Far-field super-resolution imaging has emerged as a powerful tool in biology to unravel the details of the complex molecular machinery at play at the nanoscale. However, the great majority of super-resolution techniques are based on exogenous markers (fluorophores) that demand chemical preparation protocols and further studies to determine cell viability and specificity to a targeted molecule. Most importantly, fluorescence-based tools only report on the fluorophore information—dynamical or structural—leaving open many fundamental questions on the other outnumbered unlabeled molecular species, e.g., lipids and cholesterol molecular conformation and local composition^[1–3] within lipid domains that have remained undetected in real cells, or the local composition of the species forming membrane-less organelles that are currently unknown^[4,5]. Therefore, Raman microscopies have emerged as ideal tools for probing heterogeneous biological specimens^[6] since they provide chemically resolved images using the intrinsic vibrational properties of molecules, that is, a label-free method. However, reaching fast super-resolution capabilities for opaque tissue imaging with Raman contrasts has remained challenging^[7].

In the last decade, many attempts have been made to enable vibrational far-field super-resolution in stimulated Raman scattering (SRS). Computational super-resolution methods, exploiting structured illumination microscopy (SIM), have been demonstrated for the spontaneous Raman contrast^[8]. However, the usage of an imaging spectrometer is not compatible with thick tissues, as the resolution enhancement is only provided in one dimension and the acquisition speeds are too slow for dynamic specimens. Very recently, a solution to this issue has been put forward, but the imaging methodology is still based on a wide-field geometry, potentially challenging to apply in opaque specimens^[9]. Alternatively, coherent Raman microscopies (CRMs) could provide fast acquisition speeds: with the two most known contrast mechanisms being coherent anti-Stokes Raman scattering (CARS)^[10] and SRS^[11–13]. However, there are various drawbacks that preclude biological specimen super-resolution imaging, in particular methods exploiting coherent control of vibrational dynamics^[14–25]. Furthermore, in CARS, interference artifacts complicate chemical quantification analysis^[26,27]. In the case of the background-free SRS process, the current mainstream is to exploit methods to control the dynamics of vibrational energy levels, however, using unconventional power levels that may be phototoxic for biological specimens^{[20,21,23,24,28–31]}.

While combining computational super-resolution methods with SRS technology could overcome the above-mentioned issues, it cannot provide super-resolution capabilities in thick opaque tissues. Generally, the mathematical framework of computational methods is based on wide-field illumination, which itself requires multi-pixel cameras. Unfortunately, wide-field cameras for SRS are technologically challenging because of the unconventional detection scheme of SRS: SRS requires a high-sensitivity radio-frequency lock-in amplifier (RF-LIA), which currently is only reliable in a single-pixel scheme. Despite recent developments of multi-pixel RF-LIA^[32,33], the pixel counts do not scale favorably for the architecture of thousands of pixels needed in a camera. Furthermore, wide-field illumination is not suitable for thick tissue imaging due to the lack of sectioning capabilities: even if a camera did exist for using computational methods with SRS, it would be challenging to use due to aberrations and out-of-focus light generated by the solvent or the sample itself.

To address these challenges, we present a camera-less (aka single-pixel) chemically selective super-resolution imaging methodology compatible with opaque thick biological specimens. We are interested in demonstrating that this methodology indeed breaks the diffraction limit resolution barrier of SRS microscopy. We start by describing the mathematical model used in the acquisition step, followed by a proof-of-principle aimed at demonstrating resolution beyond the theoretical diffraction limit. We then finalize the demonstration, showing that its sensitivity is compatible with imaging biological specimens. Remarkably, the framework presented here has a simple alignment procedure: it is simpler than conventional SRS microscopy, which demands the overlap of two tightly focused beams.

Specifically, we developed a single-pixel scheme (Fig. 1) compatible with computational super-resolution methods, therefore allowing for fast imaging capabilities exploiting SRS processes in the form of stimulated Raman gain (SRG) [Fig. 1(a)]. In our arrangement, a structured stationary pump beam is shaped using a spatial light modulator (SLM) and is spatially and temporally overlapped with a focused Stokes beam that scans over the sample [using a set of galvanometric mirrors, Fig. 1(b)]. After acquiring a series of SRS images with multiple structured illuminations, the data are treated with algorithms based on standard SIM mathematical framework to recover a super-resolved image^[34,35], as described below. We coin the method single-pixel blind-SIM SRS (or blind-S3 for short).

