Non-linear machine learning frequency spectrum fusion super-resolution imaging method based on step-by-step saturation

Abstract

The invention discloses a non-linear machine learning frequency spectrum fusion super-resolution imaging method based on gradual saturation, and the method enables a point spread function of emitted light (namely fluorescence) to be gradually transited from an annular shape to a Gaussian shape through adjusting the power of excitation light, thereby effectively extracting the specific spatial frequency information in an image. Meanwhile, by combining a deep recursive residual network (DRRN), the series of gradually saturated point spread functions are fused into full-space frequency spectrum information, so that the imaging speed which is twice that of a traditional method and high-quality super-resolution nanometer microscopic imaging are realized. An example verifies that under the condition of low excitation power, 33nm spatial resolution imaging of a single up-conversion nanoparticle is realized. The invention not only simplifies the construction of an optical system, but also opens up a new way for high-quality and low-phototoxicity super-resolution microscopic imaging.

Description

Technical Field

[0001] The present invention belongs to the technical field of super-resolution fluorescence microscopy, and specifically relates to a super-resolution imaging method based on stepwise saturation non-linear machine learning spectrum fusion. Background Art

[0002] Stimulated emission depletion (STED) microscopy achieves super-resolution imaging by precisely aligning a Gaussian excitation beam with an annular quenching beam. This technique has attracted wide attention in the scientific community due to its ability to image in real time without computational reconstruction. However, phototoxicity, operational complexity, and strict requirements for fluorescent dyes remain challenges for the widespread application of STED. To address these issues, a single-beam annular light super-resolution imaging technique based on upconversion nanoparticles uses the non-linear response characteristics of upconversion nanoparticles to achieve super-resolution imaging with only a single-beam annular excitation beam, thus significantly simplifying the optical configuration of the system and reducing the power of the excitation beam to well below that of the quenching beam, thereby reducing phototoxicity. To expand the frequency shift coverage range of the optical system, scientists have proposed a multi-color Fourier domain fusion algorithm that fuses Fourier components from each emission band and processes the optical transfer function containing optimized spatial information to achieve an effective point spread function. Despite these important advances, significant challenges still remain in forming the final effective point spread function. For example, directly using the annular point spread function for deconvolution may produce artifacts, and some frequency bands are lost in the Fourier domain, affecting image quality. Eliminating artifacts by setting negative side lobes to zero results in information loss and ineffective reconstruction of super-resolution images. The Fourier domain fusion method only provides two spectra with limited information content. In addition, differences in image signal-to-noise ratio lead to mismatched frequency domain information, amplifying background noise in local regions. For dense samples, the noisy background hinders the recovery of fine structures, and weak intensity points are easily filtered out directly.

[0003] Lanthanide-doped upconversion nanoparticles can effectively address the problem of limited spectral information. Upconversion nanoparticles can sequentially absorb two long-wavelength low-energy photons and convert them into a single high-energy photon, exhibiting unique optical properties by converting near-infrared laser light into short-wavelength light including visible light and ultraviolet light. In addition, upconversion nanoparticles have multiple intermediate energy levels, and the upconversion emission process can be achieved through stepwise energy transfer, which provides a non-linear response between the electron generation rate at the intermediate energy levels and the excitation power. In an ideal situation, the multiple intermediate energy levels and non-linear characteristics of upconversion nanoparticles enable a single-beam annular beam to transition from unsaturated excitation to saturated excitation, providing rich spatial frequency information for super-resolution imaging. However, when multiple fluorescent emitters overlap, information loss and background noise inevitably occur, so traditional super-resolution microscopes perform poorly in dense samples. Summary of the Invention

[0004] The present invention proposes a nonlinear machine learning spectral fusion super-resolution imaging method based on gradual saturation, which is used to improve the imaging quality of a microscopic imaging system and reduce the required excitation light power.

