condition that the membrane structures provided, especially because of their large surface-to-volume ratios (28). These EVs, appearing in the THG-contrast image (Fig. 2C) as diffraction-limited bright points, were segmented from the background (Fig. 2D) using a segmentation algorithm (Materials and Methods). A high density of EVs were found in this specific region near the tumor (12), which agrees with previous observations that cancer cells are a major producer of EVs in the tumor microenvironment (29). The quantification and analysis of EVs reveal correlations between EV density and pathological diagnoses, as discussed below.

Quantification of EVs and correlations with pathological diagnosis

Tissue specimens from 29 human breast cancer cases and 7 cancer-free breast reduction cases were imaged and analyzed for this study. An automated segmentation algorithm (Fig. 3A) was used to isolate the EVs (Fig. 3B) and subsequently quantify EV density (Materials and Methods). The segmentation algorithm was applied on THG images of purified vesicles (Fig. 3B) to validate its accuracy (Materials and Methods). Using this algorithm to segment EVs from the intraoperative images of the breast tissues, we found a clear difference in EV density (Fig. 3C) between the cancer cases (average, 142 ± 56 nl−1) compared with the breast reduction cases (average, 23 ± 8 nl−1). Furthermore, among the breast cancer cases, the quantified EV densities were correlated with the pathological diagnoses, including the histologic grade of IDC, the nuclear grade of DCIS, and the tumor-to-margin distance. The grades of IDC and DCIS are used in breast cancer pathology to evaluate the apparent aggressiveness of the cancer cells (30). A higher grade typically implies a faster and more aggressive tumor growth, and it has been suggested that more aggressive tumor cells will produce a higher number of EVs distributed throughout the tumor microenvironment (31). The quantified EV densities are plotted against the tumor-to-margin distance (Fig. 3D). As expected, there is a decreasing trend of EV density, with increasing distance from the tumor to the closest surgical margin or surface of the excised tissue mass that was imaged intraoperatively, indicating the diffusion-driven EV distribution. Moreover, data points with different histologic grades of IDC (1, 2, and 3) could be divided into three groups (shaded regions in Fig. 3D), where the curves of the region boundaries are represented by a diffusion model for EVs (Materials and Methods). By controlling the tumor-to-margin distance variation, we also performed multiway analysis of variance (ANOVA) statistics (Materials and Methods) to find that the quantified EV densities are highly related to the histologic grade of IDC (Fig. 3E), but less related to the nuclear grade of DCIS.

Fig. 3 Quantification and pathological correlations of EVs.

(A) Flowchart of EV segmentation and quantification algorithm. (B) Representative THG-contrast image acquired from the tumor microenvironment and processed binary image, highlighting the presence of EVs within the tumor microenvironment. (C) Comparison of EV density from breast cancer cases versus healthy breast reduction cases. The average EV density is 142 ± 55 nl−1 for the cancer cases, while it is only 23 ± 8 nl−1 for the healthy breast reduction cases. ****P < 0.0001 (one-sided Student’s t test). (D) EV density data from each case are registered by the distance from tumor to closest surgical margin and the cancer invasiveness grade. An overall decreasing trend of EV density is identified with increasing tumor-to-margin distance. Data points are divided into three groups (shaded areas) representing different histologic grades of IDC. (E) Relationship between EV density and IDC histologic grade. To minimize the effect of spatial heterogeneity, EV data were chosen from cases within a small range of margin distances (0 to 8 mm). Sample size of each IDC grade is indicated above each bar. ***P < 0.001, **P < 0.01, *P < 0.1 (multiway ANOVA test, multiple comparison test).

Application of EV quantification and tumor microenvironment visualization: Studying cancer invasion near desmoplasia

Desmoplastic reaction (or desmoplasia), as a pathophysiologic event occurring in the stroma of breast tissue, is often used as a histopathological risk factor for cancer invasion (3235). Multimodal nonlinear optical images were acquired at tissue sites that were identified histologically, by a board-certified pathologist, as containing evidence of desmoplasia. From the nonlinear optical images, macroscopic morphological features associated with cancer invasion were identified and validated by comparing with colocated histological images. Our findings also show that the distribution of EV densities, which are highlighted in the THG images, matches the invasion phase, based on the macroscopic morphological features.

