Thanks to the substantial reductions of tissue scatterings and absorption when fluorophores emit at a longer wavelength¹, fluorescence imaging through the second near-infrared window (NIR-II: 1000~1700 nm) has been widely developed, which includes both a series of NIR-II contrast agents^2-4 and a range of advanced imaging systems. The recent advances made in volumetric imaging systems can be compared in a pyramid of resolution, speed, depth, and volume of view. In 2010, Welsher K et al. proved that the penetration depth can be maximized when choosing the working wavelengths between 1 and 1.4 µm⁵. In 2012, Hong et al. developed single-walled carbon nanotubes for NIR-II fluorescent imaging to achieve a high spatial resolution of ~30 µm, fast acquisition of less than 200 ms, and satisfactory tissue penetration of up to 3 mm⁶. In 2014, the group made progress in through-skull fluorescence imaging of mice without craniotomy and demonstrated an imaging depth of more than 2 mm with a resolution better than 10 µm, a field of view of 25 mm × 20 mm, and a frame rate of 5.3 per second⁷. In 2019, Wenbin et al. developed a confocal microscopy system using NIR-II aggregation-induced emission dots, achieved an 800 µm-deep 3D in vivo cerebrovascular imaging of a mouse, and reached a spatial resolution of 8.78 µm at 700 µm depth⁸. Liu Y et al. reported the use of carbon quantum dots for NIR-II two-photon bioimaging with a maximum penetration depth of 500 µm⁹. In 2021, Wang et al. reported structured-illumination light-sheet microscopy with a penetration depth up to 750 µm and a spatial resolution of 1.9 µm × 1.9 µm × 1.9 µm¹⁰. With the four key performance specifications of the field of volume, imaging speed, depth resolution, and lateral resolution being quantitatively summarized in the schematic Fig. 1, the trade-off becomes obvious among the choices of three light illumination strategies, wide-field, spot scanning, and light-sheet scanning.

Light-sheet fluorescent microscopy, by illuminating a single optical section at a time, makes volumetric imaging practical with reasonable spatial and temporal resolutions¹⁴. The thin illumination layer minimizes the phototoxicity effect¹⁵, and z-scanning with an array detector reconstructs the 3D images. In 2021, Wang et al. combined light-sheet microscopy with a type of lanthanide-doped NIR-II fluorescent nanoparticles to reduce out-of-focus background and improve spatial resolution¹⁰.

There are several challenges towards extending the light sheet system for vasculature imaging at millimeter-scale in each dimension. First of all, there is a trade-off between the length and the thickness of light sheet ruled by the Gaussian beam distribution of laser. Moreover, the scattering effect prevents high-contrast imaging in millimeter-scale tissue, as the surrounding area can be also excited by the strongly scattered light through the thick tissue. The spectrum filters may not completely shield the laser scatterings and background. Consequently, the low contrast prevents the detection system from resolving vasculature's fine details.

Time-gated (TG) fluorescence imaging has attracted increasing attention^{16, 17}. It separates long-lifetime fluorescence by gating off the pulsed laser excitation. Theoretically, this method can completely remove the laser scattering light and autofluorescence toward high contrast¹⁶. In 2019, Gu et al. developed a new contrast agent of Yb³⁺ doped nanoparticles, which has near-unity quantum yield and sufficient long lifetime at NIR so that time-gated imaging can be used to achieve a high signal-to-noise ratio of >9 under an excitation power density of 1.1 mW/cm ²¹⁸. In 2020, B Nimmegeers et al. prepared NIR emitting nanoparticles for time-gated imaging to remove the autofluorescence from the biological tissues¹⁹. To implement the scheme of time gating in the visible range, we can use a modulated camera or a switchable single element detector^20-22. However, when working at the NIR-II, a time-gatable camera needs to be custom-made with high cost. Using a chopper to switch the light path^{16, 23, 24} is a low-cost and practical strategy.