We describe the forward model of the acquisition procedure. In SRS, the detected signal (ΔIS), a modulation transfer between the pump and Stokes beams, at one pixel location (of the un-processed image) is given by^[12]ΔIS∝ℑ{χ(3)}IPIS,(1)where ℑ{χ(3)} is the imaginary part of the complex-valued nonlinear susceptibility of the sample (related to the Raman cross-section), and IP and IS are the intensities of the pump and Stokes beams, respectively. In the case of the proposed acquisition method in blind-S3, a static speckle pattern generated by the pump beam spreads at the sample image plane where the Stokes beam is focused and scanned. To derive an image formation model, we assume a scalar approximation for the local intensity in one blind-S3 image, $$\begin{gathered} ΔIS(x,y)∝∬ℑ{χ(3)(x′,y′)}IP(x′,y′)IS(x−x′,y−y′)dx′dy′ \\ ∝(ℑ{χ(3)}IP)⊛IS. \end{gathered}$$(2)

An image acquired with the blind-S3 scheme obeys a forward model of the type Mi=(O×ISIM)⊛IStokes, where Mi is an SRS image from a single speckle realization, O is the optical response of the excited object (more precisely, ℑ{χ(3)}), ISIM is the spatial distribution of the structured intensity at the pump wavelength, IStokes is the effective point-spread function (PSF) of the image formation system, and ⊛ denotes a convolution operation. This means that each single image acquired follows the standard forward models in computational super-resolution frameworks of incoherent processes. We chose to work with non-sinusoidal SIM patterns in order to be compatible with thick tissues: we use speckle patterns since they are resilient in scattering specimens. While we tested the methodology with two SIM algorithms using no prior knowledge of the structured patterns ISIM^[34,35], for the results presented we used the one described in Ref. [35]. In this approximation, we disregard coherent effects as SRS processes are inherently phase-matched.

Briefly, the output power of a femtosecond laser source (Coherent, Chameleon Ultra Vision, 800 nm, 80 MHz repetition rate, 150 fs pulse length) pumps an optical parametric oscillator (APE, MIRA-OPO) that generates the Stokes beam, centered either at 1058 nm (Fig. 2, 3054cm−1 Raman-shift) or 1042 nm (Fig. 4, 2903cm−1 Raman-shift), and a small power fraction is used as the pump beam. The Stokes beam is spectrally narrowed using a combination of grating (LightSmyth, T-1000-1040) and adjustable slit width for the purpose of increasing chemical selectivity. The pump beam is also spectrally narrowed in a pulse-shaper setup using two gratings (LightSmyth Technologies, T-1400-800) and a digital micromirror device (DMD) placed in the Fourier plane (a description of the methods using DMDs for SRS spectroscopy can be found in Ref. [36]). The pump beam is amplitude-modulated at 1 MHz by an acousto-optic modulator (AA Opto-electronic, MT80-B30A1,5 VIS). The specimen is z-displaced using a piezo stage (Thorlabs, DRV517), and the signal generated by the sample is then collected by a 1.4 numerical aperture (NA) oil-immersion condenser, directed to a large-area detector (Thorlabs, DET100A2) and demodulated by a lock-in amplifier (Zurich Instruments, MFLI).

