Unified Vibrational and Multiphoton Label-Free Nonlinear Microscopy for Simultaneous Chemical and Structural Imaging

Abstract

Nonlinear microscopy enables label-free imaging by deriving contrast from the intrinsic spectroscopic responses of specimens, thereby offering a valuable tool for biomedical applications. Often, clinical imaging systems implement either multiphoton or vibrational contrast, not both. Consequently, most clinically deployed label-free nonlinear microscopes lack a robust, complementary contrast palette suitable for investigating the morphofunctional features of heterogeneous specimens. This deficiency limits not only the analytical capabilities of the nonlinear microscope but also its diagnostic utility. A main reason for this disregard is that the various imaging modalities impose distinct and stringent requirements on the excitation source and the detection chain. In this contribution, we propose a strategy that targets both multiphoton and vibrational contrasts to achieve a robust, complementary contrast palette. The approach emerges from a systematic investigation of readout schemes and provides engineering criteria to tailor the detection chain and thus maximize quantitative performance. In concert with this detection strategy, we present a compact laser source that drives vibrational coherences while simultaneously exciting multiphoton signals. We validate the resulting imaging platform using two rodent case studies: one involving a naturally occurring metastatic cancer in a mouse and another relying on an allogeneic mammary cancer model in a rat. Owing to its dimensions, cost, and versatility, we anticipate that this biophotonics tool will readily find its way into clinical applications.

I. Introduction

THE nonlinear optical microscope is a powerful analytical tool [1]. These platforms may yield spectroscopic information, and also derive contrast directly from fundamental properties of matter, thus relying on intrinsic optical signals to reveal the architecture and composition of specimens [2], [3]. This nonlinear microscope can therefore enable label-free imaging, supporting investigations of biological specimens without the need for exogenous labels [4], [5]. Such label-free capability is of great value in clinical applications, including medical procedures and image-guided surgery, for it not only expedites the analysis of specimens but also avoids the labor-intensive workflows and potential artifacts of tissue fixation, processing, sectioning, and staining [6], [7], [8], [9].

Equally important, the nonlinear optical microscope has the potential to simultaneously interrogate a specimen through multiple intrinsic nonlinear signals [10], [11], [12], [13], [14], [15]. By virtue of their heterogeneity, biological specimens mediate not one but several such contrasts, enabling yet another benefit of label-free nonlinear microscopy, namely, multimodal imaging, a feature that confers not only microstructural imaging, but also chemical specificity. Thus, a multimodal nonlinear microscope supports the spatial separation of different constituents within a sample based on their spectroscopic properties [16], [17]. This separation enables quantitative chemical imaging via spectral unmixing and statistical analysis, revealing latent patterns in downstream data-processing pipelines [18], [19].

Frequency-shifted nonlinear optical signals often form the foundation of the contrast palette of a label-free nonlinear optical microscope [20], [21]. In the spectroscopic context, frequency-shifted signals are detected in field modes that are initially vacant and therefore emerge at optical frequencies – wavelengths (colors) – different from those of the driving fields. Key frequency-shifted contrasts include second-harmonic generation (SHG) [22], multiphoton absorption fluorescence (MPAF) [23], and coherent anti-Stokes Raman scattering (CARS) [24]. SHG is highly selective for non-centrosymmetric structures (e.g., collagen, actomyosin) and is sensitive to structural changes in the extracellular matrix (ECM) [25], a tissue compartment that undergoes drastic remodeling during the progression of diseases. While several endogenous fluorophores inhabit the tissue microenvironment [26], nonlinear microscopy typically targets metabolic cofactors (NAD(P)H and FAD) whose emissions report on cellular redox state, enabling label-free readouts of cellular metabolism and heterogeneity [27]. CARS directly probes molecular bonds and is therefore well attuned to chemical composition [28], [29]. By driving specific vibrational resonances, most notably CH₂/CH₃ stretches, CARS enables rapid, label-free mapping of lipidic species and protein distributions [30], [31]. While CARS suffers from a non-resonant background [32], it is amenable to epi-detection and scalable multiplexing, making it well suited for fast exploration of freshly excised tissues or even in vivo.

Manifestly, a nonlinear microscope capable of co-registering this triad of contrasts (SHG,MPAF,CARS) offers a valuable tool for investigating the structural and chemical heterogeneity of disease within tissues, particularly cancer, a set of disorders that often exhibit both chemical and structural alterations along with inter-patient and intra-tumoral heterogeneity [33]. This contrast palette supports margin identification, grading and subtyping of resected tissues, and may even aid in predicting patient response to therapy [20]. Furthermore, the high imaging frame rates of this approach fit critical intraoperative timelines, while its label-free nature preserves tissue for downstream assays, stained histology, and immunohistochemistry. Therefore, the co-registration of multiphoton (MPAF, SHG) and vibrational (CARS) contrasts offers a powerful instrument for intraoperative assessments in surgical oncology and point-of-care procedures, a biophotonics tool that has the potential to truly revolutionize optical biopsy [34] and, in turn, make a tangible impact on human health.

Although informative, the nonlinear optical signals inherent to specific biochemical species are often weak. This weakness imposes stringent technical demands on the imaging platform, most notably on its driving source and on the detection chain. In fact, these demands constrain the nonlinear contrast palette, because parameters that favor one contrast (e.g., wavelength, pulse duration, peak intensity, detection bandwidth) can jeopardize others. Consider, for example, that CARS typically relies on picosecond (ps) pulses to attain high spectral resolution, whereas multiphoton contrasts require high peak powers achieved with ultrashort femtosecond (fs) pulses.

Consequently, the vast majority of state-of-the-art systems capture either vibrational [35], [36], [37] or multiphoton contrasts [38], [39], [40], not both. Such omission leads to a contrast palette that lacks complementarity, limiting co-registration of molecular and structural features, and reducing the ability of clinical imagers to resolve pathological states. Arguably, imaging systems capable of acquiring both multiphoton and vibrational contrasts are not yet clinically practical: depending on the design, they may rely on bulky, costly laser sources; use optical fibers with finite lifetimes that require regular replacement; or call for technical personnel to oversee and regularly maintain the equipment. These factors make it difficult to keep pace with clinical timelines and patient workflow, and most importantly, calibration, reliability, and safety requirements.

