100-Second Large-Area Chemical Imaging by Integrated Four-Channel CARS Microscopy

Abstract

Label-free coherent Raman scattering (CRS) microscopy enables chemical observation of intricate cellular and intercellular structures of biological tissues with submicron resolution and has become the most prominent stain-free alternative to gold-standard hematoxylin and eosin (H&E) histology. However, the imaging field of view (FOV) of current CRS microscopes is typically limited to less than a millimeter. Various attempts in instrumentation have been made to perform large-scale CRS mapping across an intact tissue, but the costly trade-offs are either imaging speed or spatial resolution. To address this challenge, we developed a four-channel coherent anti-Stokes Raman scattering (CARS) microscope, offering submicron imaging of 13 × 13 mm² large tissues within 100 s. Four miniaturized CARS systems, each equipped with a microelectromechanical system (MEMS) scanner, a micro-objective, and a collection fiber bundle, were fully integrated to perform ultrafast chemical mapping of intact unprocessed tissues with an electronically synchronized motorized stage. We demonstrated its capability by performing large-scale CARS imaging of mouse brain tissues and human glioma tissues. This label-free technique may pave the way for future clinical applications of stain-free pathology in the operating room.

Introduction

During tumor resection surgeries, the rapid intraoperative pathology of resected neoplastic tissues plays an indispensable role in accurately confirming benign or malignant, the subtype and stage, the infiltrated margins of a tumor, and lymph node metastasis, guiding further surgical treatments. The routine gold-standard formalin-fixed and paraffin-embedded (FFPE) histology can produce flawless morphologies of cancer cells in diseased tissue, but the complete process, demanding 24–36 h, is time-consuming. The current frozen-sectioning and the following hematoxylin and eosin (H&E) staining can provide fast pathological results in less than 30 min. However, ice-freezing induces artifacts, including cellular distortion, tissue fragmentation and loss, folding, and poor staining, frequently resulting in ambiguous morphology, inaccurate tumor grading, false-negative margin assessment, deferred diagnosis, and so forth. Moreover, the 30 min of intraoperative frozen-sectioning pathology is still far too long to carry out two or more rounds’ searching of critical surgical margins for maximal resection of a tumor in practice. A variety of emerging concepts of stain-free and sectioning-free pathology have been widely explored to replace the traditional H&E histology. Two-photon and multiphoton microscopy showed promising capability to yield H&E like images, but these nonlinear modalities provide poor morphology of the cell nucleus. Ultraviolet excitation based autofluorescence, , photoacoustic, , wide-field autofluorescence, , and other imaging modalities have exhibited various contrasts indicating the cell morphology, but some features may be interfered by the native pigments, fluorophores, or chromophores. Over the past decade, the coherent Raman scattering (CRS) microscopy, including stimulated Raman scattering (SRS) − and coherent anti-Stokes Raman scattering (CARS), − has emerged as a powerful tool in biomedical imaging due to its label-free nature, chemical specificity, high sensitivity, quantitative imaging of chemical concentrations, and capability for optical sectioning. Moreover, in clinical trials, SRS has made impressive progress in the direct and instant generation of histologic images of human cancerous tissues. With the aid of disruptive iteration power of the artificial intelligence (AI), , advanced methods like deep learning − and its widely used convolutional neural networks (CNNs), generative adversarial networks (GANs), , and present foundation models all play indispensable roles along this evolutionary road, transforming chemical specific images to near-flawless virtual H&E histology. Due to free of staining and sectioning, SRS histology has largely relieved the critical requirement of imaging speed. For instance, Zhang achieved large-area and two-color SRS imaging of a 12 × 7 mm² tissue within 8 min. This advancement in SRS imaging speed has significantly reduced the decision-making time for surgeons, but not yet sufficient to fill the gap between CRS-based virtual histology and real-time bedside pathology, which eventually hinges on whether the pathology results can really be delivered to the surgeons in 1 or 2 min.

In this study, we developed a miniaturized CARS microscopic system for fast large-scale tissue mapping. To improve the imaging speed, the galvanometer in the traditional raster-scanning CARS microscope was replaced by microelectromechanical systems (MEMSs), which can work in fast resonant mode. − To reduce the size of the imaging head, we utilized a micro-objective instead of the conventional objective. Thus, four miniaturized CARS systems were synchronized and integrated together with a motorized stage to establish a high-speed, large-FOV CARS imaging platform, which could map a 13 × 13 mm² large tissue in just 100 s, excluding excess time for data saving and processing. This technology provides a novel solution for rapid intraoperative pathological diagnosis and holds significant potential for advancing clinical imaging microscopy.