We used two configurations for SRS microscopy. Regardless of the configuration used, both beams are spatially and temporally combined at dichroic mirrors, whose location depends on the modality of SRS in use, and focused by an objective (Nikon, Plan APO IR, 60×, NA=1.27). To achieve the best compromise in terms of resolution enhancement (by having the pump beam as a structured illumination) and sensitivity (by having the Stokes beam as the demodulated beam), we have designed a layout that allows us to quickly swap the direction of the pump beam between the conventional SRS or blind-S3 configurations using a combination of a half-wave waveplate and a polarizing beam splitter cube. For the blind-S3 configuration, the pump beam is sent onto an SLM (Meadowlark Optics, HSP512L-1300) to modulate the wavefront with a random phase. The SLM throughput is higher than 80%, but it could be further enhanced by replacing the SLM by engineered diffusers since they have no absorption. Galvanometric mirror scanners are used to move either pump and Stokes beams together (conventional) or Stokes only (blind-S3). The typical average power measured before the objectives was 13 mW (conventional) and 41 mW (blind-S3) for the pump and 25 mW for the Stokes beams. However, we note that the energy density levels used for blind-S3 are inherently lower than the conventional SRS configuration: we have estimated 5 times lower effective energy densities (i.e., product of the energy densities of the pump and Stokes energy densities), taking into consideration the speckle envelope and the longer integration time in the blind-S3 procedure, when compared to the conventional method. Finally, after imaging in conventional SRS and blind-S3, we image a large field of view (FOV) to detect any sign of phototoxicity such as that shown in the wide FOV images of the biological specimens.

Samples presented in Fig. 2 were prepared by drop-casting the polystyrene beads on a coverslip and embedding them in deuterated water to decrease the spectral congestion with the water vibrational response background. The various diameters (and standard deviation) used were: 239 nm (6 nm, PS Research Particles), 372 nm (10 nm, Polysciences, Inc.), 520 nm (16 nm, Thermo Scientific), 740 nm (22 nm, Thermo Scientific), and 990 nm (30 nm, Polysciences, Inc.). Mice brain slices were kindly provided by Laurent Bourdieu, and experimental procedures were conducted in accordance with the institutional guidelines and in compliance with French and European laws and policies. All procedures were approved by the “Charles Darwin” Ethics Committee (project number 26667). More precisely, 6-month-old C57BL6 male mice were sacrificed, and the extracted brain was then stored overnight in a solution of 4% paraformaldehyde and finally rinsed in a phosphate buffer solution (PBS). Coronal slices with thicknesses of 100 µm were then cut and stored in the PBS. Prior to experiments, the slices were placed between the two coverslips with a 120-μm-thick spacer. HeLa cells (ATCC) were incubated with 400 µM oleic acid, washed, fixed with 4% paraformaldehyde, and stored at 4°C before imaging.

We assume that the theoretical transverse resolution results from the product of two focused Gaussian beams with two different wavelengths λp and λs for the pump and Stokes wavelengths, respectively. Here, we use the Raman resonance 3054cm−1 and the Rayleigh criteria to assess the resolution limit of each beam: ΔrP=1.22λP2NA=384nm and ΔrS=1.22λP2NA=508nm for the pump and Stokes beams, respectively, where NA=1.27, λP=800nm, and λS=1058nm. Therefore, the theoretical resolution limit is ΔrSRSconv=11ΔrP2+1ΔrS2=307nm for conventional SRS while it is ΔrSRSblindS3=ΔrSRSconv2=217nm for blind-S3.

We first demonstrate the improvement in the transverse resolution, surpassing the diffraction limit of usual SRS microscopy. In order to evaluate the gain in resolution, we compare blind-S3 to the conventional scanning methods. For conventional SRS, the theoretical transverse resolution is ΔrConv=307nm. This theoretical value is technically challenging to achieve with high NA objectives in the near-infrared (IR) because the wavelengths of the two beams differ by hundreds of nanometers (spectral span necessary for fast quantification of lipids, proteins, and nucleic acids in SRS microscopy), in opposition to the visible range where diffraction limited performance has been reported^[37]. Conversely, blind-S3 transverse spatial resolution results from the doubling in resolution dictated by SIM and the speckle grain size limited by diffraction, leading to ΔrblindS3=217nm. To show the superior transverse resolution in blind-S3, we imaged 239-nm-diameter polystyrene beads with the two modalities [Figs. 2(a) and 2(b)]. Clearly, conventional SRS [Fig. 2(a)] cannot resolve the beads transversely as the bead size is smaller than the theoretical resolution limit. After multiple speckle pattern illuminations, we feed the resulting images to a blind SIM algorithm to reconstruct a super-resolved image. Notably, blind-S3 methodology [Fig. 2(b)] resolves several beads in the in-focus layer. The line profiles [Fig. 2(c)] reveal the distance between the centers of the beads (242 nm), which matches well to the distance of close contact between two beads. We note that the effective region-of-interest (ROI) of blind-S3 is modulated by the speckle envelope, hence decreasing the similarity of the two images in the edges of the bead cluster, yet not affecting the resolution gain (see below). The present findings show that the blind-S3 methodology goes beyond the fundamental far-field diffraction limit resolution of SRS microscopy, by improving the resolution ≥2, without the addition of exogenous signal enhancers and using excitation energy densities lower than conventional SRS microscopy, therefore decreasing the probability of nonlinear phototoxic effects.

