A method and related apparatus for stimulated raman scattering microscopy of active tissue

Abstract

The application discloses a stimulated Raman scattering microscopic imaging method suitable for active tissue and a related device, relates to the technical field of optical imaging, and comprises the following steps: acquiring a high-speed initial stimulated Raman scattering microscopic image of the active tissue under low laser power and low exposure time, wherein the used laser power is less than a preset power, the single-pixel exposure time is less than a preset time, and the imaging speed is greater than a preset speed; processing the initial stimulated Raman scattering microscopic image by using a trained signal-to-noise ratio enhancement model to obtain a processed image; and processing the processed image by using a trained morphological feature and resolution recovery model to obtain a stimulated Raman scattering microscopic image of the active tissue. The application can obtain the stimulated Raman scattering microscopic image of the active tissue with high signal-to-noise ratio and high resolution under the premise of guaranteeing the activity of the active tissue.

Description

Technical Field

[0001] This application relates to the field of optical imaging technology, and in particular to a stimulated Raman scattering microscopy imaging method and related apparatus suitable for active tissues. Background Technology

[0002] In the field of biomedical research, high-resolution, label-free chemical imaging of ex vivo viable tissues is crucial for point-of-care diagnosis, tissue viability assessment, and subsequent transplantation therapy (such as cryopreserved ovarian tissue transplantation). Stimulated Raman scattering (SRS) microscopy, as a label-free imaging technique with high chemical specificity, holds great potential in this field.

[0003] However, traditional SRS microscopy has significant bottlenecks when applied to living tissues: on the one hand, in order to obtain high signal-to-noise ratio and high resolution images, high laser power and long single-pixel exposure time are usually required, resulting in slow imaging speed. This can cause serious photodamage to living tissues and affect their subsequent activity. On the other hand, in order to shorten the imaging time and reduce ex vivo damage to living tissues, traditional high-speed scanning methods often sacrifice image signal-to-noise ratio and resolution. Summary of the Invention

[0004] The purpose of this application is to provide a stimulated Raman scattering microscopy imaging method and related apparatus suitable for active tissues, which can obtain stimulated Raman scattering microscopic images of active tissues with high signal-to-noise ratio and high resolution while ensuring the activity of the active tissues.

[0005] To achieve the above objectives, this application provides the following solution.

[0006] In a first aspect, this application provides a stimulated Raman scattering microscopy imaging method suitable for living tissues, comprising: Acquire an initial stimulated Raman scattering microscopic image of active tissue; when obtaining the initial stimulated Raman scattering microscopic image, the laser power used is less than a preset power, the single pixel exposure time is less than a preset time, and the imaging speed is greater than a preset speed. The initial stimulated Raman scattering microscopic image is processed using a trained signal-to-noise ratio enhancement model to obtain a processed image; the trained signal-to-noise ratio enhancement model adopts a diffusion model. The trained morphological feature and resolution restoration model is used to process the processed image to obtain stimulated Raman scattering microscopic images of active tissue; the trained morphological feature and resolution restoration model adopts a recurrent adversarial generative network model.

[0007] Optionally, when performing stimulated Raman scattering microscopy on active tissue, a stimulated Raman scattering microscopy system is used, which includes: a laser, a stimulated Raman scattering microscopy optical path, a resonant scanning galvanometer, a scanning mirror, a tube mirror, a first reflecting mirror, an objective lens, a displacement stage, a condenser lens, a second reflecting mirror, a lens, a photodetector, a lock-in amplifier, and a host computer, with the active tissue located on the displacement stage; The laser emitted by the laser and the stimulated Raman scattering (SRS) microscopic imaging optical path is incident on the active tissue located on the displacement stage via the resonant scanning galvanometer, the scanning mirror, the tube mirror, the first reflecting mirror, and the objective lens, exciting a stimulated Raman scattering signal. The laser carrying the stimulated Raman scattering signal is incident on the photodetector via the condenser mirror, the second reflecting mirror, and the lens, generating an electrical signal. The electrical signal is processed by the lock-in amplifier and the host computer to obtain an initial stimulated Raman scattering microscopic image. The laser includes pump light and Stokes light. The resonant scanning galvanometer, the scanning mirror, the tube mirror, and the displacement stage are used to incident the laser at different positions on the active tissue to scan the active tissue.

[0008] Optionally, the trained signal-to-noise ratio enhancement model includes a time-step embedding unit, N downsampling blocks, and N upsampling blocks. The output of the time-step embedding unit is connected to the input of each downsampling block and the input of each upsampling block, respectively. The output of the i-th downsampling block is connected to the input of the (i+1)-th downsampling block and the input of the (N-i+1)-th upsampling block, respectively. The output of the N-th downsampling block is connected to the input of the 1-th upsampling block, and the output of the i-th upsampling block is connected to the input of the (i+1)-th upsampling block; i = 1, 2, ..., N-1. The time step embedding unit is used to encode the current time step to obtain the embedding vector of the current time step.

[0009] Optionally, the initial stimulated Raman scattering microscopic image is processed using a trained signal-to-noise ratio enhancement model to obtain a processed image, specifically including: The initial stimulated Raman scattering microscopic image is used as the noisy image, and the last denoising time step is used as the current time step; Using the noisy image and the current time step as input, the predicted noise for the current time step is determined using a trained signal-to-noise ratio enhancement model; The noisy image is denoised based on the predicted noise at the current time step to obtain a denoised image; Determine whether the current time step is the first denoising time step; if yes, use the denoised image as the processed image; if no, use the denoised image as the noise image for the next iteration, use the difference between the current time step and 1 as the current time step for the next iteration, and return to the step of "using the noise image and the current time step as input, and using the trained signal-to-noise ratio enhancement model to determine the predicted noise of the current time step".

[0010] Optionally, before processing the initial stimulated Raman scattering microscopic image using the trained signal-to-noise ratio (SNR) enhancement model to obtain the processed image, the method further includes: training the initial SNR enhancement model to obtain a trained SNR enhancement model, specifically including: Acquire the training image and use the first noisy time step as the current time step; Based on random Gaussian noise and the noise scheduling parameters of the current time step, the training image is noise-added to obtain a noisy image; Using the noisy image and the current time step as input, the prediction noise for the current time step is determined using an initial signal-to-noise ratio (SNR) enhancement model; the initial SNR enhancement model and the trained SNR enhancement model have the same structure. Determine whether the current time step is the last noisy time step, and obtain the determination result; If the judgment result is negative, the sum of the current time step and 1 is used as the current time step of the next iteration, and the step of "adding noise to the training image based on random Gaussian noise and the noise scheduling parameter of the current time step to obtain the noisy image" is returned. If the judgment result is yes, then based on the predicted noise of all noise-added time steps, determine the predicted noise image, calculate the error between the predicted noise image and the real noise image, obtain the total loss, and use the total loss to update the initial signal-to-noise ratio enhancement model to obtain the updated model; determine whether the first iteration termination condition has been met; if yes, then use the updated model as the trained signal-to-noise ratio enhancement model; if no, then use the updated model as the initial signal-to-noise ratio enhancement model for the next iteration, and return to the "obtain training image" step.