Results and Discussion

The complete imaging system is based on a commercially available 1032 nm femtosecond laser with a 260 fs pulse width and a 1 MHz repetition rate, which also serves as the Stokes laser. For CARS, the pump laser at a wavelength of 792 nm was generated by an optical parametric amplification (OPA) system, and the frequency difference with the Stokes laser matched the C–H₃ vibration at 2930 cm^–1. As shown in Figure a, the combined pump and Stokes lasers were divided and delivered into four miniaturized CARS imaging heads, which were integrated and spaced apart from each other at a distance of 6.5 mm. The spectral widths of the pump and Stokes before the glass rods are 9.7 and 9.1 nm, respectively, and the pulse widths were 583 and 439 fs, respectively (Figure b,c). After beam combination, the two beams pass through four glass rods to broaden the pulse duration to picosecond for imaging. To test the proposed spectral focusing system, we measured the CARS spectrum for dimethyl sulfoxide (DMSO) within a spectral window between 2750 and 3100 cm^–1 (Figure S1). Figure S2a presents the picture of the whole system including four channels of CARS imaging heads, MEMS drivers, PMTs, amplifiers, etc. Figure S2c shows the four-channel CARS microscope, equipped with four MEMS devices and four × 3/0.5 NA micro-objectives.

1.

System design. (a) Schematic diagram of the optical setup. HWP: half wave plate; PBS: polarizing beam splitter; M: mirror; TD: time delay; DM: dichroic mirror; F: filter; BBO: beta barium borate; YAG: yttrium aluminum garnet; PPLN: periodically poled lithium niobate; OPA: optical parametric amplification; R: rod; BS: beam splitter; MEMS: microelectromechanical systems; PMT: photomultiplier tubes. A 1032 nm pulsed laser served as the fundamental source for an OPA system, which produced a 792 nm pump beam and a 1032 nm Stokes beam. An epi-detection scheme was employed for CARS signal collection. (b) Spectral width and pulse width of a 792 nm pump beam generated by the OPA; the spectral width of the pump is 9.7 nm, and the pulse width of the pump is 583 fs. (c) Spectrum and pulse width of the 1032 nm Stokes beam; the spectral width of the Stokes is 9.1 nm, and the pulse width of the Stokes is 439 fs.

The scheme of the miniaturized laser scanning system retains the core structure of the MEMS system reported in the two-photon microscope in ref , which comprises a MEMS mirror, a scanning lens, and a micro-objective (Figure a). During the scanning process, we observed an inherent scan-induced distortion from the MEMS, manifesting as a pronounced fan-shaped laser trajectory when both scanning axes employed MEMS scanning (Figure b). This distortion was clearly evidenced when imaging a standard grid micrometer. We found that this originated from the oblique incidence (45°) of the laser beam with respect to the reflective surface of the MEMS (Figure c). The grid micrometer images revealed that the MEMS scanner introduces a curved distortion only along the Y-axis, while the X-axis scan remains linear. To mitigate this image distortion, we adopted a hybrid scanning approach; the MEMS device performs the fast-axis scanning along the X-axis, while a motorized translation stage is responsible for the slow-axis scanning along the Y-axis. Figure d,e shows that the imaging distortion can be well-fixed with this method. After correction of mapping distortion, we performed CARS imaging of human glioma tissues. Although the nonresonant background was the major interfering factor, the neural axons with rich lipids were clearly observed (Figure f).

2.

Distortion correction of the miniaturized CARS microscope. (a) Miniaturized CARS microscope assembly with the 0.5 NA micro-objective. (b) 2D laser scanning pattern generated by MEMS. (c) X–Y MEMS scanning of a micrometer target. (d) Micrometer imaging with Y-axis MEMS scanning, showing pronounced distortion. (e) Micrometer imaging with X-axis MEMS scanning, free of distortion. (f) CARS imaging of human glioma tissues (2930 cm^–1). Scale bar: 50 μm.