In the nonlinear optical and histologic images of an early phase of desmoplasia (Fig. 4, A and B), the red dashed lines mark the interfaces between tumor and desmoplasia. The dense and thick collagen fibers below the interface were recognized as being associated with desmoplasia, while the regions with densely packed cells, located on the other side of the interface, were identified as the tumor region. These dense collagen fibers of desmoplasia are tightly aligned to block tumor cell infiltration. That is because a desmoplastic reaction, at an earlier phase of carcinogenesis, is described as a secondary reaction of the human body trying to “heal” the tumor by producing dense fibrosis (35). However, it is still largely unknown whether this reaction is initiated by the tumor or is a response reaction of the body (36). At this early phase, tumor cells have not started to invade the dense ECM and, hence, are considered to be less aggressive. The EV density obtained from the THG-contrast image of this imaging site was around 144 nl−1 (Fig. 4C) on both sides of the red dashed line.

Fig. 4 Determination of phase of tumor cell invasion around desmoplasia by EV distribution.

(A) Multimodal label-free nonlinear image of desmoplasia at an early phase. There is an interface (red dashed line) between the tumor and the dense collagen fibers of desmoplasia, and the tumor cells are identified only in the tumor region (white arrows). (B) Colocated histology image of the early-phase desmoplasia. (C) Binary image of segmented EVs from the THG channel of (A). The average (AVG) EV density is quantified to be 144 nl−1, and there is no major difference (113 nl−1 versus 163 nl−1) between the EV counts from each side of the interface. (D) Multimodal nonlinear optical image of desmoplasia at a late phase. The interface between dense collagen fibers and the tumor is marked by a red dashed line, with infiltrating tumor cells being identified within the collagen region (white arrows). (E) Colocated histology image of this late-phase desmoplastic reaction. (F) Binary image of segmented EVs from the THG channel (A). The average EV density of the entire FOV is 575 nl−1, but the EV density within the dense collagen fibers (938 nl−1) is much higher than the EV density within the tumor (188 nl−1). Scale bars, 100 μm.

A later phase of tumor invasion is initiated by the secretion of ECM degrading enzymes such as matrix metalloproteinases from tumor cells (37), which can break down and remodel the dense collagen fibers produced by the desmoplastic reaction. The tumor cells subsequently use these fibers as a scaffold to facilitate their further migration (37). In this later phase (Fig. 4, D and E), small gaps were observed between the desmoplastic region located below the tumor-desmoplasia interface (red dashed lines), and a few groups of cells (white and yellow arrows) were identified within these gaps. In breast cancer histopathology, the presence of tumor cells and fibroblasts inside the newly formed ECM normally signifies a later phase of local tumor invasion (37, 38). These active tumor cells and CAFs rely on EVs to transfer intercellular information to complete targeted gene expression (39) that facilitates further tumor progression. As expected, images reveal a greater number of EVs (575 nl−1) at this site (Fig. 4E), compared with the early phase (Fig. 4C), with most EVs found inside the desmoplastic region, suggesting that more active intercellular communication is taking place.

There are several similarities and differences between the two phases of cancer invasion near sites of desmoplasia. In both cases, the thick and straightened desmoplasia-associated collagen fibers can be easily distinguished from the thin and wavy collagen fibers within the tumor regions, and there is a clear boundary separating the tumor region and the desmoplasia-associated collagen region. In the later phase, tumor cells are observed to infiltrate the remodeled desmoplasia-associated collagen region. These similarities and differences help classify and differentiate desmoplasia from the tumor and enable the direct comparison between cancer invasion phases. The concurrence of the macroscopic morphological features and the signature distributions of EV densities directly reveal the relationship between the production and distribution of EVs and macroscopic cancer invasion.


The tissue features of the breast tumor microenvironment, including collagen fibers, cells, blood vessels, mammary ducts, and lipids, are typically visualized by H&E-stained histology that requires labor- and time-intensive tissue processing. Intraoperative multimodal nonlinear optical imaging can visualize most of these features in ex vivo–resected tissue specimens in real time without the aid of any labeling or tissue preparation. The correlations with gold standard H&E-stained histology were found to validate the nonlinear optical signatures of these essential tissue features. The label-free multimodal nonlinear optical images can provide abundant details and molecular contrasts from the unperturbed breast tumor microenvironment, especially the metabolic information associated with FAD and NADH that is rapidly lost after tissue excision. Although the nonlinear optical signals from cell nuclei are not as distinguished as in stained histological sections, negative contrast (dark regions) is still evident because of the intense signal from the cell cytoplasm (12, 20).