[0011] Optionally, the recurrent adversarial generative network model includes a first generator, a second generator, a first discriminator, and a second discriminator. In this case, before processing the processed image using the trained morphological feature and resolution restoration model to obtain a stimulated Raman scattering micrograph of the active tissue, the model further includes: training the initial morphological feature and resolution restoration model to obtain a trained morphological feature and resolution restoration model, specifically including: Acquire a first training image and a second training image; both the first training image and the second training image are images obtained by stimulated Raman scattering microscopy imaging of active tissue. When obtaining the first training image, the laser power used is less than a preset power, the single pixel exposure time is less than a preset time, and the imaging speed is greater than a preset speed. When obtaining the second training image, the laser power used is greater than a preset power, the single pixel exposure time is greater than a preset time, and the imaging speed is less than a preset speed. Using the first training image as input, the first generator is used to obtain the first generated image, and using the first generated image as input, the second generator is used to obtain the first reconstructed image. Using the second training image as input, the second generator is used to obtain the second generated image, and using the second generated image as input, the first generator is used to obtain the second reconstructed image. The first discriminator is used to distinguish between the first training image and the second generated image to obtain a first discrimination result; The second discriminator is used to distinguish between the second training image and the first generated image to obtain a second discrimination result; Based on the first discrimination result and the second discrimination result, the adversarial loss and the discriminator loss are calculated respectively. Based on the first training image, the first reconstructed image, the second training image and the second reconstructed image, the cycle consistency loss is calculated. The generator loss is obtained by weighted summation of the adversarial loss and the cycle consistency loss. The generator loss is used to update the first generator and the second generator to obtain a first updated generator and a second updated generator. The discriminator loss is used to update the first discriminator and the second discriminator to obtain a first updated discriminator and a second updated discriminator. Determine whether the second iteration termination condition has been met; if yes, use the first updated generator as the trained morphological feature and resolution recovery model; if no, use the first updated generator, the second updated generator, the first updated discriminator, and the second updated discriminator as the first generator, the second generator, the first discriminator, and the second discriminator for the next iteration, respectively, and return to the step of "obtaining the first training image and the second training image".

[0012] Optionally, the first generator and the second generator have the same structure. The first generator includes an encoder, a residual module, and a decoder connected in sequence. The encoder includes multiple convolutional blocks connected in sequence. The residual module includes multiple residual blocks connected in sequence. Each residual block includes a first convolutional layer, a first instance normalization layer, a ReLU activation function layer, a second convolutional layer, a second instance normalization layer, and a splicing layer connected in sequence. The input of the splicing layer is also connected to the input of the first convolutional layer. The decoder includes multiple deconvolutional blocks and a first output layer connected in sequence. The first discriminator and the second discriminator have the same structure; the first discriminator includes multiple convolutional blocks connected in sequence and a second output layer.

[0013] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the above-described stimulated Raman scattering microscopy imaging method for active tissues.

[0014] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described stimulated Raman scattering microscopy imaging method applicable to active tissues.

[0015] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described stimulated Raman scattering microscopy imaging method applicable to active tissues.

[0016] According to the specific embodiments provided in this application, this application has the following technical effects.

[0017] This application provides a stimulated Raman scattering (SRS) microscopy imaging method and related apparatus suitable for active tissues. It acquires an initial SRS microscopic image of active tissue. When obtaining the initial SRS microscopic image, the laser power used is less than a preset power, the single-pixel exposure time is less than a preset time, and the imaging speed is greater than a preset speed. Thus, by using lower laser power, shorter single-pixel exposure time, and faster imaging speed, an initial SRS microscopic image of active tissue is obtained while ensuring the tissue's activity. At this point, the image signal-to-noise ratio and resolution are relatively low. First, the initial stimulated Raman scattering (SRS) micrograph is processed using a pre-trained signal-to-noise ratio (SNR) enhancement model to obtain the processed image. The pre-trained SNR enhancement model uses a diffusion model, which can improve the image SNR. Then, the processed image is processed using a pre-trained morphological feature and resolution restoration model to obtain the stimulated Raman scattering micrograph of the active tissue. The pre-trained morphological feature and resolution restoration model uses a recurrent adversarial generative network (ROV) model, which can improve the image morphological features and resolution. Thus, a high SNR and high resolution stimulated Raman scattering micrograph of the active tissue can be obtained while ensuring the activity of the active tissue. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 An application environment diagram of a stimulated Raman scattering microscopy imaging method for active tissues provided in an embodiment of this application; Figure 2 A schematic flowchart of a stimulated Raman scattering microscopy imaging method for active tissues provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of the optical path for stimulated Raman scattering microscopy provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a stimulated Raman scattering microscopy imaging system provided in an embodiment of this application; Figure 5 This is a schematic diagram of the network structure of a diffusion model provided in an embodiment of this application; Figure 6 A schematic diagram of the network structure of a recurrent adversarial generative network model provided in an embodiment of this application; Figure 7 This is a schematic diagram of the image processing effect provided in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] The stimulated Raman scattering microscopy method for active tissues provided in this application can be applied to, for example... Figure 1 The application environment is illustrated. The terminal communicates with the server via a network. A data storage system stores the data the server needs to process. This system can be set up independently, integrated into the server, or located in the cloud or on another server. The terminal can send an imaging request to the server. Upon receiving the request, the server acquires an initial stimulated Raman scattering (SRS) microscopic image of the active tissue. When obtaining the initial SRS microscopic image, the laser power used is less than a preset power, the single-pixel exposure time is less than a preset time, and the imaging speed is greater than a preset speed. The initial SRS microscopic image is then processed using a trained signal-to-noise ratio (SNR) enhancement model (using a diffusion model). The processed image is then further processed using a trained morphological feature and resolution restoration model (using a recurrent adversarial generative network model). The server can then feed back the obtained SRS microscopic image of the active tissue to the terminal.

[0023] Furthermore, in some embodiments, the stimulated Raman scattering microscopy method applicable to active tissues can also be implemented by a server or a terminal alone. For example, the terminal can directly process the imaging request to be processed, or the server can obtain the imaging request to be processed from the data storage system and process it.

[0024] The terminals can be, but are not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle systems. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. Servers can be implemented using independent servers, server clusters composed of multiple servers, or cloud servers.

[0025] In one exemplary embodiment, such as Figure 2 As shown, a stimulated Raman scattering microscopy imaging method suitable for active tissues is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 The following steps, S1 to S3, are used as an example to illustrate the process of using a server in the example.

[0026] Step S1: Obtain an initial stimulated Raman scattering microscopic image of the active tissue; when obtaining the initial stimulated Raman scattering microscopic image, the laser power used is less than the preset power, the single pixel exposure time is less than the preset time, and the imaging speed is greater than the preset speed.

[0027] Step S2: The initial stimulated Raman scattering microscopic image is processed using the trained signal-to-noise ratio enhancement model to obtain the processed image; the trained signal-to-noise ratio enhancement model adopts a diffusion model.

[0028] Step S3: The processed image is processed using the trained morphological feature and resolution restoration model to obtain a stimulated Raman scattering microscopic image of the active tissue; the trained morphological feature and resolution restoration model adopts a recurrent adversarial generative network model.

[0029] By implementing steps S1 to S3 above, this embodiment can obtain stimulated Raman scattering microscopic images of active tissues with high signal-to-noise ratio and high resolution while ensuring the activity of the active tissue.

[0030] Currently, there is a lack of a solution that can simultaneously achieve high-speed, large-field-of-view, and high-resolution (high signal-to-noise ratio and high resolution) pathological-grade imaging while ensuring the viability of active tissues. This directly restricts the application of SRS microscopy in the field of active tissue detection and evaluation. To address this deficiency, this embodiment proposes a stimulated Raman scattering microscopy method suitable for active tissues. The following is a detailed introduction to this stimulated Raman scattering microscopy method suitable for active tissues.