To perform microscopic imaging of larger tissues, four miniaturized CARS microscopes are tightly arranged shoulder-to-shoulder in a geometry of 2 × 2 (Figure a). The micro-objectives are spaced 6.5 mm apart, and each of them is responsible for imaging an area about 6.5 × 6.5 mm². As shown in Figure b, each CARS microscope employed a MEMS mirror for fast resonant sinusoidal bidirectional scanning along the X-axis and a motorized stage for slower continuous scanning along the Y-axis. Each scan acquires a stripe area of 350 × 6500 μm². Then, the motorized stage moves along the X axis 19 times, covering 6.5 × 6.5 mm². Because all four channels work simultaneously, they collectively achieve an imaging field of view of 13 × 13 mm² (Figure b). To validate the optical spatial resolution of our system, we performed CARS imaging of polystyrene beads with 100 nm diameter (Figure S3a,b) displays the intensity profile and the fitting curve of a polystyrene bead. The full width at half-maximum (fwhm) of the Gaussian fit measures the diameter of the bead with 100 nm to be 839 nm; therefore, we estimated the spatial resolution of our system is around 800 nm by point spread function (PSF).

3.

Four-channel microscopic imaging system. (a) Illustration of the four-channel microscopic imaging construction with four micro-objectives. (b) Large-area rapid scanning methodology via four-channel parallel imaging. Obj.: micro-objective.

Following the system validation and image quality optimization, mouse brain imaging was conducted on our platform. Figure a shows single-channel CARS mapping of a complete mouse brain section along the coronal plane with an imaging area of 13.5 × 7 mm². Each strip covers an area of 350 μm × 7 mm and takes about 5 s, consisting of 350 × 7000 pixels with a pixel dwell time of 2 μs. Total 41 strips were recorded in 205 s, depicting the intact mouse brain section with great details of the cerebral cortex, white matter, hippocampus, axonal bundles, and other structures. Figure b presents the CARS imaging results of an independent coronal hemisection of a mouse brain, with the white dashed box revealing the structure of the internal capsule shown in Figure c. To generate a homogenized tissue image, a pure DMSO sample was imaged under identical conditions to serve as the background for subsequent image normalization. Image registration was further applied to mitigate the stitching stripes without degrading the image resolution (Figure S4a). Large-area imaging of brain tissues was performed using single-channel CARS microscopy and compared with H&E staining results. The tissues used for CARS imaging and H&E staining were two consecutive adjacent sectioned slides. CARS clearly revealed the line of Gennari (indicated by yellow curves in Figure a), and cell density is consistent with that observed in the H&E-stained section. Moreover, CARS identified detailed histopathological features, including microvessels, neural cells (Figure b), which are consistent with those in H&E staining. Figure S5a,b shows the photograph and dual-channel CARS imaging results of the glioma, respectively. These results verify the feasibility of simultaneous dual-channel CARS imaging for large tissues.

4.

Single-channel CARS imaging of mouse brain sections. (a) CARS imaging results of the mouse brain section along the coronal plane. Scale bar: 1 mm. (b) CARS imaging results of another coronal hemisection from a mouse brain. Scale bar: 1 mm. (c) Zoom-in image corresponding to boxed regions in (b), showing the cellular structure in the region of the internal capsule. Scale bar: 100 μm.

5.

Single-channel CARS and H&E imaging of the human brain tissue. (a) CARS and H&E imaging results of adjacent sections of the human brain tissue. Scale bar: 1 mm. (b) Zoom-in image corresponding to the region outlined by the blue dashed box in (a), showing neural cells (blue arrow). (c) Zoom-in image corresponding to the region outlined by the green dashed box in (a), showing microvessels (green arrow). Scale bar: 100 μm.

Further, we performed four-channel label-free CARS imaging of a piece of the brain tumor tissue, which was sliced to 100 μm thickness by an oscillating blade microtome. Figure presents a total FOV of approximately 13 × 13 mm², with a total acquisition time of about 100 s, excluding additional data-processing steps. Each of the four channels imaged an area of about 6.5 × 6.5 mm² on the upper-left, lower-left, upper-right, and lower-right quadrants, respectively. The higher signal in the lower-left region corresponds to nonresonant background of water at the tissue edge.

6.

Four-channel CARS imaging of the glioma tissue. The total imaging area of 13 × 13 mm² and a total acquisition time of 100 s; the entire process, including data processing, took approximately 3 min. Scale bar: 1 mm.

Importantly, our Epi-CARS system also works for the high-resolution imaging of thick, fresh glioma specimens. Figure S6a presents an intact piece of the unprocessed glioma tissue, and we performed CARS imaging of white and black squared regions as indicated. Figure S6b,c exhibits distinct high and low signals in these two regions. The zoom-in image within the high-intensity region reveals highly abundant nerve fiber distribution (Figure S6d), while Figure S6e shows numerous tumor cells along with lipid distribution, and Figure S6f displays abundant protein-rich morphological structures. For comparison, tissues from three regions were imaged using the 2930 cm^–1 Raman channel (Figure S6g). Details of the SRS system have been described in our previously published work.