[0005] The technical solution for achieving the object of the present invention is: a nonlinear machine learning spectral fusion super-resolution imaging method based on gradual saturation, comprising the following steps:

[0006] Step 1: Build a single-beam stimulated emission depletion (STED)-like super-resolution two-photon negative-biased microscopy imaging system;

[0007] Step 2: Use an annular beam to perform single-point scanning on the upconversion nanoparticles, and capture the fluorescence signal emitted by the upconversion nanoparticles;

[0008] Step 3: Adjust the excitation light power to obtain an annular fluorescence image and a saturated Gaussian-like fluorescence image at different excitation powers;

[0009] Step 4: Use the DRRN model to fuse the high-frequency and low-frequency spatial information of the annular fluorescence image and the saturated Gaussian-like fluorescence image at different excitation powers.

[0010] Preferably, the single-beam STED-Like super-resolution two-photon negative-biased microscopy imaging system includes:

[0011] A laser, a collimating lens L1, a half-wave plate, a polarization beam splitter, a mirror M1, a mirror M2, a mirror M3, a mirror M4, a mirror M5, a vortex phase plate, a short-pass dichroic mirror, a quarter-wave plate, a high numerical aperture objective lens, a 4f optical system composed of lenses L2 and L3, an imaging lens L4, an electric flip mirror, a detector, and a single-photon counting avalanche photodiode;

[0012] The laser emits laser light, which is collimated by the collimating lens L1, then passes through the half-wave plate and the polarization beam splitter, and then forms an annular light with zero central light intensity on the focal plane after passing through the mirror M1, the mirror M2, the mirror M3, and the vortex phase plate. Then, the optical path is changed by the mirror M4 to enter the short-pass dichroic mirror, and the excitation beam is converted from linear polarization to circular polarization by the quarter-wave plate, and then focused on the sample slide by the objective lens to complete the excitation of the upconversion nanoparticles; the fluorescence signal emitted by the sample is collected by the same objective lens, separated from the excitation beam by the dichroic mirror, the optical path of the fluorescence signal is changed by the mirror M5, then the optical path is expanded by the 4f optical system and passes through the imaging lens L4, and through the electric flip mirror, it is selected to enter the detector, or coupled into a multimode optical fiber and the photons are collected by the single-photon counting avalanche photodiode.

[0013] Preferably, the upconversion nanoparticles contain 40% Yb3+ and 4% Tm 3+ Doped NaYF4 particles.

[0014] Preferably, the specific method of adjusting the excitation light power to obtain different annular emission point spread functions and saturated Gaussian-like emission point spread functions is:

[0015] According to step 1, a single-beam STED-type super-resolution two-photon negative contrast microscopy imaging system is established to scan a single upconversion nanoparticle sample;

[0016] When scanning begins, the sample is positioned on a high-precision three-dimensional translation stage and is excited in a serpentine scanning manner; when the laser scans the upconversion nanoparticles to be tested, the upconversion nanoparticles will be activated to emit ring-shaped fluorescence; as the excitation power gradually increases, the point spread function of a single fluorescent particle presents a nonlinear saturated ring shape; when the excitation power reaches the saturation threshold, the fluorescence point spread function eventually turns into a Gaussian-like shape.

[0017] Preferably, the DRRN model includes an input convolution layer, a recursive residual module and an output convolution module, and the processing process of the DRRN model is specifically as follows:

[0018] Convolve the three random particle point images with the annular fluorescence point spread function to obtain three convolved images;

[0019] A random particle image is convolved with a point-saturated Gaussian-like fluorescence image to obtain a convolved image;

[0020] The four convolved images are formed into a four-channel tensor as the input of the DRRN model, and the four-channel tensor is input into the convolution layer for feature extraction and fusion;

[0021] The output features of the input convolutional layer are sent to the recursive residual module as the initial input features for processing. The recursive residual module includes 25 recursive residual blocks, and each recursive residual block contains two convolutional layers. In each recursive residual block, the fused features first pass through the first convolutional layer, then pass through the ReLU activation function, and then pass through the second convolutional layer. The output features of each recursive residual block are added to the initial input features and enter the next recursive residual block. The feature map processed by 25 recursive residual blocks is then subjected to feature fusion through the output convolutional layer, and finally upsampled to obtain a high-resolution output image.

[0022] Compared with the prior art, the present invention has the following significant advantages:

[0023] The present invention effectively fuses high-frequency and low-frequency spatial information through a step-by-step saturation point spread function adjustment strategy, extracts more spatial frequency information during a single imaging process, and combines the global and local residual learning of a deep recursive residual network, greatly improving the quality of super-resolution microscopy imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flowchart of the present invention.