This multimodal nonlinear optical imaging method enables the direct intraoperative observation of EVs in the unperturbed ex vivo human breast tumor microenvironment. The enrichment of EVs identified in the microenvironment surrounding the resected breast tumors from human subjects provided strong evidence for the correlation of observed EV distribution with macroscopic tumor tissue events (12). It was shown that the trend of decreasing EV density correlated well with increasing tumor-to-margin distance, and the EV density determined from tumor specimens of different histologic grades of IDC can be separated by the spatial distribution curves of the EV diffusion model. Unexpectedly, EV density was more dependent on the histologic grade of IDC than on the nuclear grade of DCIS. A possible reason is that the tumor cells in DCIS are not invasive, compared with those in IDC, and hence produce fewer EVs for tumor-stroma interactions (31). Therefore, the lower number of EVs produced by DCIS offers a small contribution to the overall spatial distribution of EVs (31). A further study of EV densities and their distribution in other organ systems, and at sites of tumor metastasis, is expected to reveal more differences between IDC and DCIS in terms of EV production, kinetics, and distribution.

It is noted that in this feasibility study, the imaged tissue volumes for each breast cancer case are relatively small considering the heterogeneity of breast tissue specimens, the physical size of the resected tissue/tumor mass, and the size of the human breast. Therefore, intraoperative delineation of the tumor margin is beyond the capabilities of this imaging system. Furthermore, a direct relationship between EV density and histologic grade has not been established with controlled tumor-to-margin distances due to the limited number of cases with similar tumor-to-margin distances and different cancer invasiveness grades. The application of EV quantification for the assessment of cancer invasion near sites of desmoplasia was also limited by the number and size of images, and thus, we were unable to provide means and standard deviations (SDs) for the quantified EV densities (Fig. 4). The statistical significance can possibly be improved further in the future by implementing a fast-scanning microscope stage along with simultaneous multichannel detection to increase the imaging volume and number of imaged sites within the limited acquisition time permitted in the operating room (~5 min) and by collecting EV density data from more human breast cancer cases of different histologic grades and different tumor-to-margin distances.

The intraoperative imaging and visualization of the tumor microenvironment, and the quantification of EV density, demonstrate the strengths of our portable nonlinear optical imaging system in cancer research and clinical applications. Without the need for tissue fixation, processing, sectioning, staining, and preparation, this imaging system can provide multimodal contrast in real time to identify many tissue features that are diagnostically relevant in cancer, potentially reducing the labor and time costs associated with current surgical breast cancer diagnostics. Uniquely, the label-free in situ visualization of EVs in the ex vivo human breast tumor microenvironment, validated by immunohistochemical labeling and colocated detection (fig. S1), enables the investigation and quantification of the spatially resolved properties of EVs. As a result, the quantification of EV densities revealed relationships with pathological diagnoses, including tumor-to-margin distance and cancer invasiveness. In addition, the EV density and distribution near sites of desmoplasia were shown to be associated with macroscopic mechanisms and processes in carcinogenesis. These results suggest the feasibility and future potential for implementing intraoperative label-free nonlinear optical imaging to investigate the human breast tumor microenvironment and the spatial EV distribution, both to improve our fundamental understanding of carcinogenesis and to potentially provide new biomarkers for tumor invasiveness.


Study design

The primary objectives of this study were to intraoperatively visualize and characterize the human breast microenvironment ex vivo using label-free nonlinear optical imaging and to find the relationship between EV density and pathological diagnoses. Informed consents were obtained from all 29 patients with breast cancer and 7 cancer-free healthy patients participating in this study. During the breast cancer surgeries, the resected fresh human breast tissue specimens were imaged in the operating room with the portable system. Following guidance from the surgeon, the sites imaged on the surgical margin were chosen to be closest to the tumor within the resected specimen. Histological slides and pathological diagnoses were obtained postoperatively for correlation of image features and EV densities. Among the intraoperative images, only those collected from cancer stromal tissue were included in EV density analysis, while the images of pure adipose tissue were excluded. The inclusion criteria for the subjects in this study included a biopsy-proven diagnosis of breast cancer (DCIS or IDC) in need of surgical treatment or elective breast reduction surgery in subjects with no history of cancer. No subjects were excluded on the basis of age or race. The intraoperative imaging and EV density quantification were blindly implemented prior to obtaining the pathological diagnoses of the cancer subjects. The pathological reports were assessed afterward to correlate with the collected multimodal images and the quantified EV densities.