Computational imaging, by developing algorithms for data analysis, can significantly improve the imaging quality. For example, a vessel enhancement algorithm can facilitate to identify vessel networks from the interference of background tissue, which is usually based on the segmentation of blood vessels by imaging gradient, intensity, and texture. Hessian matrix-based method calculates the partial derivative to segment the vessel according to the mutation gradient²⁵. Principal component analysis (PCA) based method extracts vessel signals by removing the background⁵. However, both methods can be affected by the background gradient. In this work, we removed the most influence from the background gradient and developed an algorithm based on deep learning to segment the vessel.

To generate a light sheet, we set the focal spot on the designed distance to the back focal plane of the objective lens to form a lensset with low Numerical Aperture (NA). We adjust the focal spot using an objective lens (10×, NA=0.3, Nexscope planF-Apo) and reshape the Gaussian beam by a light sheet using a Powell lens array (PMMA-Powell lens array-30). Powell lens array provides more uniform sheet illumination than cylindrical lens²⁶. The FWHM (full width at half maximum) of the light sheet can be calculated by:

12ω0=2nλπNA,

where λ is the wavelength,n is the refractive index. Using the Gaussian beam model, we get the diameter and length of the beam waist, the Gaussian beam functions are shown as:

2E=E0ω0ω(z)exp(−r2ω2(z)),

3ω(z)=2ω0,

4z0=ω2πηλ,

where E is the amplitude of the light beam, E0 is the amplitude at the focal plane, z is the distance from the focal plane,r is the distance from the optical axis, ω(z) is the spot radius,ω0 is the spot radius at the focal plane. z₀ is the Rayleigh length,η is the refractive index of the medium,λ is the wavelength. The Rayleigh range is the range when the spot area becomes twice the minimum value, and the Rayleigh length of a Gaussian beam is the length from the waist to the spot of the Rayleigh range.

The surface of the mice is not flat, and therefore sufficient Rayleigh length is required for light-sheet scanning. To achieve a sufficient length of light-sheet illumination, we chose the Rayleigh length as 8 mm, so that the simulated radius of the focal spot is calculated as 100 µm in diameter. In our experiment, we achieved the focal spot with a diameter of 100.5 µm when the Rayleigh length was about 8 mm, as shown in Fig. 2. Towards a deeper illumination with the Rayleigh length of up to 26 mm, the thickness of the light sheet can be adjusted to 183.1 µm when adjusting the distance between the lens set and the objective. The comparison results of wide-field imaging and light-sheet imaging (maximum projection) are shown in Fig. 2(c, d). The result indicates that light-sheet imaging provides lower background intensity. The light sheet result shows a better contrast ratio than wide-filed illumination because of less scattering. With light-sheet mode illumination, there is only scattering light in the light sheet area. However, large surrounding scattering light overlay together, leading to a high-intensity background.

Though NIR-II light at longer wavelength leads to less optical scattering interaction with the tissue, it still shows a certain degree of unavoid scatterings. The scattered light from the laser excites the surrounding fluorescent material so that the obtained image looks thicker (reduced resolution) than expected (Fig. 3(a)). In Fig. 3(b–d), we evaluate the depth-dependent scattering effect of our light-sheet scanning system and visualize it via the 3D reconstruction of a pattern. We made four silicone phantoms containing 1550 nm-emissive downshifting nanoparticles (DSNPs, see Methods for the nanoparticle synthesis and morphology) and TiO₂ powder, in which the former serves as the contrast agent and the latter as scattering media with varying concentrations from 0 to 0.8 mg/mL (Fig. 3(b)). We used a 980 nm light to illuminate the phantoms and imaged their side profile. As the TiO₂ concentration increases, the scattering effect became severe. In the real tissue medium, the degree of scattering effect varies randomly. The distance-dependent scattering effect and the tissue section of interest will become hard to be quantified in a conventional small animal imaging system using wide-field illumination. Therefore, the resolutions in both lateral and axial directions become unpredictable. By contrast, in our case here, the application of light-sheet illumination improves our ability in optical sectioning, so that we can precisely locate the real position of the tissue section in a much small and confined area, due to the Gaussian distribution of light-sheet beam across the tissue. As the center of the light sheet has maximum power density where the fluorescent intensity reaches the maximum value, we can use the maximum position as the center of the blood vessel during the 3D reconstruction process. Here we embedded some cotton materials of 500 µm diameter into a silicone phantom to form a pattern of letters SUSTech stained with DSNPs. The letters are distributed at different depths through the phantom matrix (Fig. 3(c, d)). The 3D image reconstruction result matches the real structure of the phantom matrix, suggesting the accurate localization ability in our large-FoV imaging.