Conclusions

We developed a miniaturized CARS microscopy imaging system based on synchronized MEMS scanning and a micro-objective. By integrating four compact CARS microscopes into a multichannel configuration, we achieved a high-speed, large-area CARS imaging platform for large unprocessed tissues. This system demonstrates the capability of acquiring CARS images of 13 × 13 mm² tissue samples within just 100 s. The system exhibits a decent optical resolution of 839 nm across such a large area. Beyond CARS imaging, the platform is also compatible with other optical imaging modalities, including two-photon excitation fluorescence (TPEF) and second-harmonic generation (SHG). With the unique combination of rapid scanning, label-free imaging, submicrometer resolution, and a large FOV on unprocessed tissues, this system shows great potential for clinical pathological diagnosis in the future.

Methods

Main Optical System and Alignment

In this study, we established an optical parametric amplification (OPA) system using a femtosecond laser (KY-Vibre, Keyun) with a repetition rate of 1 MHz, a total power of 4 W, a pulse width of 260 fs, and a central wavelength of 1032 nm. We implemented a beta barium borate crystal (BBO, HG Optronics) to double the frequency of the 1032 to 516 nm laser, which served as a pump beam for the OPA system. The 1032 nm fundamental and 516 nm second-harmonic beams were then separated by a dichroic mirror (DMSP950, Thorlabs). The 1032 nm beam was subsequently focused into a yttrium aluminum garnet crystal (YAG, HG Optronics) to generate a supercontinuum, from which the 792 nm component was filtered using a bandpass filter (FBH800–40, Thorlabs) and used as the seed for the OPA. The 516 nm pump and the 792 nm seed beams were combined by another dichroic mirror (DMSP650, Thorlabs) and directed into a periodically poled lithium niobate crystal (PPLN, HC Photonics) for amplification. Finally, an output power of 110 mW at 792 nm was achieved. The 1032 nm laser served as the Stokes beam, while the 792 nm laser acted as the pump beam for CARS imaging. To achieve spectral focusing for CARS imaging, the combined beams were directed through four glass rods (SF57, HG Optronics), each 12.5 cm in length, to broaden their pulse duration to the picosecond scale (Figure a).

In the CARS imaging head, the combined beams undergo point scanning via a 1.3 kHz MEMS mirror (A7M20.2, Mirrorcle) and a scanning lens with an approximately 4 mm focal length (D0166.101, Domilight) and finally focus onto the sample through a micro-objective (D0213–3X, Domilight). All dichroic mirrors (DMR-680SP-SP, LBTEK), lenses, and micro-objectives in the miniature microscope were fixed using UV-curable adhesives. Two adjustable mirrors were applied to ensure that the laser beams could enter the system coaxially. Two-photon emission fluorescence imaging of R6G was performed to align the system. The MEMS scanner was mounted on a five-axis adjustable stage via a custom adapter, allowing precise adjustment of its 3D position and 2D tilt angles to maximize and center the two-photon signal. The MEMS was then permanently fixed to the microscope body by using UV-curable adhesives. Epi-CARS signals were collected by a customized multimode fiber bundle (0.62 NA; 3 mm in diameter). After transmission through the fiber bundle, the signal passed through optical filtering (FBH650–40, Thorlabs) and collimation modules before reaching the PMT detector (H7422–40, Hamamatsu). The PMT output current was converted to a voltage signal by a custom-designed amplifier with an optimized frequency response (2 μs) and further digitized by a data acquisition card.

To assemble a four-channel CARS system, the combined pump and Stokes beams were split into four channels using two 50:50 nonpolarized flat beam splitters (BS2R55-B, LBTEK). Each beam was directed into a miniature microscope via two adjustable mirrors. To compensate for the temporal delays between pump and Stokes beams introduced by the BS, optical windows were added to equalize the path lengths across all channels. The four pairs of laser beams were independently aligned into four CARS imaging heads equipped with four identical micro-objectives (NA = 0.5). The CARS signals generated from these four systems were collected by four multimode fiber bundles and detected by four separated PMTs. A custom sample holder with four adjustable set screws was designed to tilt the sample to match the focal plane. The holder was mounted on a vertical translation stage for coarse Z-axis adjustment.