In this paper, we present a strategy for the generation and detection of frequency-shifted nonlinear optical signals, a strategy that targets both multiphoton and vibrational contrasts, aiming to reveal the morphofunctional properties of the pristine tissue/tumor microenvironment. This strategy results from a systematic investigation and refinement of the imaging system hardware, beginning with the performance of different types of photomultiplier tubes (PMTs) together with various readout schemes (analog, lock-in, and photon counting), which are the bases for optimizing the detection of the weak nonlinear signals. Following this, we provide engineering criteria to tailor the detection chain and thus maximize quantitative performance. In concert with this detection strategy, we present a refinement of a novel laser source introduced by Boppart and colleagues [41], which shifts the optical source from a fiber-based oscillator to a ytterbium (Yb)–based solid-state laser. This source provides both the frequency detuning required to drive the vibrational coherences that elicit the CARS contrast and the excitation for multiphoton signals. We evaluated this strategy in two rodent cancer case studies: a naturally occurring metastatic carcinoma in a mouse and an induced, allogeneic mammary carcinoma in a rat. For validation, we conducted two studies to compare the resulting label-free images of mammalian tissues with histology, the gold standard in diagnostic pathology. Because this strategy yields a compact, cost-effective, and robust biophotonics imaging platform, we believe it not only sets a precedent for broad clinical translation but also offers a practical venue toward adoption as a label-free complement to the current standard of care of stained histopathology.

II. Photomultiplier Tubes and Readout Schemes for Label-Free Nonlinear Microscopy

Because of their single-photon sensitivity, high electronic bandwidth (fast temporal response), and low dark counts, photo-multiplier tubes (PMTs) are the detectors of choice for recording nonlinear signals in label-free nonlinear microscopy [42]. These vacuum-tube light detectors convert ultra-weak light into measurable electrical signals with high gain. PMTs can be read out in two ways, analog (current, AO) and photon counting (digital, PC), and are classified accordingly.

A. Analog Output PMTs and Their Synergy With the Lock in Technique

AO-PMTs measure the continuous anode current $I$. The signal-to-noise ratio (SNR) of this type of PMT is given by [43], [44]

$${\text{SNR}}_{\text{AO}}=\frac{I_s}{\sqrt{2eFB\left(I_s+I_b+I_d\right)+{\sigma }_{\text{inst}}^2}},$$ (1)

where $I_s$ is the mean signal photocurrent, $I_b$ the mean background photocurrent, $I_d$ the dark current, $e$ the elementary charge, ${\sigma }_{\text{inst}}^2$ the variance of additive instrumental noise, and $B$ the measurement bandwidth. The first term in the denominator, $2eFB\left(I_s+I_b+I_d\right)$, arises from random photoelectron arrivals, which produce shot noise with a flat current power spectral density, i.e., $S_i\left(f\right)=2eFI$ with $I=I_s+I_b+I_d$. A crucial component of Eq. (1) is the PMT excess-noise factor $F\ge 1$, which models dynode gain fluctuations. The implication of this term is that AO-PMTs exhibit non-negligible excess (multiplication) noise: dynode gain fluctuations raise the shot-noise floor above the Poisson limit.

AO-PMTs are also susceptible to the relative-intensity noise (RIN) of the laser [45], [46]. The RIN, which is a distinctive characteristic of a given laser, increases the noise floor above the shot-noise limit, thereby degrading the SNR [47]. Because the RIN is larger at low frequencies, modulation transfer techniques, such as lock-in amplification (LIA) [48], help shift detection to a higher frequencies where $\text{RIN}(f)$ is smaller. This shift improves the SNR at low photon fluence. Given that they deliver analog signals, AO-PMTs interface naturally with LIA: by modulating the optical signal at $f_m$ and demodulating at the same frequency, detection is moved away from low-frequency noise and confined to a narrow bandwidth $B\approx 1/(4\tau )$ set by the LIA time constant $\tau$. Figure 1 shows the RIN of lasers typically used for label-free nonlinear microscopy. Note how in all instances, with the exception of the fiber oscillator, the RIN reaches a base line around 1MHz. This RIN floor is a good range for detecting the nonlinear signals.

Fig. 1.

Relative intensity noise (RIN) of various sources typically used in label-free nonlinear microscopy. The red curve depicts the RIN of the laser driving the experiments discussed in this contribution. Yb: ytterbium. OPO: optical parametric oscillator.

B. Photon Counting PMTs

Unlike analog readout, photon counting does not measure a continuous waveform. Instead, a discriminator registers discrete photon events (rates $S$). This means $I_k=e{S}_k$ for $k\in \{s,b,d\}$ [42]. In ideal counting, instrumental/electronic noise is typically negligible (${\sigma }_{\text{inst}}^2\to 0$), and $F\approx 1$ because the discriminator suppresses pulse-height fluctuations. Therefore, the SNR of a PC-PMT is

$${\text{SNR}}_{\text{PC}}=\frac{S_s}{\sqrt{2B\left(S_s+S_b+S_d\right)}}.$$ (2)

To verify whether off-the-shelf hardware obeys Eqs. (1) and (2), we co-registered the CARS signal from a reference sample of microplastic beads, recording the undemodulated and lock-in–demodulated outputs from an AO-PMT along with the edges from a PC-PMT. To ensure nearly identical power at the detector planes, a 50:50 beam splitter divided the signal. The AO-PMT photocurrent was passed to a transimpedance amplifier (TIA), whereby the resulting voltage was split into two branches: one sent to an analog input of a data-acquisition (DAQ) card, and the other to a LIA, whose demodulated output was routed to a second analog input. The edges from the PC-PMT were fed directly to the counter module of the same DAQ. With the analog and counter modules synchronized, we ensured that each readout from all channels corresponded to the same location in the object plane. Figure 2 summarizes the results of this test at moderate (Panel a) and ultralow (Panel b) photon fluences.

Fig. 2.