Intraoperative real-time multimodal label-free nonlinear imaging system

A portable imaging system integrating four nonlinear optical imaging modalities (Fig. 5, A to C) was designed to collect label-free multimodal imaging data in the operating room during cancer surgeries. The laser source in this imaging system provided transform-limited 55-fs laser pulses at a 70-MHz repetition rate and with a spectral range of 1040 to 1100 nm, which can excite the four nonlinear optical processes with high efficiency and avoid the potential laser damage and photo bleaching in the tissue specimens. With this excitation spectral range, the emitted nonlinear optical signals from the four modalities were separately detected in different spectral windows (Fig. 5D), achieved by four optical bandpass filters. The color code for each imaging modality was chosen to represent the approximate wavelengths of the emitted nonlinear optical signals from the specimens. Using a pair of galvanometer scanning mirrors, a nonlinear optical image (800 by 800 pixels, 500 by 500 nm2 per pixel) was acquired within 80 s for each imaging modality. By sequentially switching between filters for each imaging modality, a complete set of images from a single field of view (FOV) (400 × 400 μm2) was acquired in approximately 5 min.

Fig. 5 Intraoperative label-free multimodal imaging system.

(A) A photograph of the compact and portable intraoperative label-free multimodal nonlinear imaging system. (B) Software interface of the imaging system. (C) System schematic. (D) Spectral range and display color of the four nonlinear optical imaging modalities. L, lens; GM, galvanometer-scanning mirror; DM, dichroic mirror; OBJ, objective; FW, filter wheel; PMT, photomultiplier tube. (Photo credit: Yi Sun, Biophotonics Imaging Laboratory, University of Illinois at Urbana-Champaign.)

To effectively observe EVs, the lateral resolution of THG imaging (~322 nm) was sufficiently high to visualize and resolve microvesicles (~500 to 1000 nm) and detect some exosomes (40 to 120 nm) (6). Under the spatial Nyquist sampling condition, the pixel size should be set below half of the THG imaging resolution (~161 nm) to assure sufficient sampling. However, because of some residual image jitter caused by inevitable cart vibration from mechanical elements and room equipment, acquiring images with small pixel sizes would considerably reveal vibration-induced artifacts, yield low image quality and fidelity, and increase image acquisition time. Therefore, the pixel size was somewhat compromised and set to 500 nm by 500 nm, which was still sufficient to visualize EVs as diffraction-limited bright points in the acquired images due to the strong THG signal they emitted. Furthermore, to assure the depth-resolved sectioning of EVs, the axial resolution of THG imaging in this imaging system was estimated to be 1.0 μm, which was later used to calculate the imaging volume and EV density.

The entire imaging system was housed in a compact and portable cart (90 cm by 90 cm by 120 cm, 90 kg) that was comparable in size and weight to other intraoperative equipment, such as intraoperative ultrasound, intraoperative optical coherence tomography, and anesthesia carts used within the intraoperative working environment. The system was able to be readily moved throughout the hospital and clinical environment and operated by a single person. The optical components were aligned in a robust way to withstand floor obstacles and vibrations during transportation, eliminating the need for realignment before image acquisition. The system was designed so all optical components and electronics were contained within the cart enclosure, and imaging was performed in an inverted microscope configuration where the specimen was simply placed on a clear glass window inlaid in the top surface of the cart, covered with a light-tight box-shaped cover, and imaged using an objective located within the cart and below the glass window. Because of the strictly controlled lighting conditions in the operating room, light concealment during imaging was considered a priority when the imaging system was designed and built, so as to obviate the need for the surgeon and staff to change lighting conditions during surgery and delay the procedure. Specifically, switching between modalities was automated using a motorized filter wheel, and focus adjustment was accomplished by a piezoelectric linear stage to eliminate the need to open the cart doors. As a result, noise from background light was minimized, and the laser beam was confined within the imaging cart for laser safety. On the other hand, necessary ventilation was maintained to reduce the thermal noise of the PMT (H7421-40, Hamamatsu Photonics K.K.). With the minimized background and thermal noise, the average signal-to-noise ratio (SNR) measured from the intraoperative THG images was approximately 19 ± 4 dB, sufficient to visualize the tissue structures.