Laser Settings and Data Acquisition

The laser powers were measured after the micro-objective for all experiments. For CARS imaging of the grid micrometer, the pump and Stokes lasers were set at 792 nm (1 mW) and 1032 nm (1 mW), respectively. The field of view was 300 × 300 μm² with 300 × 300 pixels (Figure c); 350 × 350 μm² with 350 × 350 pixels (Figure d); and 450 × 450 μm² with 450 × 450 pixels (Figure e), and the dwell time for imaging was 2 μs/pixel. For CARS imaging of biological samples, the laser powers were 5 mW for both the pump and Stokes beams. The dwell times and pixel settings were 10 μs/pixel and 400 × 400 pixels for glioma (Figure f); the field of view and pixel settings were 1 μm/pixel and the dwell time for imaging was 2 μs/pixel (Figures a,b, , , S4a,b, S5b, and S6b,c). For nanoparticle imaging, the pump and Stokes lasers were set at 792 nm (2 mW) and 1032 nm (2 mW), respectively. As shown in Figure S3a, the field of view was 5 × 5 μm² with 30 × 30 pixels, and the dwell time for imaging was 10 μs/pixel. For SRS imaging of glioma (Figure S6g), we applied a water immersion objective with N.A. 1.2, with pump and Stokes beam laser powers of 50 (800 nm) and 80 mW (1045 nm), respectively. The field of view was 200 × 200 μm² with 800 × 800 pixels, and the dwell time was 10 μs/pixel.

Sample Preparation

The nanoparticles (L815936, Macklin) were diluted to an appropriate concentration by using deionized water and resuspended via vortexing. The nanoparticle suspension was then sonicated for 20 min at room temperature to achieve a homogeneous solution, followed by a 1:1000 dilution with deionized water. Prior to imaging, a droplet of the diluted nanoparticle solution was placed on a glass slide and allowed to dry. The dried nanoparticles were subsequently sealed between two glass coverslips.

The study was approved by the Institutional Ethics Committee of Beijing Ditan Hospital, Capital Medical University, with written informed consent (approval no. DTEC-KY2025- 066–01). Human tissues samples and mouse brains were fixed in 4% paraformaldehyde (BL539A, Biosharp) for 24 h, sectioned into 100 μm slices using a vibratome (VT1200, Leica), and washed once by PBS (BL310A, Biosharp) before sealing glass slides (Citotest) and coverslips (VWR) with double-sided adhesives. The residual glioma tissue was processed for formalin-fixed and paraffin-embedded H&E staining (Figure a). In Figure S6, the fresh glioma tissue is simply rinsed with PBS, placed on a glass slide, and sealed between a slide and a coverslip using double-sided adhesive tape. The complex preparation procedures such as fixation and sectioning are unnecessary.

Synchronization and Data Acquisition

Four data acquisition cards (PCIe-6251, NI) were employed to independently sample the signals of four channels. The DAQ cards were synchronized through jumper wire configurations. The clock of the arbitrary function generator was synchronized by the 1 MHz clock signal from the laser. The DAQ cards that control the MEMS scanning and CARS signal acquisition were synchronized by the function generator, ultimately achieving synchronization between the laser pulses, MEMS, and all acquisition cards.

Image Distortion Correction

Since the scanning signal provided to the MEMS mirror follows a sinusoidal waveform, the dwell time of the laser focus on the sample varies spatially. To correct the image distortion, we proposed a pixelwise algorithm, which redistributes the integration time of the pixels. Specifically, the pixels corresponding to fast mirror scanning (i.e., near the linear region of the sine wave) are assigned a longer integration time, whereas those in slow scanning regions (i.e., near the sine wave peaks) receive proportionally shorter integration time. This dynamic compensation effectively calibrates the sinusoidal distortion in the acquired images.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/cbmi.6c00016.

• CARS spectrum for DMSO (Figure S1); the system construction (Figure S2); imaging resolution of miniaturized objectives (Figure S3); image normalization (Figure S4); dual-channel CARS imaging of glioma (Figure S5); and comparing CARS and SRS imaging of fresh glioma (Figure S6) (PDF)

W.J. and Z.H. conducted all experiments and analyzed data. S.Y. built the OPA system. C.Y. helped on biological sample preparation. Y.Z. performed image stitching and corrected image distortion. W.J., Z.H., P.W. discussed all results and wrote the manuscript with input from all authors. P.W. conceived the idea and designed experiments. P.W., J.W., X.Y. supervised the project. W.J. and Z.H. contributed equally.

The authors declare no competing financial interest.