[0031] (I) Constructing an ultra-high-speed, low-optical-damage stimulated Raman scattering microscopic imaging system to acquire initial stimulated Raman scattering microscopic images. In this embodiment, a high-speed initial stimulated Raman scattering microscopic image of the active tissue is obtained under low laser power and low exposure time. Specifically, the initial stimulated Raman scattering microscopic image obtained by performing stimulated Raman scattering microscopy on the active tissue is obtained. When performing stimulated Raman scattering microscopy on the active tissue to obtain the initial stimulated Raman scattering microscopic image, the laser power used is less than the preset power, the single pixel exposure time is less than the preset time, and the imaging speed is greater than the preset speed, which can ensure the activity of the active tissue.

[0032] In this embodiment, an ultra-high-speed confocal scanning imaging system based on a resonant scanning galvanometer and a displacement stage (which can be a high-precision displacement stage) is first constructed and integrated into the stimulated Raman scattering microscopy imaging optical path to obtain an ultra-high-speed, low-optical-damage stimulated Raman scattering microscopy imaging system. The stimulated Raman scattering microscopy system includes a laser (which can be a high-power narrowband femtosecond laser or other types of lasers), a stimulated Raman scattering microscopy imaging optical path for generating picosecond hyperspectral data (including a modulator for high-frequency modulation of the laser beam (which can be an electro-optic modulator or other modulation devices with high-frequency modulation capabilities), a hyperspectral modulation system for adjusting the Raman frequency shift (which can be a spectral focusing system based on a laser chirp-delay modulation system or other hyperspectral systems with hyperspectral modulation capabilities), a displacement stage for time-domain modulation of the laser pulse, and other optical components (such as lenses, dichroic mirrors, and mirrors) to form the complete optical path), a high-speed resonant scanning system (including a resonant scanning galvanometer (which can be a high-speed resonant scanning galvanometer), a scanning mirror, a tube mirror, and a displacement stage), a high numerical aperture confocal microscopy system (including a first mirror, an objective lens (which can be a microscope objective), a condenser, a second mirror, and a lens), and an image acquisition system (including a photodetector, a lock-in amplifier, and a host computer (such as a computer)).

[0033] At this time, as Figure 3 As shown, the laser includes a pump light laser and a Stokes light laser. The pump light passes through two lenses and a hyperspectral adjustment system that can adjust the target Raman frequency shift before entering the dichroic mirror. The Stokes light passes through a lens, a modulator, a lens, a hyperspectral adjustment system that can adjust the target Raman frequency shift, two mirrors located on a displacement stage for time-domain adjustment, and a mirror before entering the dichroic mirror. The laser output from the dichroic mirror passes through two lenses and a mirror before being output as laser light, and then enters the subsequent resonant scanning galvanometer.

[0034] At this time, as Figure 4As shown, when performing stimulated Raman scattering microscopy on active tissue, a stimulated Raman scattering microscopy system is used. The stimulated Raman scattering microscopy system includes: a laser, a stimulated Raman scattering microscopy optical path, a resonant scanning galvanometer, a scanning mirror, a tube mirror, a first reflecting mirror, an objective lens, a displacement stage, a condenser lens, a second reflecting mirror, a lens, a photodetector, a lock-in amplifier, and a host computer. The active tissue is located on the displacement stage.

[0035] The initial laser emitted by the laser is converted into a laser by the stimulated Raman scattering (SRS) microscopy imaging optical path. The laser (i.e., the laser emitted by the laser and the SRS microscopy imaging optical path) is incident on the active tissue located on the displacement stage through the resonant scanning galvanometer, scanning mirror, tube mirror, first reflecting mirror and objective lens, exciting a stimulated Raman scattering signal. The laser carrying the stimulated Raman scattering signal (which can be called the carrier laser at this time) is incident on the photodetector through the condenser mirror, second reflecting mirror and lens, generating an electrical signal. The electrical signal is processed by the lock-in amplifier and the host computer to obtain the initial stimulated Raman scattering microscopic image. The laser includes pump light and Stokes light. The resonant scanning galvanometer, scanning mirror, tube mirror and displacement stage are used to incident the laser at different positions of the active tissue to scan the active tissue.

[0036] It should be noted that during imaging, the fresh, viable tissue to be tested should be placed in physiological saline or PBS (Phosphate-Buffered Saline) environment. The system parameters of the stimulated Raman scattering microscopy system should be set, using a laser power below the photodamage threshold for excitation. For example, a femtosecond laser with a wavelength of 700nm-900nm can be used as the pump light, and a femtosecond laser with a wavelength of 1000nm-1200nm can be used as the Stokes light. The laser power received at the viable tissue should be controlled to be no higher than 60mW (i.e., the preset power, which needs to be below the photodamage threshold to avoid photodamage to the viable tissue). High-speed linear scanning is achieved through a resonant scanning galvanometer and scanning mirror, combined with stepping using a displacement stage, thereby achieving high-speed imaging with a large field of view. During scanning, the single-pixel exposure time should be set to no higher than 100ns (i.e., the preset time to avoid photodamage to the viable tissue), and the imaging speed should be no less than 0.1mm. 2 / s (i.e., the preset speed to avoid photodamage to active tissue). Thus, scanning of a large area of ​​active tissue is completed in a very short time, acquiring one or more original low-dose, low-signal-to-noise-ratio, low-resolution SRS images (i.e., initial stimulated Raman scattering microscopic images).

[0037] (ii) Signal-to-noise ratio enhancement based on diffusion model In this embodiment, the initial stimulated Raman scattering microscopic image is processed using a trained signal-to-noise ratio enhancement model to obtain a processed image. The trained signal-to-noise ratio enhancement model adopts a diffusion model.

[0038] This embodiment receives the initial stimulated Raman scattering microscopic image acquired by the stimulated Raman scattering microscopic imaging system and inputs it into a trained signal-to-noise ratio enhancement model based on the denoising diffusion probability model (DDPM).

[0039] The trained signal-to-noise ratio enhancement model adopts a diffusion model, which is used as a noise prediction network. It can adopt an improved U-Net2D architecture. In this case, the trained signal-to-noise ratio enhancement model (i.e., diffusion model) includes a time-step embedding unit, N downsampling blocks and N upsampling blocks. The output of the time-step embedding unit is connected to the input of each downsampling block and the input of each upsampling block. The output of the i-th downsampling block is connected to the input of the (i+1)-th downsampling block and the input of the (N-i+1)-th upsampling block. The output of the N-th downsampling block is connected to the input of the 1-th upsampling block. The output of the i-th upsampling block is connected to the input of the (i+1)-th upsampling block. i = 1, 2, ..., N-1, where N is the number of downsampling blocks or upsampling blocks, which can be user-defined.

[0040] The temporal step embedding unit encodes the current time step to obtain its embedding vector. The temporal step embedding unit comprises a sinusoidal position encoding unit and a multilayer perceptron connected in sequence. This embedding vector is directly injected into each downsampling block and each upsampling block. Specifically, the temporal step embedding unit encodes the time steps during the diffusion process. First, the sinusoidal position encoding unit maps the time step to a high-dimensional feature representation. Then, a nonlinear transformation is performed through the multilayer perceptron to obtain the embedding vector. This embedding vector is injected as conditional information into each layer of the U-Net network and fused with convolutional features after linear projection. This guides the U-Net network to learn the corresponding noise prediction function based on the current diffusion stage, improving the accuracy of noise estimation. The injection process is a mature existing technique and will not be elaborated upon here.