Here in our system, the low scattering effect of light-sheet illumination leads to a high resolution in a deep depth. Moreover, the residue laser scattering can be removed by the time-gated detection by collecting the signal from the long lifetime DSNPs (2 ms, see Supplementary information Methods) after the pulsed laser excitation off, towards a high imaging contrast. As shown in Fig. 4(a), in the detection optical path of the system, we inserted a chopper in between the first lens with an 8 mm focus length and the second lens with a 25 mm focus length to realize the time-gating. The detection path only opens when the laser excitation is switched off so that any short-lived background becomes negligible. The chopper used here (C995, Terahertz Technologies) is designed to have a blade consisting of 8 slots with a duty cycle of 2.23:1, making sure the most excitation light is detected. Though the chopper blades could remove all the excitation light, it causes part of the emission light to be lost. To maximize the luminous flux in detection, we optimize the excitation time, detection time, and detection frequency, according to the fluorescence lifetime and integral calculation (see Supplementary information Methods). When operating at a frequency of 250 Hz, it gave a rotational speed of ~31.25 rev/s. A 6-mm diameter pinhole aperture was attached very close to the chopper blade at a radius of 4.2 cm so that an ON/OFF switching time of 1856 µs was achieved for the signal light with any stray light being removed. The chopper outputs a TTL signal, generated from the slotted optical switch built in the chopper head, to trigger a homemade pulse synchronizer. The latter delivered pulses of 1858 µs duration to the laser controller/driver to switch on the 980 nm laser when the detection path was blocked by the chopper blade so that a SWIR camera became effectively time-gated.

Figure 4(b) shows the time-gated (TG) imaging result in comparison with the continuous wave (CW) mode with average laser power from 54 mW to 486 mW. The high peak power of TG mode enables it to excite a brighter fluorescent signal than CW mode illumination with the same average laser power. Thus, TG mode imaging provides higher imaging contrast. To quality the imaging contrast, we calculated the intensity contrast using the following function:

5Icontrast=(Imax−Imin)/Imax,

where the I_contrast is the index of intensity contrast, I_max is the mean intensity of the fluorescent area, I_min is the mean intensity of the background area. We compare the I_contrast and intensity range of TG mode and CW mode with average laser power from 54 mW to 486 mW. Furthermore, we study the photothermal effect of these two modes. TG mode shows a lower photothermal effect than CW mode and the comparison results are shown in the supplementary information.

To obtain the whole-body image of the vessel structure in tissue volume, we move the stage during light-sheet scanning, 100 μm a step, after capturing the image with an exposure time of 500 ms. It takes about 6.25 minutes to complete 750 steps for the whole mouse scanning. We then use Imaris software to reconstruct the 3D image.

A conventional method to enhance the vessel structure is to calculate the partial derivative of the image by Hessian matrix²⁵. However, when the vessels are presented against a complex background, the resolving power of Hessian matrix based method is limited. To enhance the vessel contrast and resolve the small vessels, we first make a compensation against the background gradient by rotating the profile to a horizontal angle and then mark the vessel by filtering pixels with the shape of uplifting. Although this approach can well judge the small blood vessels under a low contrast, it may misjudge the noise and the calculation speed is slow. To solve this problem, we introduce a deep learning model of DenseNet²⁷. We modify the results of the above algorithm by denoising, and training the DenseNet model with raw images and the corresponding results. Finally, we obtain a depth neural network model which can improve the contrast between the blood vessel and background tissue. The 3D reconstruction result and its vessel enhancement result are shown in Fig. 5. The pixels in the white block are used to calculate the mean intensity of the background. There is a very weak vessel signal pointed out by the white arrow. This algorithm enables us to obtain the vessel structure from images with I_contrast down to 0.07 (−0.012+0.007)I_contrast, and to reveal more vessels when their intensities are close to the background. In a 55 mm × 75 mm × 20 mm volume, the finest size of the vessels was around 100 µm. Using light-sheet center location and vessel enhancement algorithm, we provide more depth information with more details about the vessel network than the traditional small animal imaging system. In Fig. 6, we reconstruct the 3D model of mice vessel. By measuring each length of the vessels, we conclude that the total length of the reconstructed vessel is 527.7 mm.