[0025] Figure 2 is a single-beam STED-Like super-resolution two-photon negative bias microscopy imaging system.

[0026] Figure 3 is a conceptual diagram of super-resolution imaging based on step-by-step saturation non-linear machine learning spectral fusion.

[0027] Figure 4 are experimental results of upconversion nanoparticles under sub-diffraction volume. DETAILED DESCRIPTION OF THE INVENTION

[0028] The technical details and operation steps of this example of the present invention will be described in detail below in combination with relevant drawings to ensure that technicians can fully understand and effectively implement this technology. The drawings will cover key experimental setups and imaging results, from basic system construction to complex data processing and image fusion processes, to help those skilled in the art accurately understand and implement the present invention.

[0029] A super-resolution imaging based on step-by-step saturation non-linear machine learning spectral fusion, the process of this method is as Figure 1 shown, and includes the following steps:

[0030] Step 1: Build a single-beam STED-Like super-resolution two-photon negative bias microscopy imaging system, as Figure 2 shown. The required optical elements and specific methods are as follows:

[0031] A 980 nm diode laser coupled with a single-mode optical fiber is used to emit laser light. After passing through the collimating lens L1, the laser beam is collimated. Then it passes through a half-wave plate and a polarization beam splitter. This configuration enables precise regulation of the laser power by electronically rotating the half-wave plate. Subsequently, after passing through the mirrors M1, M2, M3 and the vortex phase plate, an annular light with zero central light intensity on the focal plane is formed. Then, the optical path is changed by the mirror M4 to enter a short-pass dichroic mirror, and the excitation beam is converted from linear polarization to circular polarization through a quarter-wave plate. Then it is focused on the sample slide by a high numerical aperture objective lens to complete the excitation of the upconversion nanoparticles. The fluorescence signal emitted by the sample is collected by the same objective lens and separated from the excitation beam by the dichroic mirror. The fluorescence signal changes the reflection optical path through the mirror M5, then passes through a 4f optical system to expand the optical path and passes through the imaging lens L4. It can be selectively entered into the detector through an electrically actuated flip mirror, or coupled into a multimode optical fiber and the photons are collected by a single-photon counting avalanche photodiode. In addition, the multimode optical fiber can be switched to a spectrometer equipped with an EMCCD detector.

[0032] Step 2: Use the annular beam to perform single-point scanning on the upconversion nanoparticles and capture the fluorescence signal emitted by the upconversion nanoparticles, as Figure 3 shown.

[0033] The scanning process is mainly as follows: The laser is preliminarily shaped by a collimating objective lens, and then the polarization state is adjusted through a half-wave plate and a polarization beam splitter. The beam is converted into a vortex annular beam with an optical center singularity through a vortex phase plate, and then the annular beam is focused on the sample through a high numerical aperture objective lens. Subsequently, the precise movement of the sample is controlled by a precise scanning stage. When the upconversion nanoparticles are excited by the annular light and emit fluorescence, the fluorescence signal is collected by the same objective lens as the excitation optical path and separated from the excitation beam by the dichroic mirror. The separated fluorescence signal is further shaped by a 4f optical system and introduced into a multimode optical fiber for transmission and collected by a single-photon counter, thereby realizing point-by-point scanning of the sample by the beam.

[0034] The upconversion nanoparticles contain 40% Yb 3+ and 4% Tm 3+ doped NaYF4 particles, and non-linear saturation excitation of 800 nm fluorescence emission can be achieved through stepwise energy transfer.

[0035] Among them, the preparation method of the upconversion nanoparticles is as follows:

[0036] Step 2.1: Preparation of Y-OA: Weigh 2.5 mmol of YCl3·6H2O, 10 ml of oleic acid and 15 ml of octadecane. Put these reaction mixtures into a 100 ml three-necked flask equipped with a stir bar, install a heating device, connect a condenser and a temperature probe. Continuously introduce nitrogen and gradually heat to 70 °C. Keep at this temperature for 30 minutes, then slowly raise the temperature to 150 °C and keep for 1 hour until the solution becomes clear and transparent. Let the solution cool naturally to room temperature, and then transfer it to a glass bottle for storage at room temperature.