[0041] The downsampling block consists of a convolutional layer, a downsampling layer, a normalization layer, and an activation function layer connected in sequence. The upsampling block consists of an upsampling layer (whose input is the input feature map output by the previous upsampling block; for the first upsampling block, it is the input feature map output by the Nth downsampling block), a concatenation layer (whose input is the input feature map output by the downsampling block and the input feature map output by the upsampling layer), a convolutional layer, a normalization layer, and an activation function layer connected in sequence.

[0042] Preferably, in this embodiment, some downsampling blocks and some upsampling blocks can be designed to use attention mechanism layers. Specifically, the I1th to Nth downsampling blocks can be designed to use attention mechanism layers. In this case, the downsampling block includes a convolutional layer, a downsampling layer, a normalization layer, an activation function layer and an attention mechanism layer connected in sequence. The 1st to I2th upsampling blocks can be designed to use attention mechanism layers. In this case, the upsampling block includes an upsampling layer, a splicing layer, a convolutional layer, a normalization layer, an activation function layer and an attention mechanism layer connected in sequence. The values ​​of I1 and I2 can be customized by the user.

[0043] When N is 6, I1 is 5, and I2 is 2, as follows Figure 5As shown, the trained signal-to-noise ratio enhancement model specifically includes: (1) an input layer with an input size of 512 pixels × 512 pixels and 1 input channel (e.g., grayscale image) or 3 (multi-channel, e.g., color image). The output of the input layer is connected to the input of the first downsampling block; (2) a downsampling path, which is as follows: the first downsampling block (DownBlock2D, 128 channels, size 256), the second downsampling block (DownBlock2D, 128 channels, size 128), the third downsampling block (DownBlock2D, 256 channels, size 64), and the fourth downsampling block (DownBlock2D, 256 channels, size 64). ock2D, 256 channels, size 32), fifth downsampling block (AttnDownBlock2D, 512 channels, size 16, including attention mechanism layer), sixth downsampling block (AttnDownBlock2D, 512 channels, size 8, including attention mechanism layer), first downsampling block, second downsampling block, third downsampling block, fourth downsampling block, fifth downsampling block and sixth downsampling block are connected in sequence; (3) upsampling path, in sequence: first upsampling block (AttnUpBlock2D, 512 channels, size 16, including attention mechanism layer), second upsampling block (AttnUpBlock2D, 512 channels, size 32, including attention mechanism layer), third upsampling block (UpBlock2D, 256 channels, size 64), fourth upsampling block (UpBlock2D, 256 channels, size 128), fifth upsampling block (UpBlock2D, 128 channels, size 256), sixth upsampling block (UpBlock2D, 128 channels, size 512), the first upsampling block, second upsampling block, third upsampling block, fourth upsampling block, fifth upsampling block and sixth upsampling block are connected in sequence; downsampling at each level in the downsampling path. The output of the sample block is connected to the input of the corresponding level upsampling block in the upsampling path through jump connections to form a U-shaped symmetrical structure, that is, the first downsampling block is connected to the sixth upsampling block, the second downsampling block is connected to the fifth upsampling block, the third downsampling block is connected to the fourth upsampling block, the fourth downsampling block is connected to the third upsampling block, the fifth downsampling block is connected to the second upsampling block, and the sixth downsampling block is connected to the first upsampling block; (4) The time step embedding unit is connected to each downsampling block and each upsampling block to generate the embedding vector of the current time step and inject the embedding vector of the current time step into each downsampling block and each upsampling block.

[0044] In this embodiment, the initial stimulated Raman scattering (SRS) microscopic image can first be preprocessed using a data preprocessing module. This module includes a data augmentation unit and a tensor transformation unit. The data augmentation unit performs random horizontal flipping, random vertical flipping, and random cropping operations on the image data. The cropping size is preferably 512 pixels × 512 pixels. The tensor transformation unit is connected to the data augmentation unit and converts the enhanced image data into tensor format for subsequent processing. At this point, the initial SRS microscopic image is preprocessed to obtain a preprocessed image, which is then input into a trained signal-to-noise ratio (SNR) enhancement model. The data preprocessing includes data augmentation and tensor transformation; data augmentation includes random horizontal flipping, random vertical flipping, and random cropping operations.

[0045] During training, a noise scheduling module and a forward noise addition module are designed. The noise scheduling module adopts a DDPMScheduler (Denoising Diffusion Probability Model Scheduler) structure, storing a noise scheduling table including noise scheduling parameters for T noise addition time steps, where T is the number of noise addition time steps, preferably 1000. It also provides a noise addition interface, facilitating the subsequent addition of noise of corresponding intensity to the image data based on the noise scheduling parameters of the current time step. The forward noise addition module is electrically connected to the data preprocessing module and the noise scheduling module, receiving image data transmitted from the data preprocessing module. (i.e., training images), generating and Random Gaussian noise ε of the same dimension (i.e., the same size) receives noise scheduling parameters for the current time step t (t∈[0,T-1]) transmitted by the noise scheduling module. Through formula Generate a noisy image at the current time step t. The noise prediction network receives the noise image at the current time step t. Given the current time step t, through its downsampling and upsampling paths including the attention mechanism layer, combined with the skip connections and the embedding of the time step embedding vector, the added noise (i.e., prediction noise) ε_θ is accurately predicted. Through the above process, the prediction noise for each noisy time step can be predicted, and the prediction noise is gradually added to the training image in ascending order of the noisy time steps, resulting in a prediction noise image containing the prediction noise.

[0046] During training, a loss calculation and optimization module is designed. The loss function calculation unit within this module calculates the mean square error (MSE) between the predicted noise image (containing prediction noise) and the real low signal-to-noise ratio reference image (containing real noise, such as the initial stimulated Raman scattering micrograph) to obtain the total loss. The gradient backpropagation unit within the loss calculation and optimization module calculates the gradient of the total loss with respect to each network parameter in the noise prediction network. Based on the gradient of the total loss with respect to each network parameter in the noise prediction network, the parameter optimization unit within the loss calculation and optimization module uses the AdamW optimization algorithm (with a learning rate preferably 1×10⁻⁶) to optimize the algorithm. -5 Update the parameters of each network in the noise prediction network to update the noise prediction network.

[0047] Repeat the above steps of data preprocessing, noise addition, prediction, and optimization until the preset number of training rounds (preferably 1000 rounds) is reached.

[0048] During training, a model storage and output module is designed. The periodic storage unit of the model storage and output module saves the current network state, including model network parameters and optimizer state, every preset period (preferably 10 rounds). The training log recording unit of the model storage and output module records the training loss value (i.e., total loss) for each round and generates a loss curve (the curve of total loss changing with the number of rounds).

[0049] At this point, in this embodiment, before processing the initial stimulated Raman scattering microscopic image using the trained signal-to-noise ratio enhancement model to obtain the processed image, the stimulated Raman scattering microscopic imaging method applicable to active tissues in this embodiment further includes: training the initial signal-to-noise ratio enhancement model to obtain a trained signal-to-noise ratio enhancement model, specifically including: (1) Obtain the training image and take the first noisy time step as the current time step.

[0050] The training images can be noise-free stimulated Raman scattering microscopy images.

[0051] (2) Based on random Gaussian noise and the noise scheduling parameters of the current time step, the training image is denoised to obtain the denoised image (i.e., the noise image).