Moreover, to resolve the depth information of blood vessels from the mice skin, we designed a scattering-based modality to acquire the 3D surface profile of the mice skin. The height information of the mice vessel is shown in Fig. 6(c), and the mice skin is shown in Fig. 6(d). In our experiment, we used a 1550 nm laser to form a light sheet, as this wavelength is consistent with the emission wavelength of DSNPs to avoid chromatic error. Thus, the fluorescent imaging mode is to map the vessels and the scattering scanning mode is to image the skin profile, which can be integrated into the same system with the same FoV as shown in Fig. 6(a). To make sure the light sheet generating device is stable and consistent, both illumination optics share a single-mode fiber, and switching between the two modes can be realized by feeding the laser input from the other end of the fiber. By merging the two results of vessel image and skin surface image of the mice, the depth information of vessels under the skin surface is finally shown in Fig. 6(e). The longest depth of vessels we see from the skin is 1.95 mm. We analyzed three areas of the vessel part which show different depths under the skin. The red part and orange part are of similar width in the view of biology (they are symmetrical on the mice), however, the leaning gesture of the mice makes them have different distances to the skin from the side of view. More or less, can describe the relationship between fluorescence scattering and vascular depth.

We demonstrate that the synergistic use of the time-gated imaging, light-sheet scanning and deep learning algorithm can effectively minimize the influence of laser scattering and out-of-focus background noise, so that vasculature can be clearly resolved by NIR-II volumetric imaging through deep tissue. We show the volumetric imaging system can successfully reconstruct the total length of 527.7 mm of vessels from the whole mice thick tissue within 2 mm from the skin, including the smallest feature size of 0.1 mm diameter vessels. The whole mice scale 3D imaging with acceptable depth resolution of 0.1 mm fills the gap between the conventional 2D microscopic and macroscopic imaging, which enables the studies of the whole-body pathology models. With the large volume of information, the system provides capability of analyzing the microscopic details of macroscopic region of interest in the context of the intact body²⁸, such as the transport of cancer cells in blood vessels²⁹. High speed in vivo volumetric imaging with high contrast and high resolution offers superior advantages compared with the current tissue transparency techniques used to avoid scattering effects in imaging the details of fixed organs.

There are several directions for further improvements. The emerging field of in vivo volumetric imaging will highly demand NIR cameras to be developed with high resolution and sensitivity. For example, the maximum number of pixels from the best available NIR-II cameras is currently limited at 640 × 512, which is far from being sufficient in resolving vessels with the smallest diameter towards only five micrometers³⁰. Though the power of NIR-II imaging has been improved in imaging the gastrointestinal tract and bones of mice³¹, as the result of strong absorption and scattering by the tissue, the penetration depth for NIR-II illumination and emission detection from the NIR-II contrast agent through deep tissue remain limiting the optical system to resolve the vessel networks deeper than 2 mm from the skin. Our current design of the light-sheet scanning only allows one side of the mice to be illuminated, which leads to the mice skin partially obscured, and rotating scanning scheme can be used to cover the entire angles of the mice. Techniques, such as adaptive optics, multi-sheet illumination imaging, structured light illumination imaging, etc, can be adopted to enhance the performance. Last but not the least, the rapid progress in material sciences with bright and more efficient luminescent contrast agents to be developed will continue to accelerate both the fundamental research and pre-clinical translations of NIR-II volumetric imaging, as well as the field of imaging guided minimally invasive diagnostic and therapeutic procedures³².