Coregistration of CARS with AO-PMT, AO-PMT+LIA, and PC-PMT. (a)–(b) CARS maps at 2950 cm⁻¹ of polystyrene beads imaged with pump powers of 1.2 mW and 6.5 mW at the sample, respectively. Scale bars: $10\mu \text{m}$. Insets below each panel show intensity profiles along the yellow line. The plots on the right report image SNR as a function of pixel dwell time. For AO+LIA, the LIA time constant $\tau$ was set so that the detection bandwidth satisfies the Nyquist criterion with respect to the pixel rate. The black curve (AO, raw) shows no improvement because the pixel binning factor was fixed at 1 (no averaging).

Invariably, the PC-PMT outperformed the AO-PMT in terms of image SNR, defined as ${\text{SNR}}_{\text{dB}}\equiv 20{\text{log}}_{10}\left(\frac{{\mu }_{\text{signal}}-{\mu }_{\text{bg}}}{{\sigma }_{\text{bg}}}\right)$, where ${\mu }_{\text{signal}}$ and ${\mu }_{\text{bg}}$ are the mean signal and background levels, and ${\sigma }_{\text{bg}}$ is the background standard deviation. At moderate photon fluence, AO + LIA and PC produced comparable image quality, but at low photon fluence, the PC-PMT outperformed both AO approaches, consistent with Eqs. (1)–(2). Figure 2(b), which shows an image acquired with an AO PMT, reveals an odd spatial-frequency pattern. This pattern arises not from the sample but from systematic electronic noise picked up in the detection chain. This observation ascertains the limited performance of the AO PMT detection approach relative to its counterpart using LIA and PC detection. Notably, for short pixel dwell times (fast scans), AO with lock-in detection approached PC performance because the LIA mitigates low-frequency RIN and instrumental noise.

C. Spectroscopic Capabilities of PMTs

Because a nonlinear microscope performs localized spectroscopy, we investigated the performance of AO and PC-PMTs as analytical tools. CARS lends itself to this experiment, for this signal scales as $I_{\text{CARS}}\propto {\left|{\chi }^{(3)}\right|}^2{I}_{\text{p}}^2{I}_{\text{s}}$, where $I_{\text{p}}$ and $I_{\text{s}}$ are the pump and Stokes intensities, and ${\chi }^{(3)}$ is the third-order nonlinear susceptibility, whose resonant component relates to analyte concentration [49]. With the Stokes intensity held fixed and in the resonant limit (negligible nonresonant background as in the CH stretching), CARS obeys power-law scaling in both pump power and concentration, i.e., $I_{\text{CARS}}\propto I_{\text{p}}^2$ and $I_{\text{CARS}}\propto C^2$. This means that, in either case, a slope of 2 should be observed on log–log plots.

To determine whether our instrumentation in fact captures the spectroscopic essence of CARS, we once again split the CARS signal with a 50:50 beam splitter and coregistered measurements from the AO-PMT, AO+LIA, and PC-PMT. For these experiments, we used glycerol, an oily solvent characterized by strong Raman signals in the CH stretch. We varied either the concentration or the pump power, as Fig. 3 left and right columns show, respectively. As expected, AO + LIA and PC delivered the lowest noise, but the noise statistics differed, with AO+LIA exhibiting Gaussian fluctuations and the PC-PMT following Poisson statistics. Surprisingly, in both set of experiments, only photon counting recovered the slopes that approached the expected value of 2 in the log–log fits, suggesting that in the low- to mid-fluence regime PC-PMTs provide the more reliable quantitative readout.

Fig. 3.

Coregistered CARS signals from glycerol measured simultaneously with AO (raw), AO + LIA, and PC readouts. The left column shows CARS intensities for glycerol dilutions, with pump and Stokes powers fixed at 6.5 mW and 15.3 mW, respectively. The right column shows CARS readouts as a function of pump power; in this series the Stokes power and glycerol concentration were held fixed at 15.3 mW and 6.8 mM, respectively. The insets show log–log fits and their corresponding slopes $b$.

D. Implications of Analog-Output and Photon-Counting PMT Readouts for Label-Free Tissue Imaging

The tissue microenvironment is a highly heterogeneous substrate. This heterogeneity is reflected in the spectroscopic features of the signals. Chief among these is variation in intensity. Such is the variability of intensity that it displays strong intra-organ dependence: a given channel readout may be either saturated, photon-starved, or completely dark over distances of only a few micrometers ($\mu \text{m}$). These intensity swings have practical consequences for detector choice. Ideally, the detection chain should accommodate order-of-magnitude intensity variations within the same field of view without modifying the experiment, e.g., attenuating the excitation or the signals. Achieving this feat depends on the dynamic range of the detector, meaning, the span between the smallest signal the detector can distinguish from noise and the largest signal it can measure without undergoing saturation.

Although AO-PMTs do not match the sensitivity and analytical reliability of their PC counterparts, they do offer one critical advantage for tissue imaging: flexibility. This flexibility stems from the adjustable gain, which tunes the dynamic range of the detector by controlling the anode current. By changing the gain it is possible to prevent saturation of bright structures, adapting the instrument to the large intensity variations characteristic of heterogeneous tissues. In practice, where a PC-PMT would require inserting attenuators prior to the photocathode, an AO-PMT could dim the photocurrent by lowering the gain with a knob or under computer control. At the expense of sensitivity, this trivial feature grants adaptability to the AO readout.

We illustrate this argument with co-registered CARS images of a freshly excised murine kidney acquired with AO-PMT+LIA and a PC-PMT. Specifically, we imaged arbitrary regions of a sagittally cut kidney at various AO-PMT gains (950, 750, 550 V; Fig. 4(a)–(c). We also imaged, with an AO-PMT gain set to 390 V (Fig. 4(d)), periadrenal tissue, targeting adipocytes, for these lipid-rich cells emit strong CARS signals.

Fig. 4.

Coregistered CARS images and intensity distributions of murine kidney acquired with AO-PMT+LIA and PC-PMT. Left column: split images (AO+LIA on the left, PC on the right). Right column: intensity normalized histograms. (a-c) Arbitrary regions of a sagittally sectioned kidney at AO-PMT HV settings of 950, 750, and 550, respectively. (d) Periadrenal adipose tissue at AO-PMT HV = 390. Scale bars: $100\mu m$.