[0052] (3) Using the noisy image and the current time step as input, the prediction noise of the current time step is determined by the initial signal-to-noise ratio enhancement model. The initial signal-to-noise ratio enhancement model and the trained signal-to-noise ratio enhancement model have the same structure.

[0053] (4) Determine whether the current time step is the last time step with noise added, and obtain the determination result.

[0054] (5) If the judgment result is negative, the sum of the current time step and 1 is used as the current time step of the next iteration, and the step of "adding noise to the training image based on random Gaussian noise and the noise scheduling parameter of the current time step to obtain the noise-added image" is returned.

[0055] (6) If the judgment result is yes, then based on the predicted noise of all noise-adding time steps, determine the predicted noise image, calculate the error between the predicted noise image and the real noise image, obtain the total loss, and use the total loss to update the initial signal-to-noise ratio enhancement model to obtain the updated model. Determine whether the first iteration termination condition has been met. If yes, then use the updated model as the trained signal-to-noise ratio enhancement model. If no, then use the updated model as the initial signal-to-noise ratio enhancement model for the next iteration and return to the "obtain training image" step.

[0056] The termination condition for the first iteration can be reaching the first maximum number of iterations (e.g., 1000 rounds).

[0057] Before performing the above training process, this embodiment can design a pre-trained diffusion model loading unit and an optimizer state recovery unit. The pre-trained diffusion model loading unit reads the pre-trained weight file through the model parameter storage interface. The state dictionary loader is electrically connected to the U-Net2D model structure and loads the pre-trained weight file into the noise prediction network. The optimizer state recovery unit loads the optimizer parameters from the pre-training stage to maintain the continuity of the training state, thereby pre-training the noise prediction network before performing the above training process.

[0058] In the inference application, an SRS image with a high signal-to-noise ratio (i.e., the processed image) is generated and output through an iterative denoising process, aiming to restore the image details and contrast lost due to low-power excitation.

[0059] In this embodiment, the trained signal-to-noise ratio enhancement model is used to process the initial stimulated Raman scattering microscopic image to obtain the processed image, specifically including: (1) The initial stimulated Raman scattering microscopy image is used as the noisy image, and the last denoised time step is used as the current time step.

[0060] (2) Using the noisy image and the current time step as input, the predicted noise of the current time step is determined by the trained signal-to-noise ratio enhancement model.

[0061] (3) Denoise the noisy image based on the predicted noise at the current time step to obtain the denoised image.

[0062] Existing denoising formulas can be used to denoise noisy images, which will not be elaborated here.

[0063] (4) Determine whether the current time step is the first denoising time step. If yes, use the denoised image as the processed image. If no, use the denoised image as the noise image for the next iteration, and use the difference between the current time step and 1 as the current time step for the next iteration. Return to the step of "using the noise image and the current time step as input, and using the trained signal-to-noise ratio enhancement model to determine the predicted noise of the current time step".

[0064] (III) Morphological feature and resolution restoration based on CycleGAN model In this embodiment, the trained morphological feature and resolution restoration model is used to process the processed image to obtain stimulated Raman scattering microscopic images of active tissue. The trained morphological feature and resolution restoration model adopts a recurrent adversarial generative network model.

[0065] This embodiment receives the processed image generated by the diffusion model and inputs it into a trained CycleGAN-based morphological feature and resolution restoration model. This trained morphological feature and resolution restoration model has learned the mapping relationship between a large number of high-speed, low-quality SRS images and standard high-quality SRS images, mastering the ability to restore the morphological features and resolution of the image. It processes the input processed image, correcting motion artifacts or deformations that may be introduced by the high-speed scanning of the resonant scanning galvanometer, while simultaneously sharpening image features and improving its resolution, making it visually and structurally closer to images of traditional pathological staining (such as HE staining) (i.e., standard high-quality SRS images). This ensures the morphological features and resolution stability of the final output stimulated Raman scattering microscopy image of active tissue, meeting the requirements of pathological diagnosis. The final output is an SRS image (i.e., stimulated Raman scattering microscopy image of active tissue) that simultaneously possesses a high signal-to-noise ratio, high resolution, clear tissue structure, and meets the requirements of pathological diagnosis.

[0066] The recurrent adversarial generative network model includes a first generator G. A Second generator G B First discriminator D A Second discriminator D B The specific structure and parameter configuration of the generator and discriminator are as follows: (1) Generator structure and parameter configuration First generator G A Second generator G B They have the same structure, all adopting a symmetrical structure, and the generators include: 1) Encoder: A sequence of convolutional layers is used. The sequence of convolutional layers includes multiple convolutional blocks connected in sequence. Each convolutional block includes a convolutional layer, a batch normalization layer and a ReLU activation function layer connected in sequence. The convolutional layer sequence is used as a feature extraction unit. The number of input channels is 1 (grayscale image) and the number of output channels is ngf (preferably 32). The feature dimension is increased layer by layer. 2) Residual Module: A residual block sequence is used, which includes num_resnet residual blocks (preferably 6). Each residual block adopts a residual connection structure, including a first convolutional layer (keeping the number of input and output channels consistent), a first instance normalization layer, a ReLU activation function layer, a second convolutional layer, a second instance normalization layer, and a concatenation layer (implementing skip connections, directly adding the input to the output, and achieving identity mapping). 3) Decoder: It adopts a deconvolution layer sequence, which includes multiple deconvolution blocks connected in sequence and an output layer. Each deconvolution block includes a deconvolution layer, a batch normalization layer and a ReLU activation function layer connected in sequence to gradually restore the image size. The output layer adopts the Tanh activation function and the output range is [-1, 1], which matches the range of the input image data.

[0067] This embodiment designs a weight initialization unit, which is configured to use a normal distribution for initialization. The generator is designed with initial values ​​that conform to a normal distribution, with the mean preferably being 0.0 and the standard deviation preferably being 0.02, to ensure stability in the early stages of training.

[0068] (2) Discriminator structure and parameter configuration First discriminator D A Second discriminator D B They have the same structure, both using the PatchGAN architecture, and the discriminator includes: 1) Convolutional layer sequence: includes multiple convolutional blocks connected in sequence. Each convolutional block includes a convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer (preferably with a slope of 0.2) connected in sequence. The first convolutional block has 1 input channel and ndf (preferably 64) output channels. The number of output channels of the convolutional layers in subsequent convolutional blocks increases by doubling layer by layer until a preset maximum value is reached (which can be customized by the user). After reaching the preset maximum value, the number of output channels of the convolutional layers in subsequent convolutional blocks is also the preset maximum value. 2) Output layer: The output of the output layer is the probability that the input image is a real image.

[0069] This embodiment designs a weight initialization unit that uses the same normal distribution initialization strategy as the generator. It is configured to use normal distribution initialization to design initial values ​​for the discriminator that conform to a normal distribution, with the mean preferably being 0.0 and the standard deviation preferably being 0.02, to ensure stability in the early stages of training.

[0070] In this embodiment, the first generator and the second generator have the same structure. The first generator includes an encoder, a residual module, and a decoder connected in sequence. The encoder includes multiple convolutional blocks connected in sequence. The residual module includes multiple residual blocks connected in sequence. Each residual block includes a first convolutional layer, a first instance normalization layer, a ReLU activation function layer, a second convolutional layer, a second instance normalization layer, and a splicing layer connected in sequence. The input of the splicing layer is also connected to the input of the first convolutional layer. The decoder includes multiple deconvolutional blocks and a first output layer connected in sequence.