EV segmentation

On the basis of the characteristics of EVs in the THG-contrast images, an automated segmentation algorithm (Fig. 3A) was developed to extract the EV signal from the background and subsequently quantify EV density. The principles of this segmentation algorithm relied on the spatial and nonlinear optical properties of the EVs: that they are small (40 to 2000 nm) (6), point-like, and generate exceptionally strong THG signal due to the good phase-matching condition provided by their large surface-to-volume interface ratio (28). Therefore, an intensity threshold was used to segment the EVs from the background (Fig. 3A). This threshold was automatically set to be the pixel intensity value at a fixed percentage out of the intensity histogram generated from each THG image, and this specific percentage for all images was deliberately determined to leave only the in-focus EVs while suppressing the background noise. By applying this algorithm to THG-contrast images of the tumor microenvironment (Fig. 2C), binary images (Fig. 2D) were generated to reveal the spatial distribution and density of the EVs. These black points in the binary image were subsequently quantified to represent the density of EVs in each FOV.

To validate this imaging and segmentation method for EV detection and quantification, we acquired THG-contrast images of EVs purified from human cancer cell lines with a known density of 3 × 1010 ml−1, measured by a standardized technique (40) with a commercial instrument (NS3000, NanoSight Ltd.). A representative example of a THG-contrast image of purified EVs was processed to highlight the EVs (Fig. 3B) using this segmentation algorithm. Considering the axial resolution and imaging FOV, the three-dimensional imaging volume of each image of purified EVs was approximately 100 μm by 100 μm by 1 μm = 10−8 ml. The density of EVs was then calculated to be 1.5 × 1010 ml−1 using the EV counts (152 ± 10), quantified from five THG-contrast images of purified EVs. To explain the density discrepancy, it is likely that some EVs underwent refractive index matching due to the diffusion and permeation of glycerol through the EV membrane and, thus, ceased to provide the phase-matching condition necessary for THG signal generation. Nevertheless, the EV density quantified from the THG images using the segmentation algorithm was of the same magnitude as the known density of EVs measured by the commercial instrument. Furthermore, the isolated EVs were mixed with human cells in culture. The drastic increase of EV density after mixing was identified by the quantification algorithm based on the THG images (fig. S2) and validated by the NanoSight measurement. Therefore, the EV densities quantified from the intraoperative THG-contrast images faithfully represented the distribution and density of EVs in the tumor microenvironment.

EV diffusion model

To help explain and demonstrate the relationship between EV density and tumor-to-margin distance, the distribution of EVs in the tumor microenvironment was treated as a process of diffusion for small particles (41). Assuming that the EV distribution is spherically symmetric and emanating from the center of the tumor and that the tumor cells are the only major source of EVs, we can derive the time-dependent three-dimensional diffusion equation to describe EV density C(r, t) with the boundary conditionsEmbedded Imagewhere D is the diffusion coefficient, r is the radial distance from the tumor boundary, t is time, and A is the initial number of EVs. Under the same assumptions, the boundary conditions for this partial differential equation were set to be a constant value A at the tumor boundary and zero at infinite distance. As for the initial condition of the EV density distribution, functions such as an exponential decay function, power function, and Gaussian function were compared with the fitted results of EV density versus tumor-to-margin distance.

The exponential decay function was chosen as an initial condition to generate the region boundary curves shown in Fig. 3D. Considering the much longer time scale of tumor growth (42) compared with EV diffusion (41), it was also assumed that EV diffusion was always at an equilibrium state (changing extremely slowly over time) at the time the tissue specimens were imaged. The boundary condition value A was used to represent the cancer invasiveness based on the observation that more aggressive cancer cells tend to produce more EVs at the boundary (31). The two EV distribution curves calculated from two different values of A served as the boundary curves of the EV density data points from cases with different histologic grades of IDC (Fig. 3D).