Remarkably, super-resolution in blind-S3 comes with intrinsic z-sectioning capabilities. In each illumination during the blind-S3 procedure, an image is formed based on a wide-field geometry model, that is, an object is convoluted with a linear PSF. In a hypothetical conventional wide-field SRS microscope using multi-pixel cameras, the excitation beams would overlap a large volume. This would in turn reduce the sensitivity due to out-of-focus shot noise, therefore deteriorating image quality and resolution due to the background noise (shot noise). Conversely, the blind-S3 methodology improves sectioning, without resorting to cameras, as the nonlinear optical response is local in the longitudinal direction: because SRS signals are only generated within the overlap region of the two beams, each SRS image does not contain appreciable out-of-focus shot noise. To demonstrate the sectioning capabilities of blind-S3, we probe a thin film of oil (of a few micrometers) by scanning it in the longitudinal direction. Note that we collect the signal generated for each z-position on a ≈10mm-wide detector, hence, not in a confocal geometry. Clearly, conventional SRS and blind-S3 give a peaked response, which means that those two techniques have inherent longitudinal sectioning [Fig. 2(d)]. Indeed, the conventional SRS microscope can show such z-sectioning capability due to its nonlinear longitudinal PSF.

Contrary to conventional methods in super-resolution microscopy, in blind-S3 it is not straightforward to compare the reconstructed images with a “ground truth” object. This arises from the fact that the FOV in blind-S3 is modulated by the speckle envelope, which is smaller than conventional SRS microscopy. Therefore, we devised another methodology to ensure that the reconstruction was indeed reaching super-resolution capabilities. We imaged commercially available calibrated polystyrene beads of various sizes, which are well-known to aggregate and form close-packed structures. Therefore, we can use the bead close contact distance as a proxy for the bead diameter. We measured the close contact distances of several beads for several sizes ranging from smaller than to several times the resolution limit, and also for two different objectives with different NAs. Although the method is somewhat subjective, we were careful to choose “spot centers” that had the smallest distances possible. Following this procedure, we noticed that the maximum spot-to-spot centers were indeed limited by the bead size, that is, in the 360-nm-bead diameter we did not see 240 nm spatial fluctuations. The outcome of this procedure is shown in Fig. 3, and the agreement between the nominal bead diameter and the retrieved diameter, therefore, confirms that the features observed in the blind-S3 reconstructions indeed correspond to physical features beyond the diffraction limit.

To demonstrate blind-S3 compatibility with biological specimens, we image standard cell lines and mouse brain tissues. Conventional SRS reveals several μm-large droplets within the cell in the FOV [Fig. 4(a)]. Close-up images show different cluster morphology [Fig. 4(b)] and increased resolution gain with blind-S3 from the line profiles of the selected ROI [Fig. 4(c)]. To demonstrate capabilities for aberrant and opaque tissues, we have further imaged highly scattering brain slices at 8-μm-deep in the sample [Figs. 4(d) and 4(e)] with line profiles demonstrating increased resolution power of the myelin structures [Fig. 4(f)]. The close-up images with super-resolution capabilities [Fig. 4(e)] reveal that the structure of the myelin in the tissue is actually not as symmetrically perfect as inferred from the low-resolution images. These results show that the method is compatible with thick tissue imaging, despite being completely opaque, a situation in which using a hypothetical SRS widefield camera approach may fail due to background shot noise.