[0037] Step 2.2: Preparation of NaTEA-OA: Weigh 4 mmol of sodium trifluoroacetate and 10 ml of oleic acid, and put them into a 100 ml three-necked flask equipped with a stir bar. Stir at room temperature for 2 hours until the solution becomes clear and transparent. Transfer the solution to a glass bottle for storage at room temperature.

[0038] Step 2.3: Preparation of upconversion nanocrystal nuclei (NaYF4:Yb,Tm): First, weigh 0.56 mmol of YCl3·6H2O, 0.4 mmol of YbCl3·6H2O and 0.04 mmol of TmCl3·6H2O. Add these reaction mixtures to a 100 ml three-necked flask containing 4 ml of oleic acid and 15 ml of octadecane, and the flask is equipped with a magnetic stirrer. Place the three-necked flask in a heating device, connect a condenser and a temperature probe, and continuously introduce nitrogen. Slowly heat to 70 °C, evacuate for 30 minutes, then gradually heat to 100 °C and 120 °C, and keep at each temperature for 20 minutes. Switch to a nitrogen atmosphere, heat to 140 °C and keep for 60 minutes until the powder is completely dissolved into a clear solution. Let the reaction mixture cool naturally to room temperature, then add a methanol solution (10 ml) containing NaOH (2.5 mmol) and NHF (4 mmol), and stir for 30 minutes. Next, evacuate and heat to 70 °C for 40 minutes. Switch to ventilation protection, quickly heat to 285 °C, and react for 15 minutes. Cool to room temperature, precipitate the solution with anhydrous ethanol, centrifuge at 4000 rpm for 5 minutes, and wash twice with ethanol. Disperse the precipitate in 10 ml of cyclohexane and store at 4 °C.

[0039] Step 3: Adjust the excitation light power to obtain the annular fluorescence images and saturated Gaussian-like fluorescence images at different excitation powers. The specific process is as follows:

[0040] By taking advantage of the multiple long-lived intermediate energy levels in the upconversion nanoparticles, the sensitizer Yb 3+ can directly absorb 980 nm photons in the near-infrared region. Through a step-by-step energy transfer process, it pumps the ground-state electrons of the activator Tm 3+ to the excited state, and finally realizes the near-infrared 800 nm two-photon-like nonlinear fluorescence emission, as shown in Figure 3 (b).Figure 3 (a) is a schematic diagram of the negative confocal setup. A vortex phase plate is used to generate a ring-shaped excitation beam on the sample plane, and then a single-photon counting avalanche photodiode captures the fluorescence intensity at different excitation powers. Figure 3 (b) is nanoparticle NaYF4: 40% Yb 3+ , 4% Tm 3+ The inset shows the nanoparticle model and a simplified energy level diagram. The multiphoton near-infrared upconversion emission mainly comes from the two-photon excited state (800nm, 3 H4→ 3 H6). Figure 3 (c) is the power-dependent saturation intensity curve of 800nm ​​emission. The inset is the simulation result of the power-dependent point spread function pattern of 800nm ​​emission of a single upconversion nanoparticle under 980nm annular beam. Figure 3 (d) is the normalized central cross-sectional profile of the optical transfer function, and the inset is the optical transfer function of the Gaussian point spread function, the annular point spread function, and the optical transfer function fusion result at different powers. Figure 3 (e) in the figure is the Deep-Fusion training process. Based on the deep recursive residual network, Deep-Fusion takes four captured images from the Gaussian point spread function and the annular point spread function at different powers as the input of the generator and generates super-resolution results. The scale bar is 1μm.

[0041] According to step 1, a single-beam STED-like super-resolution two-photon negative contrast microscopy system was established to scan single-particle samples, and the experiment detected 800nm ​​fluorescence emission with a two-photon-like process. When the system starts scanning, the sample is located on a high-precision three-dimensional translation stage, and the sample is excited in a serpentine scanning manner. When the laser scans the upconversion nanoparticles to be tested, the particles will be activated to emit ring-shaped fluorescence. As the excitation power gradually increases, the fluorescence point spread function presents a nonlinear saturated ring shape. When the excitation power reaches the saturation threshold, the fluorescence point spread function eventually turns into a Gaussian-like shape, such as Figure 3 (c) as shown.