[0071] The first discriminator and the second discriminator have the same structure. The first discriminator includes multiple convolutional blocks connected in sequence and a second output layer.

[0072] During the training process, this embodiment designs a first image domain acquisition module A (resonance galvanometer image domain), a second image domain acquisition module B (ordinary galvanometer image domain), a loss calculation module, and a model training and storage module.

[0073] The first image domain acquisition module A acquires the first training image, which is an image acquired based on the activity of the active tissue without considering the image signal-to-noise ratio and sharpness. This is the same as the initial stimulated Raman scattering microscopy image acquisition method, i.e., a high-speed, low-quality SRS image. The second image domain acquisition module B acquires the second training image, which is an image acquired based on the image signal-to-noise ratio and sharpness without considering the activity of the active tissue, i.e., a standard, high-quality SRS image.

[0074] like Figure 6 As shown, the training process flow is as follows: (1) Data loading and preprocessing: The first image domain acquisition module A acquires multiple first training images to form a first image set, and the second image domain acquisition module B acquires multiple second training images to form a second image set. The first image set and the second image set do not need to be paired and can be acquired from different active tissues. The image data in the first image set and the second image set are read in batch processing by the data loader. The batch size is preferably 2. Eight parallel working threads are used. The training set and the validation set are divided according to a preset ratio. When reading 50 samples (including one first training image and one second training image), preferably, the first 40 samples are used as the training set and the last 10 samples are used as the validation set.

[0075] (2) Forward propagation of generator: First generator G A Receive the first training image real A Generate the first generated image (fake). B =G A (real) A ), second generator G BReceive the second training image real B Generate a second fake image A =G B (real) B Simultaneously, a loop reconstruction is performed, with the first generator G... A Receive the second generated image (fake) A Generate the first reconstructed image recon B =G A (fake) A ), second generator G B Receive the first generated image (fake) B Generate a second reconstructed image. A =G B (fake) B ).

[0076] (3) Discriminator discrimination: First discriminator D A Receive the first training image real A Second generated image fake A Output the first discrimination result, which includes the first training image real. A The probability of a real image and the probability of a fake image generated in the second generation. A The probability of a real image, where the real image refers to the first training image. A Second discriminator D B Receive the second training image real B And the first generated image fake B Output the second discrimination result, which includes the second training image real. B The probability of a real image and the probability of the first generated image being fake. B The probability of a real image, where the real image refers to the second training image. B .

[0077] (4) Loss calculation and gradient update 1) Loss calculation: Completed by the loss calculation module. a) Generator loss: The adversarial loss is calculated using the mean squared error between the first discrimination result and the all-1 label and the mean squared error between the second discrimination result and the all-1 label. The cycle consistency loss is calculated using the L1 distance between the first training image and the second reconstructed image and the L1 distance between the second training image and the first reconstructed image. The cycle consistency loss is multiplied by the weight coefficient λ (preferably 10) and then added to the adversarial loss to obtain the generator loss.

[0078] The formula for calculating the generator loss is: LG =L GA +L GB +λ A ×L cycleA +λ B ×L cycleB ; Among them, L G For generator loss; L GA For the first confrontation loss; L GB For the second type of adversarial loss; λ A The first weighting coefficient is 10, which is preferred; L cycleA λ is the consistency loss for the first cycle; B The second weighting coefficient is preferably 10; L cycleB This represents the consistency loss during the second cycle.

[0079] First confrontation loss L GA The mean squared error between the second discrimination result and the all-1 label is calculated using the following formula: ; Where E is the expectation; xA is the first training image; pA is the distribution of the first training image; D B (G A (xA)) is the second discriminator D B Output the first generated image G A (xA) represents the probability of the real image, which belongs to the second discrimination result.

[0080] Second confrontation loss L GB The mean squared error between the first discrimination result and the all-1 label is calculated using the following formula: ; Where xB is the second training image; pB is the distribution of the second training image; D A (G B (xB)) is the first discriminator D A Output the second generated image G B (xB) represents the probability of the real image, which belongs to the first discrimination result.

[0081] First cycle consistency loss L cycleA The L1 distance between the first training image and the second reconstructed image is calculated using the following formula: ; Among them, G B (G A (xA)) is the second reconstructed image, G A (xA) is the first generated image.

[0082] Second cycle consistency loss LcycleB The L1 distance between the second training image and the first reconstructed image is calculated using the following formula: ; Among them, G A (G B (xB)) is the first reconstructed image, G B (xB) is the second generated image.

[0083] Cyclic consistency loss can ensure that the image retains its original content features after bidirectional mapping.

[0084] b) Discriminator Loss: The first discriminator loss is calculated using the mean square error between the discrimination result of the first discriminator on the first training image and the all-1 label, and the mean square error between the discrimination result of the first discriminator on the second generated image and the all-0 label. The second discriminator loss is calculated using the mean square error between the discrimination result of the second discriminator on the second training image and the all-1 label, and the mean square error between the discrimination result of the second discriminator on the first generated image and the all-0 label. The discriminator loss is obtained by summing the first discriminator loss and the second discriminator loss.

[0085] The formula for calculating the discriminator loss is: L D =L DA +L DB ; Among them, L D For discriminator loss; L DA The loss of the first discriminator; L DB This is the loss of the second discriminator.

[0086] The formula for calculating the loss of the first discriminator is: ; Among them, D A (xA) represents the first discriminator D. A Output the probability that the first training image is a real image, belonging to the first discrimination result; D A (G B (xB)) is the first discriminator D A Output the second generated image G B (xB) represents the probability of the real image, which belongs to the first discrimination result.

[0087] The formula for calculating the loss of the second discriminator is: ; Among them, D B (xB) represents the second discriminator D. B Output the probability that the second training image is a real image, belonging to the second discrimination result; DB (G A (xA)) is the second discriminator D B Output the first generated image G A (xA) represents the probability of the real image, which belongs to the second discrimination result.

[0088] 2) Gradient update: Completed by the model training and storage module. Update the generator parameters and discriminator parameters using the generator loss and discriminator loss respectively, until the second maximum number of iterations is reached, and then save the generator parameters and discriminator parameters at this point.

[0089] At this point, in this embodiment, before processing the processed image using the trained morphological features and resolution restoration model to obtain a stimulated Raman scattering microscopic image of the active tissue, the stimulated Raman scattering microscopic imaging method for active tissue in this embodiment further includes: training the initial morphological features and resolution restoration model to obtain a trained morphological features and resolution restoration model, specifically including: (1) Acquire the first training image and the second training image. Both the first training image and the second training image are images obtained by stimulated Raman scattering microscopy of active tissue. When obtaining the first training image, the laser power used is less than the preset power, the single pixel exposure time is less than the preset time, and the imaging speed is greater than the preset speed. When obtaining the second training image, the laser power used is greater than the preset power, the single pixel exposure time is greater than the preset time, and the imaging speed is less than the preset speed.

[0090] (2) Using the first training image as input, the first generator is used to obtain the first generated image, and using the first generated image as input, the second generator is used to obtain the first reconstructed image.

[0091] (3) Using the second training image as input, the second generator is used to obtain the second generated image, and the second generated image is used as input to obtain the second reconstructed image using the first generator.

[0092] (4) Use the first discriminator to discriminate the first training image and the second generated image to obtain the first discrimination result.

[0093] (5) Use the second discriminator to discriminate between the second training image and the first generated image to obtain the second discrimination result.