While setting the AO-PMT gain unnecessarily high (Fig. 4(a)–(b)), specimen details are lost to brightness plateaus caused by saturation. Instead, setting the gain to a value appropriate for the sample brightness restores contrast (Fig. 4(c). Nevertheless, even when the time constant $\tau$ of the LIA satisfies the Nyquist sampling criterion relative to the pixel dwell time, AO+LIA images remain smoothed. This softening of the contrast derives from the last stage of the lock-in technique, a low-pass filter. Furthermore, as expected from Eqs. (1)–(2), the noise floor of the AO+LIA images is higher than those from the PC-PMT.

Equally informative are the intensity-normalized histograms [50] on the right of Fig. 4(a)–(c). The AO-PMT+LIA distributions (green) are continuous and approximately Gaussian, consistent with a linear analog chain whose noise is the sum of many independent contributions (shot, Johnson/dark, RIN), yielding near-Gaussian statistics. In contrast, the PC-PMT distributions (red) are Poissonian.

Finally, the flexibility of AO-PMTs is evidenced by Fig. 4(d). In a field of view where the PC-PMT was overwhelmed by photon events (in fact saturated, Fig. 4(d) right panel), the AO-PMT easily adjusted to the intense CARS of the fat cells. By recording both bright and dim structures, AO+PMT maintains operational flexibility, thereby avoiding saturation, which has negative implications for downstream data processing pipelines as it leads to the effective loss of information.

From these results we draw the following corollary for steady-state label-free nonlinear microscopy: photon counting may be favored when precise quantification or imaging of dim structures is required, while analog-output detection, coupled with a modulation transfer technique, may be preferred for tissues exhibiting substantial brightness swings. Table I summarizes the key differences between AO-PMT and PC-PMT readouts.

TABLE I.

Core Distinctions: AO-PMTs Vs. PC-PMTs. TIA: Transimpedance Amplifier. LIA: Lock-In Amplifier. ADC: Analog-to-Digital Converter

	AO	PC
Observable	Anode current	Pulses (rate)
Key noise (rms)	$\sqrt{2\mathit{\text{eFBI}}+{\sigma }_{\text{inst}}^2}$	$\sqrt{2B\left(S_s+S_b+S_d\right)}$
Dynode gain fluctuations	Present	Absent (ideal)
Instrumental noise	Present	Largely rejected
Best regime	High fluence	Low fluence
Typical hardware	PMT + TIA + LIA + ADC	PMT + Discriminator + Counter

III. The Multimodal Label-Free Imaging System

Three components comprise the backbone of our multimodal label-free system: (i) a laser source, (ii) an imaging platform, and (iii) a detection chain. Below, we describe each.

A. The Laser Source

Our design for the laser source takes advantage of one of the breakthroughs of modern laser technologies, namely, the solid-state laser based on ytterbium-doped (Yb-doped) crystals. The broad absorption and emission bandwidth of the Yb³⁺ ion enable the generation of fs-pulses, while its reduced thermal load, along with the perfect overlap between the emission band of InGaAs diode-lasers, results in a compact and efficient source [51]. In addition, the solid-state platform exhibits an outstandingly low RIN, thereby outperforming Yb-doped fiber lasers, see the red and olive curves in Fig. 1. Such low RIN levels augur superior performance compared with Yb-doped fiber lasers and support low-noise applications.

In particular, the source of our imaging system starts with a Yb-based solid-state laser (Flint, Light Conversion) operating at a 10 MHz repetition rate and delivering up to 5 W average power, whose output is split into two synchronized branches. One branch with approximately 1.5 W is loosely focused into a 10 mm-long MgO-doped periodically poled lithium niobate (MgO:PPLN; Covesion Ltd.) crystal, which is stabilized by an oven at 40°C to maintain the integrity of the crystal and the stability of the beam. Because LiNbO₃ is ferroelectric, periodic inversion of its ${\chi }^{(2)}$ domains enables quasi-phase matching [52], providing access to wavelengths otherwise unattainable with birefringence-based phase matching. In this way, we obtain emission centered at 1520 nm with up to 25% conversion efficiency (Fig. 5(b)), to produce the Stokes pulses.

Fig. 5.

The imaging system and spectra of its laser source. (a) Schematic of the label-free multimodal microscope. L, lens. F, filter. PPLN, MgO-doped periodically poled lithium niobate. MO, microscope objective. DM, dichroic mirror. G, grating. S, slit. ${\text{PMT}}_{AO}/{\text{PMT}}_{PC}$: analog-output/photon-counting PMT. (b) Left panel: spectra of the fundamental and narrowband pump. Right panel: Stokes spectrum.

The other branch of the fundamental (~250 mW) is sent to a spectral shaper implemented in a $4f$ configuration: a blazed grating (GR25–1210, 1200 grooves/mm; Thorlabs) disperses the broadband fundamental, an $f=120$ mm lens forms the Fourier plane, and a slit at that plane transmits a narrow spectral slice. The selected narrowband is retroreflected through the same optics, recombined, and then directed to the microscope. For clarity, the schematic in Fig. 5(a) depicts the effective optical path. Because the spectral resolution in CARS is set by the pump bandwidth, the narrowband output of the shaper serves as the pump (Fig. 5(b)). By narrowing the slit, the pump bandwidth can be reduced to achieve a spectral resolution of ~15 cm⁻¹. For the spectroscopic experiments, the slit was narrowed, but for imaging tissues we widened the slit slightly, trading spectral resolution for shorter pulses and stronger multiphoton signals. The pump and Stokes are combined with a short pass filter (FF01-1326/SP-25, Semrock), and guided to the imaging platform. The pump pulse duration immediately after the combiner was ~1330 fs with the slit closed and ~150 fs with the slit open, while the Stokes pulse duration was ~300 fs. The pump still generates SHG and MPAF, signals that can be resolved by considering the engineering criteria for detector choice discussed in Section II, a strategy that strengthens the contrast palette of our label-free nonlinear microscope. Thus, for tissue imaging the pump and the Stokes were set to 150 fs and 300 fs, respectively.