Multiway ANOVA

Multiway ANOVA (anovan; MATLAB) was performed to examine the relationship between EV density and the corresponding grade of IDC/DCIS. Noticing that the tumor-to-margin distance also contributes to the EV density, we only included in the multiway ANOVA the cases with a tumor-to-margin distance between 0 and 8 mm to minimize the contribution by distance and maintain a sufficient number of data points. Statistically significant correlation was found between the EV density and the histological grade of IDC (P = 0.0002), but not with the nuclear grade of DCIS (P = 0.1835). Furthermore, with statistical significance existing between the EV density and the histologic grade of IDC, a multiple comparison test (multcompare, critical value: “tukey-kramer,” MATLAB) was used to analyze the differences of EV densities between different histologic grades of IDC (Fig. 3E).

Human tissues

Human tissues were obtained under a protocol approved by the Institutional Review Boards at the University of Illinois at Urbana-Champaign and Carle Foundation Hospital, Urbana, Illinois. A total of 29 breast cancer human subjects and 7 healthy (and no history of cancer) human subjects undergoing breast reduction surgeries were included in this study (Table 1). During the breast cancer surgeries, and immediately following resection, the fresh human breast tissue specimens were directly passed to the research team member in the operating room for ex vivo imaging. On the basis of guidance from the surgeon, the location on the surgical margin surface that was deemed to be closest to the tumor within the resected mass was identified and selected as the site from which images were collected. The distance between the imaged surgical margin surface and the tumor mass (tumor-to-margin distance) was later measured postoperatively by the pathologists.

Table 1 Demographic data on human subjects and pathological diagnoses.

A total of 29 breast cancer subjects and 7 healthy cancer-free subjects undergoing breast reduction surgery were included. Healthy subjects are indicated by NML (normal) under the histologic grade column.

View this table:

The multimodal label-free images of the unperturbed human breast specimens were acquired within a time window of less than 30 min between the time of surgical excision and tissue fixation for histopathological processing, without disrupting or delaying the surgical procedure. Following imaging, the imaged regions on the surgical margin surfaces of the breast tissue specimens were marked with surgical ink for later registration with histological slides for image feature correlations. After intraoperative imaging, the intact specimens were sent to the pathology laboratory for standard processing and diagnosis.


Supplementary material for this article is available at

Fig. S1. Validation of THG imaging of EVs by immunohistochemical-based detection.

Fig. S2. Increase of EV density in cell culture by adding isolated EVs.

This is an open-access article distributed under the terms of the Creative Commons Attribution-NonCommercial license, which permits use, distribution, and reproduction in any medium, so long as the resultant use is not for commercial advantage and provided the original work is properly cited.


Acknowledgments: We thank Carle Foundation Hospital and its physicians, surgeons, nursing, and research staff for their clinical collaboration and assistance with this translational research study. Funding: Research reported in this publication was supported by the National Institute for Biomedical Imaging and Bioengineering and the National Cancer Institute of the NIH under award numbers R01EB023232, R01CA166309, and R01CA213149. One hundred percent of the total project costs was financed with federal money, and 0% of the total costs was financed by nongovernmental sources. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH. This work was also supported by an award from the Cancer Scholars for Translational and Applied Research (CSTAR) program of Carle Foundation Hospital, the University of Illinois at Urbana-Champaign, and the Cancer Center at Illinois. Author contributions: H.T. and S.A.B. conceived the project of intraoperative nonlinear optical imaging. Y.S., S.Y., and D.R.S.J. designed and built the portable imaging system. Y.S., S.Y., and J.W. conducted imaging in the operating room. E.J.C., M.M., J.L., and R.B. assisted with image collection and communicated with surgeons, nurses, and pathologists for specimen handling and pathological reports. M.M. and S.A.B. wrote the protocol for intraoperative imaging of breast cancer specimens. E.J.C. prepared the histological slides from imaging sites. Z.G.L. provided expert advice on pathological interpretation of the image data. A.M.H. and K.A.C. performed the breast cancer surgeries and provided surgical guidance on tumor localization. N.N.L. performed the breast reduction surgeries. Y.S. obtained histological correlations for the intraoperative images and analyzed EV density data. Y.S. and S.A.B. wrote the manuscript. Competing interests: H.T., S.Y., and S.A.B. are inventors on patents filed by the University of Illinois at Urbana-Champaign (2018/0286044 A1, 4 October 2018) related to the laser source technology and the imaging and quantification of EVs. All other authors declare that they have no competing interests. Data and materials availability: All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. Additional data available from authors upon request.

Stay Connected to Science Advances

Navigate This Article