[0094] (6) Based on the first discrimination result and the second discrimination result, the adversarial loss and the discriminator loss are calculated respectively. Based on the first training image, the first reconstructed image, the second training image and the second reconstructed image, the cycle consistency loss is calculated. The adversarial loss and the cycle consistency loss are weighted and summed to obtain the generator loss.

[0095] (7) Update the first generator and the second generator using the generator loss to obtain the first updated generator and the second updated generator. Update the first discriminator and the second discriminator using the discriminator loss to obtain the first updated discriminator and the second updated discriminator.

[0096] (8) Determine whether the second iteration termination condition has been met. If yes, use the first updated generator as the trained morphological feature and resolution recovery model. If no, use the first updated generator, the second updated generator, the first updated discriminator and the second updated discriminator as the first generator, the second generator, the first discriminator and the second discriminator for the next iteration, respectively, and return to the step of "obtaining the first training image and the second training image".

[0097] The second iteration termination condition can be reaching the second maximum number of iterations.

[0098] During the inference application, the optimal weight file of the first generator is loaded from the storage path to load the trained morphological feature and resolution restoration model. The processed image to be converted is input into the trained morphological feature and resolution restoration model, and the corresponding ordinary galvanometer style image (i.e. stimulated Raman scattering microscopy image of active tissue) is output, which retains the anatomical structure information of the original image and has the uniform sampling characteristics of ordinary galvanometer.

[0099] This embodiment provides an AI-assisted ultra-high-speed stimulated Raman scattering (SRS) microscopy imaging method suitable for active tissues, aiming to solve the problem that related technologies struggle to balance imaging speed, image signal-to-noise ratio (SNR), and tissue activity when imaging active tissues. First, an ultra-high-speed confocal scanning imaging system based on a resonant scanning galvanometer and a displacement stage is constructed, combined with the SRS microscopy imaging optical path, to achieve large field-of-view, high-speed image data acquisition. During this process, laser power several times lower than that of related methods is used for low-optical-damage imaging to protect the activity of active tissues. Then, an AI model based on a diffusion model and CycleGAN is used to process the acquired raw low SNR image (i.e., the initial SRS microscopy image). The diffusion model significantly improves the image's SNR, while the CycleGAN model restores and ensures the morphological features and resolution stability of the image under large field-of-view, high-speed scanning. By employing the aforementioned technical approach of "low-damage acquisition + AI enhancement," it is possible to obtain stimulated Raman scattering microscopic images with high signal-to-noise ratio, high definition, and meeting the requirements of pathological diagnosis, while ensuring the viability of the active tissue. This is of great significance for detection scenarios that require maintaining the subsequent viability of the active tissue, such as safety assessment before cryopreservation of active tissue.

[0100] Compared with related technologies, this embodiment has the following beneficial effects: (1) Extremely low photodamage and good preservation of tissue viability: By controlling the laser power to an extremely low level (e.g., a total of 60mW) and combining it with ultra-high-speed scanning technology (e.g., completing a 2mm×2mm scan in 40 seconds), the photodamage to viable tissues during the imaging process is greatly reduced. Experimental results show that the viable oocytes in fresh ovarian tissue scanned by this method are still present, and viable oocytes can still be obtained 3 weeks after transplantation back into the body, proving that the tissue viability and function are well preserved.

[0101] (2) Excellent image quality: The original low signal-to-noise ratio image is processed by AI model (diffusion model + CycleGAN), which can significantly improve the image quality. The signal-to-noise ratio of the final output image (i.e. stimulated Raman scattering microscopy image of active tissue) is much higher than that of the original low signal-to-noise ratio image. The structural similarity index (SSIM) with the real tissue structure is not less than 0.7, and the peak signal-to-noise ratio (PSNR) can be increased to more than 15 times that of the original low signal-to-noise ratio image. The image clarity is sufficient to identify the fine structure of the tissue. The consistency rate with the HE staining image of histopathology is high, which can meet the requirements of pathological diagnosis.

[0102] Specific effects are as follows Figure 7 As shown in a and b, Figure 7 The first three images in section 'a' show the original low signal-to-noise ratio images under different laser powers. Figure 7 The last image in section 'a' (i.e., with a laser power of 100-200mw) shows a standard high-quality SRS image and can be used as a reference image. Figure 7 b in the text shows the... Figure 7 The first three images of 'a' in the model are processed by the AI ​​model and the resulting output images are similar to standard high-quality SRS images, demonstrating that good results have been achieved.

[0103] (3) Ultra-high-speed large field-of-view imaging capability: Combining resonant scanning galvanometer and displacement stage, an image depth of not less than 0.1 mm is achieved. 2 The imaging speed is [missing information - likely a speed measurement value] / s. Compared to traditional SRS microscopy, the complete imaging time for a sample of the same size (e.g., 4mm × 4mm) is reduced from approximately 1.5 hours to approximately 40 seconds, improving imaging efficiency by about 90 times. This significantly shortens the time of ex vivo exposure of active tissue and reduces the negative impact of the environment on the activity of active tissue. Meanwhile, traditional SRS microscopy requires approximately 1.5 hours to completely image a 4mm × 4mm sample, and the laser power and single-pixel exposure time used result in approximately 120 times the light dose received per unit area of ​​tissue compared to this method, demonstrating that this method can reduce photodamage to active tissue.

[0104] (4) Breakthrough in related technical bottlenecks: For the first time, a feasible solution is provided that can achieve pathological-level imaging without damaging the living tissue. This provides strong technical support for solving key medical issues such as tumor cell screening before cryopreservation and transplantation of ovarian tissue, and is expected to break through the safety limitations in this field.

[0105] This application also provides an application scenario in which the above-described stimulated Raman scattering microscopy imaging method for active tissues is applied. Specifically, the stimulated Raman scattering microscopy imaging method for active tissues provided in this embodiment can be applied in an imaging scenario. The imaging scenario includes an imaging stage and a display stage. The imaging stage is used to generate stimulated Raman scattering microscopic images of active tissues, and the display stage is used to display the stimulated Raman scattering microscopic images of active tissues to the user. The stimulated Raman scattering microscopy imaging method for active tissues provided in this embodiment belongs to the imaging stage.

[0106] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 8 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a stimulated Raman scattering microscopy imaging method suitable for active tissues.

[0107] Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0108] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described stimulated Raman scattering microscopy imaging method for active tissues.

[0109] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the above-described stimulated Raman scattering microscopy method for active tissues.

[0110] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described stimulated Raman scattering microscopy method for active tissues.

[0111] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations and be authorized by the owner of the corresponding device.

[0112] Any references to memory, databases, or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may take many forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0113] The databases involved in the various embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases.

[0114] The processors involved in the various embodiments provided in this application may be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited thereto.

[0115] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0116] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A stimulated Raman scattering microscopy imaging method suitable for living tissues, characterized in that, include: Acquire initial stimulated Raman scattering microscopic images of active tissues; When obtaining the initial stimulated Raman scattering microscopic image, the laser power used is less than the preset power, the single pixel exposure time is less than the preset time, and the imaging speed is greater than the preset speed. The initial stimulated Raman scattering microscopic image is processed using a trained signal-to-noise ratio enhancement model to obtain a processed image; the trained signal-to-noise ratio enhancement model adopts a diffusion model. The trained morphological feature and resolution restoration model is used to process the processed image to obtain stimulated Raman scattering microscopic images of active tissue; the trained morphological feature and resolution restoration model adopts a recurrent adversarial generative network model.