[0042] Furthermore, the Gaussian-like emission point spread function and the annular emission point spread function at different powers are converted into optical transfer functions in the Fourier domain. Compared with their corresponding Gaussian-shaped point spread functions, the annular emission point spread functions at different powers contain more information at high spatial frequencies, while the Gaussian-shaped point spread function supplements the missing information in the mid- and low-frequency parts, such as Figure 3 Therefore, by spectrally fusing the fluorescence point spread functions with specific responses in the Fourier domain, the resulting emission point spread function comprehensively covers both low-frequency and high-frequency information.

[0043] Step 4: The DRRN model is used to fuse the high-frequency and low-frequency spatial information of the fluorescence point spread function at different excitation powers. The DRRN model is as shown in Figure 3 (e).

[0044] The module design of the DRRN model enables it to effectively process multi-channel input data and enhances the feature expression ability through layer-by-layer feature extraction and residual connections. The DRRN model mainly includes an input convolutional layer, a recursive residual module, and an output convolutional module. The training of the DRRN model needs to be completed by a dataset constructed by simulation. The actual training dataset includes 500 groups of images for training and 100 groups of images for validation. Each group of the dataset contains four images, including three convolved images obtained by convolving three random particle point images with the annular fluorescence point spread function respectively, and one convolved image obtained by convolving a random particle point image with the saturated Gaussian-like fluorescence point spread function;

[0045] The four convolved images are formed into a four-channel tensor. Then, these images are synthesized into a four-channel tensor as the network input, and then feature extraction and fusion are performed through the input convolutional layer.

[0046] The output of the input convolutional layer is sent as the initial input feature into the recursive residual module. The recursive residual module includes 25 recursive residual blocks for processing, and each recursive residual block contains two convolutional layers; the initial input feature is processed through 25 recursive residual modules. In each recursive residual block, the feature first passes through the first convolutional layer, then through the ReLU activation function, and then through the second convolutional layer. This process gradually enhances the feature representation ability, enabling the network to better learn the detailed information of the image. During the processing of each recursive residual block, the output feature of each recursive residual block is added to the initial input feature and then enters the next recursive residual block to achieve residual learning, realize "parameter sharing", and obtain more convolutions with fewer parameters. The feature map after being processed by 25 recursive residual blocks is then subjected to feature fusion through the output convolutional layer and finally upsampled to obtain a high-resolution output image.

[0047] In addition, the network is optimized using the MSE loss function, and finally a super-resolution image with more details, textures, and higher resolution is generated. Validation on multiple datasets shows that DRRN has superiority in terms of image restoration quality and model efficiency, which is particularly crucial for image processing applications.

[0048] The present invention uses a single-beam annular beam to gradually saturate and excite upconversion nanoparticles, and by adjusting the excitation power, different fluorescence point spread functions are generated. Subsequently, through the DRRN model, the high-frequency and low-frequency spatial information captured at different powers is fused, and the frequency of the progressive saturation point spread function is incorporated into the full-space spectral information, thereby realizing the reconstruction of super-resolution images. This method achieves a spatial resolution of 33 nm, corresponding to 1 / 29 of the excitation wavelength, and doubles the typical imaging speed.

[0049] Example:

[0050] A self-made single-beam STED-like super-resolution two-photon negative-contrast confocal microscopy system was used to scan upconversion nanoparticles with non-linear progressive saturation excitation. First, 800-nm fluorescence progressive saturation annular scanning imaging of upconversion nanoparticles was performed using a standard two-photon-like confocal system. Thanks to the non-linear excitation characteristics of upconversion, the experimental results Figure 4 (a)-(d) show the transition from a unique multi-order annular point spread function to a Gaussian-like point spread function. Figure 4 In (a)-(d), the images of the 800-nm emission band of a single upconversion nanoparticle under the conditions of 80 mW, 10 mW, 5 mW, and 3 mW of a 980-nm annular beam, respectively. (e) Super-resolution imaging results (deconvolution fusion) obtained by fusing the optical transfer functions of (a)-(d) in the Fourier domain. (f) Output results of the DRRN architecture (Decon.Fusion). (g)-(j) The images of the 800-nm emission band of double-fluorescent nanorods under the conditions of 80 mW, 10 mW, 5 mW, and 3 mW of a 980-nm annular beam, respectively. (k) Super-resolution imaging results (Decon.Fusion) obtained by fusing the optical transfer functions of (g)-(j) in the Fourier domain. (l) Output results of the DRRN architecture (Deep-Fusion). (m)-(p) The images of the 635-nm emission band of upconversion nanoparticles doped with 2% Er 3+ and 20% Yb 3+ under the conditions of 80 mW, 10 mW, 5 mW, and 3 mW of a 980-nm annular beam, respectively. (q) Super-resolution imaging results (deconvolution fusion) obtained by fusing the optical transfer functions of (m)-(p) in the Fourier domain. (r) Output results of the DRRN architecture (Deep-Fusion). The inset is the enlarged area of the white square. (s) Line profiles of two adjacent upconversion nanoparticles from Fig. (a) to Fig. (f). (t) Line profiles of four adjacent upconversion nanoparticles from Fig. (g) to Fig. (l). (u) Line profiles of a single UCNP from Fig. (m) to Fig. (r). The scale bars for all Figs. (a)-(r) are 1 micron, and the inset is 200 nm.

[0051] Due to the limitation of the diffraction limit, Gaussian-like imaging cannot resolve individual nanoparticles within the diffraction limit ( Figure 4 (a) inset). In contrast, negative-contrast confocal imaging provides higher frequency information for resolving the positions of individual upconversion nanoparticles ( Figure 4 (b)-(d) inset). In addition, this experiment applied the Fourier-domain fusion algorithm, which combined the optical transfer functions of Gaussian-like imaging and negative-offset confocal imaging, to obtain a super-resolution result with rich frequency-domain information ( Figure 4 (e)), achieving a resolution of up to 48 nm. By inputting the Gaussian-like emission image and the negative-contrast confocal image into the network, the Deep-Fusion model trained based on the simulated dataset produced the Figure 4 output shown in (f). The experimental results demonstrated the ability to resolve discrete nanoparticles within a distinguishable distance of 33 nm. Further verification of the feasibility of the method of the present invention was carried out using dual-fluorescent nanorods with a fixed spacing of 210 nm ( Figure 4 (g)-(j)). The magnified comparison images marked by white boxes ( Figure 4 (g)-(l) inset) showed that the Fourier-domain fusion method failed to accurately recover the particles in the high-concentration sample and produced image artifacts ( Figure 4 (k) inset), but Deep-Fusion effectively solved these problems and clearly identified pairs of fluorescent particles ( Figure 4 (l) inset). The particle line profiles along the white dashed line ( Figure 4 (t)) showed that the resolution of individual upconversion nanoparticles was significantly improved to 74 nm. This method utilized the non-linear saturation response of upconversion nanoparticles and convolutional neural networks to effectively improve the image resolution and interpretability without increasing the system complexity.

[0052] To further demonstrate the versatility of this method, in this example, upconversion nanoparticles doped with Er 3+ and Yb 3+ (NaYF4: 2% Er 3+ , 20% Yb 3+ ) were used for single-ring beam scanning in the 650 nm fluorescence channel. The imaging results are shown in Figure 4 (m)-(p). The experiment scanned over-saturated Gaussian-like imaging and negative-offset imaging at different powers. After being processed by the Fourier-domain fusion algorithm and convolutional neural networks, the results are shown in Figure 4 (q)-(r). The local area within the white box was further magnified ( Figure 4 (m)-(r) inset) and the intensity distribution along the white dashed line was analyzed ( Figure 4(u)), The quantitative results show that the full width at half maximum is significantly enhanced from 288 nm to 41 nm (about λ / 24).