By coupling our solid-state Yb laser with the PPLN crystal, the system generates the pump–Stokes frequency detuning required to drive vibrational coherences in the C–H region, eliciting efficient CARS in the 2800–3000 cm⁻¹ range. Additionally, shifting the Stokes to the deep near-infrared, avoids electronic transitions and reduces the detrimental effects of the nonresonant background. Importantly, the Stokes beam alone did not produce noticeable signals that could leak into the detectors of the pump-driven multiphoton channels [53].

B. The Imaging Platform

After the combiner, the pump and Stokes beams are telescoped down and sent to a galvanometer scanner (GVS012, Thorlabs). A scan lens and a tube lens form an afocal relay that images the scanner to the back focal plane of the objective, ensuring telecentric scanning. An air objective (Nikon CFI60 Plan Apochromat Lambda D 20X, NA = 0.75) was chosen over oil or water immersion because mixed wavelengths (pump and Stokes) undergo significant chromatic aberrations with immersion optics. Additionally, water absorbs near ~1500 nm, close to the Stokes wavelength, and rapidly evaporates.

The nonlinear signals were collected in the epi direction and separated from the excitation with a dichroic mirror (FF875-Di01, Semrock). The CARS signal was then split from the remaining nonlinear signals using a dichroic beamsplitter (Di02-R785, Semrock), placing a collection lens and a bandpass filter (FF02–809/81–25, Semrock) prior to the CARS detector. A short-pass filter (cutoff ≈780 nm) passed SHG and MPAF. SHG was isolated with a dichroic (Di02-R532, Semrock) and a bandpass filter (FF01–509/22, Semrock). MPAF was selected with a bandpass filter (FF01–559/34, Semrock), targeting two-photon absorption fluorescence in the 540–580 nm range. Large-area scanning of the specimens was achieved using a motorized microscope stage (FTP-2000, ASI). Table II summarizes the targeted nonlinear signals and their corresponding spectral ranges.

TABLE II.

Target Nonlinear Signals and Corresponding Spectral Ranges

Target nonlinear signal	Spectral range (nm)
SHG	500–520
MPAF	540–580
CARS	765–850

Naturally, this platform has limitations. Chief among these is the lack of microscope objectives optimized for both the 1030 nm and 1520 nm beams, whereby chromatic aberrations reduce the phase matching of the CARS signal, effectively reducing the field of view achievable with the CARS contrast relative to the two-photon processes. Similarly, because microscope objectives are primarily optimized for transmitting the visible and near-infrared ranges (i.e., below 1100 nm), they attenuate the 1520 nm beam, negatively impacting the generation efficiency of the CARS contrast. While reflective objectives could ameliorate or even address these limitations, the restricted penetration depth at 1520 nm precludes volumetric imaging. Similar arguments apply when one desires to use visible light, in concert with the 1520 nm beam, to probe fluorophores with sub-500 nm excitation.

C. The Detection Chain

Because the laser source is optimized for driving vibrational coherences in the CH stretching, the CARS signal can be very strong but also very weak across a given field of view. Taking into account this large variation in brightness, we used an AO-PMT (H16722–50, Hamamatsu) coupled with a TIA (TIA60, Thorlabs), and LIA (Moku Pro, Liquid Instrument). For the CARS coregistration experiments discussed in Section II, this same AO-PMT was used and synchronized with its photon counting head homologue (H16721–50, Hamamatsu).

The SHG showed a rather similar trend as the CARS signal: some fields of view were SHG-rich while others SHG-starved. Therefore, the SHG channel relied on another AO-PMT (H16722–40, Hamamatsu) coupled to a TIA (TIA60, Thorlabs), and connected to the LIA, a multichannel instrument that could demodulate both CARS and SHG simultaneously.

The criterion for the detection choice of the MPAF channel was different from that of the coherent signals. While MPAF was very strong in some compartments (e.g., liver, kidney, blood), it produced low photon-fluences in most cases, a finding that is consistent with the fact that the pump pulses are relatively long. For this reason, the MPAF channel relied on a PC-PMT (H16721–40, Hamamatsu), for a photon-counting approach maximizes sensitivity.

Note that our system (Fig 5(a)) does not include a modulator. This is not an omission. Contrary to heterodyne detection of conventional pump-probe techniques [54], [55], [56], [57], [58], in which the signal sits on top of the driving fields as a very small intensity modulation, homodyne detection relies on the occupation of previously empty optical modes. This means that detection of frequency-shifted signals simply requires (optically) filtering the excitation and measuring the incoming radiation. Because our Yb-laser delivers a train of pulses, the excitation is intrinsically modulated, and the frequency-shifted fields inherit this modulation. Thus, by demodulating at the repetition rate of the laser, we implement the lock-in technique without an external modulator. The repetition rate of our laser was customized to 10 MHz, which lies in a region where its RIN is at its minimum, see red curve in Fig 1. Additionally, this low repetition rate yields high pulse energy and peak power for nonlinear excitation while still allowing fast scan rates and thermal relaxation between pulses. Therefore, our LIA takes its reference frequency directly from the synchronization output of the laser.

Finally, the detection chain was controlled by an in-house LabVIEW program that synchronized the readouts using two DAQ boards (PCI-6323, National Instruments).

IV. Simultaneous Acquisition of Multiphoton and Vibrational Contrasts: Preclinical Validation

The capabilities of the imaging system were evaluated over the course of two studies: (i) a naturally occurring metastatic carcinoma of unknown origin in a mouse and (ii) an induced mammary gland carcinoma in a rat. In both cases, SHG, MPAF, and CARS signals were acquired simultaneously from freshly excised tissues and coregistered using the strategy and system described in Sections II and III. After label-free imaging, specimens were processed for histology. Below, we detail the materials and methods, followed by the imaging results.

A. Ethics Statement

All animal procedures were approved by the Institutional Animal Care and Use Committee (IACUC) at the University of Illinois Urbana–Champaign (protocol IDs: #24063, #24014, #25042. This study adhered to the NIH Guide for the Care and Use of Laboratory Animals. Animal welfare was prioritized by minimizing pain and distress and by following the principles of replacement, reduction, and refinement (3Rs). Importantly, all animals used in this work were part of other IACUC-approved projects and no additional animals were euthanized specifically for this study.