2. The stimulated Raman scattering microscopy imaging method for living tissues according to claim 1, characterized in that, When performing stimulated Raman scattering microscopy on active tissue, a stimulated Raman scattering microscopy system is used. The stimulated Raman scattering microscopy system includes: a laser, a stimulated Raman scattering microscopy optical path, a resonant scanning galvanometer, a scanning mirror, a tube mirror, a first reflecting mirror, an objective lens, a displacement stage, a condenser lens, a second reflecting mirror, a lens, a photodetector, a lock-in amplifier, and a host computer. The active tissue is located on the displacement stage. The laser emitted by the laser and the stimulated Raman scattering (SRS) microscopic imaging optical path is incident on the active tissue located on the displacement stage via the resonant scanning galvanometer, the scanning mirror, the tube mirror, the first reflecting mirror, and the objective lens, exciting a stimulated Raman scattering signal. The laser carrying the stimulated Raman scattering signal is incident on the photodetector via the condenser mirror, the second reflecting mirror, and the lens, generating an electrical signal. The electrical signal is processed by the lock-in amplifier and the host computer to obtain an initial stimulated Raman scattering microscopic image. The laser includes pump light and Stokes light. The resonant scanning galvanometer, the scanning mirror, the tube mirror, and the displacement stage are used to incident the laser at different positions on the active tissue to scan the active tissue.

3. The stimulated Raman scattering microscopy imaging method for living tissues according to claim 1, characterized in that, The trained signal-to-noise ratio enhancement model includes a time-step embedding unit, N downsampling blocks, and N upsampling blocks. The output of the time-step embedding unit is connected to the input of each downsampling block and the input of each upsampling block. The output of the i-th downsampling block is connected to the input of the (i+1)-th downsampling block and the input of the (N-i+1)-th upsampling block. The output of the N-th downsampling block is connected to the input of the 1-th upsampling block. The output of the i-th upsampling block is connected to the input of the (i+1)-th upsampling block. i = 1, 2, ..., N-1. The time step embedding unit is used to encode the current time step to obtain the embedding vector of the current time step.

4. The stimulated Raman scattering microscopy imaging method for living tissues according to claim 3, characterized in that, The initial stimulated Raman scattering microscopic image is processed using a trained signal-to-noise ratio enhancement model to obtain a processed image, specifically including: The initial stimulated Raman scattering micrograph is used as the noisy image, and the last denoised time step is used as the current time step; Using the noisy image and the current time step as input, the predicted noise for the current time step is determined using a trained signal-to-noise ratio enhancement model; The noisy image is denoised based on the predicted noise at the current time step to obtain a denoised image; Determine whether the current time step is the first denoising time step; if yes, use the denoised image as the processed image; if no, use the denoised image as the noise image for the next iteration, use the difference between the current time step and 1 as the current time step for the next iteration, and return to the step of "using the noise image and the current time step as input, and using the trained signal-to-noise ratio enhancement model to determine the predicted noise of the current time step".

5. The stimulated Raman scattering microscopy imaging method for living tissues according to claim 3, characterized in that, Before processing the initial stimulated Raman scattering microscopic image using the trained signal-to-noise ratio (SNR) enhancement model to obtain the processed image, the method further includes: training the initial SNR enhancement model to obtain a trained SNR enhancement model, specifically including: Acquire the training image and use the first noisy time step as the current time step; Based on random Gaussian noise and the noise scheduling parameters of the current time step, the training image is noise-added to obtain a noisy image; Using the noisy image and the current time step as input, the prediction noise for the current time step is determined using an initial signal-to-noise ratio (SNR) enhancement model; the initial SNR enhancement model and the trained SNR enhancement model have the same structure. Determine whether the current time step is the last noisy time step, and obtain the determination result; If the judgment result is negative, the sum of the current time step and 1 is used as the current time step of the next iteration, and the step of "adding noise to the training image based on random Gaussian noise and the noise scheduling parameter of the current time step to obtain the noisy image" is returned. If the judgment result is yes, then based on the predicted noise of all noise-adding time steps, determine the predicted noise image, calculate the error between the predicted noise image and the real noise image, obtain the total loss, and use the total loss to update the initial signal-to-noise ratio enhancement model to obtain the updated model; determine whether the first iteration termination condition has been met; if yes, then use the updated model as the trained signal-to-noise ratio enhancement model; if no, then use the updated model as the initial signal-to-noise ratio enhancement model for the next iteration, and return to the "obtain training image" step.

6. The stimulated Raman scattering microscopy imaging method for living tissues according to claim 1, characterized in that, The recurrent adversarial generative network model includes a first generator, a second generator, a first discriminator, and a second discriminator. Before processing the processed image using the trained morphological feature and resolution restoration model to obtain a stimulated Raman scattering micrograph of the active tissue, the model further includes: training the initial morphological feature and resolution restoration model to obtain a trained morphological feature and resolution restoration model, specifically including: Acquire a first training image and a second training image; both the first training image and the second training image are images obtained by stimulated Raman scattering microscopy imaging of active tissue. When obtaining the first training image, the laser power used is less than a preset power, the single pixel exposure time is less than a preset time, and the imaging speed is greater than a preset speed. When obtaining the second training image, the laser power used is greater than a preset power, the single pixel exposure time is greater than a preset time, and the imaging speed is less than a preset speed. Using the first training image as input, the first generator is used to obtain the first generated image, and using the first generated image as input, the second generator is used to obtain the first reconstructed image. Using the second training image as input, the second generator is used to obtain the second generated image, and using the second generated image as input, the first generator is used to obtain the second reconstructed image. The first discriminator is used to distinguish between the first training image and the second generated image to obtain a first discrimination result; The second discriminator is used to distinguish between the second training image and the first generated image to obtain a second discrimination result; Based on the first discrimination result and the second discrimination result, the adversarial loss and the discriminator loss are calculated respectively. Based on the first training image, the first reconstructed image, the second training image and the second reconstructed image, the cycle consistency loss is calculated. The generator loss is obtained by weighted summation of the adversarial loss and the cycle consistency loss. The generator loss is used to update the first generator and the second generator to obtain a first updated generator and a second updated generator. The discriminator loss is used to update the first discriminator and the second discriminator to obtain a first updated discriminator and a second updated discriminator. Determine whether the second iteration termination condition has been met; if yes, use the first updated generator as the trained morphological feature and resolution recovery model; if no, use the first updated generator, the second updated generator, the first updated discriminator, and the second updated discriminator as the first generator, the second generator, the first discriminator, and the second discriminator for the next iteration, respectively, and return to the step of "obtaining the first training image and the second training image".

7. The stimulated Raman scattering microscopy imaging method for living tissues according to claim 6, characterized in that, The first generator and the second generator have the same structure. The first generator includes an encoder, a residual module and a decoder connected in sequence. The encoder includes multiple convolutional blocks connected in sequence. The residual module includes multiple residual blocks connected in sequence. Each residual block includes a first convolutional layer, a first instance normalization layer, a ReLU activation function layer, a second convolutional layer, a second instance normalization layer and a splicing layer connected in sequence. The input of the splicing layer is also connected to the input of the first convolutional layer. The decoder includes multiple deconvolutional blocks and a first output layer connected in sequence. The first discriminator and the second discriminator have the same structure; the first discriminator includes multiple convolutional blocks connected in sequence and a second output layer.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the stimulated Raman scattering microscopy imaging method for active tissues as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the stimulated Raman scattering microscopy method for active tissues as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the stimulated Raman scattering microscopy method for active tissues as described in any one of claims 1-7.

Images