We consider the effects of the idle speckle upon illumination of the sample. One potential hazardous effect is due to heating. While the Stokes is point scanning, there are other regions of the speckle pattern that constantly illuminate the sample, therefore potentially presenting a phototoxic effect. Here, we assume the peak power of the speckle is too weak to induce nonlinear photodamage (at least not observed in our experiments and shown by the large FOV images after the acquisition procedure), and we consider what the temperature rise is due to the small, but non-negligible, absorption of water at the pump wavelength. This is a safe assumption as the speckle power used in the experiments is spread over a region that is over 10× larger than the spot size of the conventional SRS imaging system: this would represent a 100× weaker peak power. To analyze the heat effects, we use a well-established heat propagation framework used in the context of optogenetics^[38], which has been shown to be accurate in previous experiments. The simulation was performed with a speckle illumination power of 1 W (20× higher than the experiments) and an envelope size of 10 µm. Since heat-diffusion is faster than image acquisition, the speckle grains are washed out, and only the speckle envelope is important to consider. Hence, we plot the maximum of this thermal envelope in Fig. 5 for an illumination that stopped at 10 s and a continuous one. One can see that the steady-state value saturates at 0.023KW−1MHz−1, which corresponds to temperature rise <0.1K in our experiments. Note that the recently reported photothermal SRS signal^[39] could be present and, therefore, have a higher temperature rise than the pump speckle itself. However, in the simulations above, we consider the pump speckle envelope, which does not go through the thermal SRS effect as a whole. Furthermore, this thermal SRS heat mechanism would be highly local and would quickly dissipate as the overlap of the pump and Stokes beams is small and as we use water (fast thermal dissipation). Finally, to reduce the temperature rise due to the idle speckle illuminating the sample, one could alleviate this by conjugating the excitation speckle plane with a scanning mirror (for instance, using a DMD).

Here, we have demonstrated above that the proposed method reaches a resolution beyond the diffraction limit. We further demonstrated that the low excitation power used ensures low phototoxicity. Nevertheless, we discuss below potential issues in the methodology and how to overcome them with further engineering.

Further technical improvements could greatly overcome current limitations in these proof-of-principle experiments. In regards to speed and/or increasing the FOV, the current implementation used classical frequency-domain spectral narrowing methods, which have lower throughput, therefore limiting the largest FOV available (i.e., a larger FOV requires more laser power to keep the energy density constant). To address this issue, we envision an improvement using spectral focusing methods, therefore increasing the FOV of the measurement as it can use all laser power available. Despite these proposed follow-up improvements, the blind-S3 technique can super-resolve in strongly opaque biological tissues: indeed, in biomedical applications, aberrations deteriorate the PSF of the microscope, perhaps explaining the lower resolution attained in the biological specimens observed here. Here, while the penetration depth of brain tissues was limited to 8 µm, with recent advances of deep SRS imaging^[40], this penetration depth could in principle be extended. Finally, we speculate that blind-S3 could be a route to reach far-field nanoscopy (sub-100-nm resolution). A straightforward route for nanoscopy is to decrease the excitation wavelengths in the ultraviolet (UV)^[37]. However, UV radiation is known to enhance phototoxicity (for instance, by generating DNA photoradicals^[41]). With blind-S3, one could still use excitation lasers in the visible range and reach resolutions below the 100 nm barrier because of the structured illumination approach.

In conclusion, we have designed and demonstrated a single-pixel super-resolution technique that is straightforward to implement, i.e., simpler than conventional SRS microscopy. We demonstrated that this technique can image beyond the diffraction limit (2), and that it did not show phototoxic effects in the imaged biological specimens. blind-S3 is a universal approach in the sense that it depends neither on specific vibrational mode ultrafast dynamics^[29], nor does it require vibrational signal enhancers^[42,43] or a priori knowledge about the specimen, for instance, as gained in the training of neural network methods^[44,45]. Therefore, these achievements have overcome a decade-long challenge in SRS imaging, paving the way for investigating matter in its most natural environments.

After the initial submission of the current work^[46], an alternative computational approach has been experimentally demonstrated for SRS super-resolution of biological specimens^[47] using deconvolution methods.