B. Mice

Upon necropsy, a naturally occurring solid tumor mass was detected in the lower abdomen of a 1-year, 22-day-old retired breeder 5xFAD strain female mouse (MMRRC strain #034840-JAX), see (Fig. S1). As a non-carrier for the familial Alzheimer’s disease Tg(APPSwFlLon,PSEN1*M146L*L286 V)6799 V (also known as Tg6799) mutation, this mouse had been paired with a 5xFAD male positive for the Tg6799. Liver, lung, and kidney from both mice were harvested for gross macroscopic and histological evaluation following ex vivo imaging.

C. Allogeneic Orthotopic Mammary Tumor Model (rat)

Passage 2 (P2) of semi-adherent 13762 MAT B III rat mammary adenocarcinoma cells (ATCC, #CRL-1666), originally derived from a Fischer 344 rat mammary tumor, were maintained in McCoy’s 5 A medium containing L-glutamine (2 mM) and D-glucose (17.5 mM), supplemented with 10% fetal bovine serum (FBS; Cytiva, #SH30396.03) and 1% Antibiotic–Antimycotic solution (Cytiva, #SV30079.01), until cultures reached ~90% confluence in T-75 flasks (Corning). Following mechanical detachment, cell viability and concentration were measured on a Vi-CELL XR Cell Viability Analyzer (Beckman Coulter), per the manufacturer’s instructions. Approximately 1 × 10⁶ cells (viability ≈ 96.5%) were pelleted at $400\times g$ for 5 min at 4 °C, and the pellet was resuspended in $200\mu \text{L}$ of sterile PBS (pH 7.0). The cell suspension was injected intraductally into the right fourth (abdominal) mammary pad of a female retired breeder Sprague–Dawley rat (Charles River Laboratories; body mass 335 g; fed ad libitum). The contralateral pad received PBS alone (control). Tumor burden was estimated as $V=0.52\times L\times W^2$, where $V$ is tumor volume (mm³), $L$ is length (mm), and $W$ is width (mm) (both measured with a digital caliper). On day 9 post–cell inoculation, control skin and tumor-bearing skin were harvested.

1). Ex Vivo Tissue Imaging:

Prior to same-day imaging, freshly excised tissues were submerged in phenol red–free DMEM supplemented with 10% FBS and 25 mM HEPES and held at 4 °C. To flatten the margins, approximately half of the rat mammary tumor was dissected free from the skin bed to remove the dome and provide a planar surface. For imaging, tissues were transferred to glass-bottom dishes, oriented with the surface of interest facing the objective, kept hydrated with the same medium, and gently stabilized with a coverslip spacer to avoid compression. Label-free multimodal datasets were acquired as follows: CARS (red), SHG (green), and MPAF (yellow). Although the criterion for the selection of the color map is rather arbitrary, we used colors corresponding to what an observer would experience when looking at light centered at 515 nm (green), 560 nm (yellow), and 800 nm (dark red). Acquisition parameters (field of view: $100\mu \text{m}\times 90\mu \text{m},360\times 400$ pixels, pixel dwell $\text{time}=20\mu \text{s}$, and detector gains of 500V) were kept constant within each specimen. No exogenous labels or stains were used. Representative fields and zoomed regions of interest were recorded for each condition (control, peritumoral, tumor).

D. Histological Tissue Analysis

The mouse tumor was bisected to obtain a cross-section. The rat tumor was cut along its long axis to obtain a longitudinal section. Tissues were fixed in 10% PBS-buffered formalin for 72 h before processing. All specimens were embedded with the surface of interest facing downward. Paraffin-embedded blocks were sectioned at $5–7\mu \text{m}$ on a rotary microtome. To visualize nuclei, cytoplasm, lipids, red blood cells, muscle, and collagen fibers, deparaffinized sections were stained either with H&E (Harris hematoxylin; Electron Microscopy Sciences, #26041–05; and Reserve Multichrome eosin; StatLab, #SL201) or with Masson’s Trichrome or PicroSirius Red stains according to standardized protocols. Slides were digitized on a NanoZoomer 2.0-RS scanner and viewed with NDP.View v2.7.25 (Hamamatsu). Polarized light images were captured using Zeiss Axio Observer microscope equipped with 5x objective, circular and linear polarizers and ZEN v.2.6 (blue edition) software.

E. Case i: Metastatic Carcinoma in a Retired Breeder Female Mouse

Upon sacrificing the animals, kidneys, livers, and the tumor, in the case of the female, were harvested and prepared for imaging.

We first examined the tumor of the female rodent, comparing the label-free images with Masson’s Trichrome staining, a histologic technique that colors nuclei black, cytoplasm red, and collagen blue. In this tumor, the SHG channel revealed a fine mesh in almost every field of view. This structure may originate from tumor-derived type I collagen organized into thin, mechanically stiff fibrils that can promote cancer cell migration in solid tumors [59], [60]. PicroSirius Red staining shows the collagen fibers organized into thin type III or thicker mechanically stiff type I fibrils (Fig. S3). These structures were also observed across the entire tumor section on Masson’s Trichrome staining. Fig. 6(a) shows a whole-slide Masson’s Trichrome image with color-coded regions of interest, while Fig. 6(b) shows high-resolution images of those areas (top row) along with the corresponding high-resolution multimodal nonlinear images (bottom row). Because the CARS channel covers the CH₃ band (~2930 cm⁻¹), the contrast likely arises from protein-dense cytoplasm, consistent with the reddish staining of cytoplasmic regions on Masson’s Trichrome. In the multimodal images, the MPAF channel highlights cell-sized structures, potentially from FAD within tumor cells.

Fig. 6.

Abdominopelvic solid tumor of uncertain primary origin (intestine vs cervix) of a wild-type female mouse. (a) Whole-mount Masson’s Trichrome image with regions of interest marked with color-coded rectangles. (b) The top row shows high-resolution images of the regions of interest in (a), while the bottom row depicts multimodal images of related structures of the intact tumor (CARS = red, SHG = green, MPAF = yellow). Scale bars: (a) $1000\mu \text{m}$; (b) $200\mu \text{m}$.