Claims

1. A non-linear machine learning spectral fusion super-resolution imaging method based on gradual saturation, characterized in that The following steps are involved: Step 1: Build a single-beam stimulated emission depletion-like super-resolution two-photon negatively polarized microscopy imaging system; Step 2: Use an annular beam to perform single-point scanning on the upconversion nanoparticles to capture the fluorescence signal emitted by the upconversion nanoparticles; Step 3: Adjust the excitation light power to obtain annular fluorescence images and saturated Gaussian-like fluorescence images at different excitation powers; Step 4: The high-frequency and low-frequency spatial information of the annular fluorescence images and saturated Gaussian-like fluorescence images under different excitation powers are fused through the DRRN model.

2. The method for super-resolution imaging of spectrum fusion based on stepwise saturation nonlinear machine learning according to claim 1, wherein The single beam STED-Like super-resolution two-photon negatively polarized microscopy imaging system comprises: Laser, collimating lens L1, half wave plate, polarization beam splitter, reflector M1, reflector M2, reflector M3, reflector M4, reflector M5, vortex phase plate, short-pass dichroic mirror, quarter wave plate, high numerical aperture objective lens, 4f optical system composed of lens L2 and lens L3, imaging lens L4, electric flip mirror, detector and single photon counting avalanche photodiode; The laser emits a laser, and the laser beam is collimated after passing through the collimating lens L1, and then passes through a half-wave plate and a polarization beam splitter, and then passes through a reflector M1, a reflector M2, a reflector M3 and a vortex phase plate to form a ring light with a central light intensity of zero on the focal plane, and then changes the light path through the reflector M4 to enter a short-pass dichroic mirror and converts the excitation beam from linear polarization to circular polarization through a quarter-wave plate, and then is focused on a sample slide by an objective lens to complete the excitation of up-conversion nanoparticles; the fluorescence signal emitted by the sample is collected through the same objective lens and separated from the excitation beam by a dichroic mirror, and the fluorescence signal changes the light path through the reflector M5, and then passes through a 4f optical system to expand the light path and passes through an imaging lens L4, and through an electric flip mirror, is selected to enter a detector, or is coupled into a multimode optical fiber and the photons are collected by a single-photon counting avalanche photodiode.

3. The method for super-resolution imaging of spectrum fusion based on stepwise saturation nonlinear machine learning according to claim 2, wherein, The upconversion nanoparticles contain 40% Yb 3+ and 4% Tm 3+ doped NaYF4 particles.

4. The method for super-resolution imaging of spectrum fusion based on stepwise saturation nonlinear machine learning according to claim 2, wherein The specific method of adjusting the excitation light power to obtain different annular emission point spread functions and saturated Gaussian emission point spread functions is: According to step 1, a single-beam STED-type super-resolution two-photon negative contrast microscopy imaging system is established to scan a single upconversion nanoparticle sample; When scanning begins, the sample is positioned on a high-precision three-dimensional translation stage and is excited in a serpentine scanning manner; when the laser scans the upconversion nanoparticles to be tested, the upconversion nanoparticles will be activated to emit ring-shaped fluorescence; as the excitation power gradually increases, the point spread function of a single fluorescent particle presents a nonlinear saturated ring shape; when the excitation power reaches the saturation threshold, the fluorescence point spread function eventually turns into a Gaussian-like shape.

5. The method for super-resolution imaging of spectrum fusion based on stepwise saturation nonlinear machine learning according to claim 2, characterized in that, The DRRN model includes an input convolution layer, a recursive residual module and an output convolution module. The processing process of the DRRN model is specifically as follows: Convolve the three random particle point images with the annular fluorescence point spread function to obtain three convolved images; A random particle image is convolved with a point-saturated Gaussian-like fluorescence image to obtain a convolved image; Four convolved images are formed into a four-channel tensor as the input of the DRRN model, and the four-channel tensor is subjected to feature extraction and fusion by the input convolutional layer; The output features of the input convolutional layer are sent to the recursive residual module as the initial input features for processing. The recursive residual module includes 25 recursive residual blocks for processing, and each recursive residual block contains two convolutional layers. In each recursive residual block, the fused features first pass through the first convolutional layer, then through the ReLU activation function, and then through the second convolutional layer. The output features of each recursive residual block are added to the initial input features and then enter the next recursive residual block. The feature map after 25 recursive residual blocks is then subjected to feature fusion by the output convolutional layer and finally upsampled to obtain a high-resolution output image.

Images