Although possibly afflicted by neurodegeneration, the male mouse was a healthy specimen. Since it was contemporaneous with the female mouse and lived under the same conditions, it served as a perfect control. The top row in Fig. 7 shows hematoxylin and eosin (H&E) staining and label-free multimodal images of the kidney and liver of the male mouse. The kidney images exquisitely resolve tubular lumens and glomeruli, as expected in the renal cortex, while the liver images show hepatocytes (red in the multimodal images) and tiny intense (yellow) puncta in the nonlinear images, a contrast that may emerge from lipopigments.

Fig. 7.

Cross-validation of our label-free multimodal imaging system with H&E. Top and middle rows show the kidney and liver from a male and a female mouse, respectively. The left panels show the H&E image, whereas the right panels show the corresponding multimodal image (CARS = red, SHG = green, MPAF = yellow). The bottom panel shows a whole-slide H&E image of the female kidney, highlighting a nodule and perinodular stroma, with the adjacent multimodal image from the boxed region delineating a margin. All scale bars are $200\mu \text{m}$, except the whole-slide image in the bottom row, which is $1000\mu \text{m}$.

Conversely, both the kidney and liver of the female mouse exhibit radical architectural and chemical remodeling (middle row of Fig. 7). In the kidney, the H&E image reveals disruption of cortical organization, along with irregular tubular profiles and altered glomeruli. Similarly, the MPAF channel (yellow) of the multimodal image depicts abnormal tubular profiles covered with round, high-intensity structures which, owing to the sensitivity of the CARS channel to CH₂ vibrations (red), are likely lipid droplets. In the liver, H&E demonstrates sheets of tumor cells (little dark spots) replacing hepatocytes, while the multimodal image reveals intensified CARS with an SHG-enhanced boundary. Both the multimodal images and the H&E sections indicate morphofunctional remodeling in the viscera of the female mouse.

Finally, the multimodal images, particularly the MPAF channel, locate a margin, with adjacent sites showing radically different microenvironments (see the bottom panel of Fig. 7). A whole-slide H&E image not only confirms the structural abnormalities in the kidney of the female mouse but also locates a large nodule with altered morphology, supporting the findings of our label-free multimodal system.

F. Case ii: Imaging the Pathogenesis of Cancer in an Allogeneic Orthotopic Breast Cancer Animal Model

To assess the performance of our system in a more defined oncologic context, a mammary tumor was induced in a Sprague–Dawley rat by inoculating breast adenocarcinoma cells into the left abdominal mammary gland. Although this cell line originated from a different rat strain, it proved remarkably aggressive, producing sizable tumors within a few days with 80% take even in the presence of an intact immune system. As a testament to this aggressiveness, on day 9 post-inoculation the rat had a tumor measuring 11.7 mm × 17.9 mm, located in the left abdominal mammary gland (Fig. S2). At this stage of relatively low clinical tumor burden (V =1275.6 mm³), the animal was euthanized to capture early chemical and structural transformations induced by the malignant cells. Nonlinear imaging was performed consecutively on the freshly excised control site (right abdominal mammary gland), left abdominal mammary gland bearing the primary tumor, the overlying skin, and primary tumor mass.

We started by imaging normal (PBS-treated) control tissue to establish a baseline, see Fig. 8(a)). Thus, we imaged the right abdominal mammary gland, for this tissue was far from the tumor-bearing site. Although the H&E image located the adipocyte ghosts, the CARS contrast mapped these polyhedral fat cells, thereby revealing their packing, arrangement, and even texture in their native microenvironment. The SHG highlighted the collagenous connective stroma, whereas the MPAF channel delineated structures, potentially elastin-rich stroma or vascular fibers, with additional punctate autofluorescence, which may derive from hemoglobin or FAD-rich cells. The nonlinear image also exposed a duct, with a conspicuous lack of signals, running around the adipocytes. This profile is consistent with a mammary duct, its central lumen (the hollow interior) appearing dark. In this specimen, both the H&E and label-free multimodal images revealed a rather organized substrate.

Fig. 8.

Allogeneic orthotopic mammary tumor in rat. Left column: H&E images of (a) control (right mammary gland), (b) peritumoral bed (underlying tissue beneath the lesion), and (c) tumor. Middle column: higher-magnification views of regions of interest marked in the left column. Right column: corresponding label-free multimodal images of freshly excised specimens. Bottom-row insets correspond to regions of interest marked in panel (c) – left H&E and right CARS. Note the circular corpuscles, which are consistent with cells exhibiting a high nuclear-to-cytoplasmic ratio. Scale bars: first column, $500\mu \text{m}$; second and third columns, $150\mu \text{m}$; bottom row, $25\mu \text{m}$.

Next, we imaged the tissue underneath the tumor, aiming for a margin. Because the tumor extruded from its origin, exhibiting exophytic growth, delineating the margin was challenging. Regardless, the microscopic landscapes exposed by both H&E and the nonlinear images (Fig. 8(b)) were already different from the control (Fig. 8(a)). In the nonlinear images, the SHG channel offers a most notable feature, namely, an increased alignment and bundling of SHG-positive fibrils, reminiscent of collagen-rich (desmoplastic) stroma at the tumor interface of breast cancers [59], [60]. Across the imaged fields of view, the CARS channel showed a reduced adipocyte population, with the residual fat cells presenting a rounder shape instead of their characteristic sharp polyhedral geometry. Additionally, the multimodal images revealed an elevated MPAF signal, whose origin may be attributed to either epithelial/stromal species or blood vessels. Furthermore, no ductal lumina was observed, hinting to architectural transformation. Similarly, the H&E images also reveal abnormal features. Chief among these are rounder adipocytes; a higher density of microvessels; and an explosion of dark-purple round cells with a high nuclear-to-cytoplasmic ratio.

As a final dataset, we imaged the tumor (Fig. 8(c)). H&E shows a densely cellular and architecturally uniformity, revealing that the tumor is mainly composed of small cells with a high nuclear-to-cytoplasmic ratio, features that are consistent with high-grade, poorly differentiated mammary carcinoma. Instead, the label-free multimodal images resolve additional features. In particular, the SHG contrast reveals a thin collagen mesh commonly observed in solid tumors [59], [60] (compare with Fig. 6), while the MPAF channel highlights a microvascular structure. CARS demonstrates near-complete depletion of adipocytes and, at higher magnification, it reveals small cellular structures, which may be cancer cells, a fact that is substantiated by histology.

V. Conclusion

To conclude, our results highlight the chemical and structural heterogeneity of the tissue microenvironment, reiterating the need for simultaneous acquisition of both vibrational and multiphoton contrasts. With such a versatile, complementary contrast palette, a multimodal microscope can provide a thorough evaluation of tissue, revealing both chemical composition and architecture. Because chemical changes often precede morphological alterations in cancer pathogenesis [61], an imaging system of this kind has strong potential to make an impact in intraoperative consultations and image-guided surgery.

Our preclinical studies also reinforce the suitability of label-free nonlinear microscopy for clinical workflows. Although we do not aim or suggest replacing informative histologic techniques, a multimodal label-free nonlinear microscope expedites acquisition, reducing labor and potential processing artifacts inherent to histology. In particular, a label-free multimodal imaging system enables imaging of fresh, unprocessed tissue within minutes, thereby addressing the Achilles’ heel of histopathology: time-to-result with minimal manual effort. Furthermore, as our data show, the techniques are complementary: owing to the non-destructive nature of label-free nonlinear microscopy, specimens are left intact for downstream traditional histopathology.

Finally, the spatial brightness swings observed across biological specimens reiterate the necessity of tailoring the imaging system to the target tissue, including both the excitation source and the detection chain, the former to elicit the nonlinear signals, the latter to avoid saturation while maintaining sensitivity. Thus, building on a thorough investigation of detectors and readout schemes, we provided engineering considerations for detector choice in label-free nonlinear microscopy. In addition, we discussed a refined laser source that enables the full set of nonlinear contrasts while maintaining a compact, cost-effective platform that facilitates clinical translation. Notably, this source costs about one quarter as much as current commercial solutions and occupies roughly half their footprint. Therefore, the aforementioned strategy provides a blueprint for designing and constructing imaging platforms well suited for clinical deployment and ready to advance digital histopathology via optical biopsy.

Supplementary Material

This article has supplementary material provided by the authors and color versions of one or more figures available at https://doi.org/10.1109/JSTQE.2025.3650148.

Biographies

Alejandro De la Cadena was born in Xalapa, Veracruz, Mexico, in 1987. He received the B.S. degree in electronics and instrumentation engineering from Veracruz University, Xalapa, Mexico, in 2010, the M.Sc. degree in electronics from the National Polytechnic Institute, Mexico City, Mexico, in 2012, and the Ph.D. (with honors, Magna Cum Laude) degree from Friedrich Schiller University Jena, Jena, Germany, in 2018. He was a Postdoctoral Researcher, advancing broadband lasers and multimodal nonlinear microscopy with Politecnico di Milano, Italy. In 2022, he joined Prof. Stephen Boppart’s Lab at the University of Illinois at Urbana-Champaign to apply optical imaging to disease research. In the Beckman Institute for Advanced Science and Technology, he currently leads the Imaging Tech Group, a multidisciplinary team focused on developing biophotonics tools to study cancer and neurodegeneration. In 2023, he received the Marie Skłódowska-Curie Global Postdoctoral Fellowship, was named a Beckman–Cancer Center Postdoctoral Fellow, and was awarded the Mistletoe Research Fellowship.

Edita Aksamitiene received the B.S. degree in biology, the M.S degree in molecular biology and biotechnology, and the Dr. Sc. degree in biochemistry from the Faculty of Natural Sciences, Vytautas Magnus University (VDU), Kaunas, Lithuania, in 2001, 2003, and 2007, respectively. She was a Postdoctoral Researcher of cancer signal transduction as a NIH R01 and NIH NIAAA T32 grant-funded Trainee with the Department of Pathology, Anatomy and Cell Biology, Thomas Jefferson University (TJU), Philadelphia, PA, USA. In 2014, she joined the Department of Otolaryngology – Head and Neck Surgery, TJU, as a Research Instructor. Since 2019, she was a Research Scientist with the Beckman Institute for Advanced Science and Technology, which is a unit of University of Illinois Urbana-Champaign, Urbana, IL, USA, dedicated to interdisciplinary research. Her research interests include method validation and optimization strategies to achieve repeatable and reproducible cell signaling, wound healing-, and nanoparticle-related studies. Dr. Aksamitiene is a long-term member of Wound Healing Society, American Society for Biochemistry and Molecular Biology, American Society for Investigative Pathology, and the International Society for Extracellular Vesicles.

Stephen A. Boppart (Fellow, IEEE) received the B.S. degree in electrical and bioengineering and the M.S. degree in electrical engineering from the University of Illinois Urbana-Champaign, Urbana, IL, USA, in 1990 and 1991, respectively, the Ph.D. degree in medical and electrical engineering from the Massachusetts Institute of Technology, Cambridge, MA, USA, in 1998, and the M.D. degree from Harvard Medical School, Boston, MA, USA, in 2000. Since 2000, he has been a Faculty Member with the University of Illinois Urbana-Champaign. He is currently a Grainger Distinguished Chair of engineering as a Professor with the Departments of Electrical and Computer Engineering, and Bioengineering. He is the Director of the NIH/NIBIB P41 Center for Label-free Imaging and Multiscale Biophotonics (CLIMB), GSK Center for Optical Molecular Imaging, and the Interdisciplinary Health Sciences Institute. He has authored more than 450 invited and contributed publications, and more than 1000 invited and contributed presentations. He holds more than 50 patents. His research interests include the development of novel optical imaging technologies for translation to medical applications, and for fundamental biological discovery. He is a fellow of AAAS, OSA, SPIE, AIMBE, IAMBE, and BMES, and a member of the National Academy of Inventors. He was named one of the top 100 innovators in the world by the MIT’s Technology Review Magazine for his research in medical technology. He was the recipient of the IEEE Engineering in Medicine and Biology Society Early Career Achievement Award and the IEEE Technical Achievement Award.