Rapid hyperspectral light field Raman body imaging system and method

Abstract

The invention relates to a rapid hyperspectral light field Raman body imaging system and method, and the system employs wide-field imaging, combines a multi-dimensional information coding unit with a CCD detector, carries out the integrated coding of the panoramic depth information of a sample, and employs an adjustable spectrum filtering sub-module to achieve the integrated coding of the panoramic depth information of the sample. And high-quality acquisition of three-dimensional space and Raman hyperspectral information under the wavelength required by the sample can be realized only by single exposure. According to the imaging method, a light field Raman body imaging algorithm is provided, data priori of different stages in the light field imaging process is constructed respectively, and the training processes of light field denoising, a supramolecular network and a three-dimensional reconstruction sub-network are guided in sequence, so that rapid and high-fidelity isotropic resolution three-dimensional reconstruction of a low-signal-to-noise-ratio light field image is realized. According to the method, a multi-dimensional image rapid and high-resolution three-dimensional visualization processing technology system integrating traditional optics, information optics and computational optics is constructed, the imaging efficiency and precision are remarkably improved, and efficient three-dimensional information capture and high-quality reconstruction are achieved.

Description

Technical Field

[0001] This invention belongs to the field of microscopic imaging technology, specifically relating to a rapid hyperspectral light field Raman volume imaging system and method. Background Technology

[0002] Raman 3D imaging technology is based on the Raman scattering effect and generates images by integrating spectral and spatial information. It can obtain chemical composition information of samples non-destructively without fluorescent labeling and is widely used in materials science, biomedicine, environmental monitoring and other fields.

[0003] Currently, traditional Raman 3D imaging techniques are mainly achieved through confocal microscopy, light sheet microscopy, and projection tomography. Confocal microscopy effectively suppresses stray light interference by selectively focusing on a specific focal plane of the sample, thus achieving high signal-to-noise ratio tomographic imaging. However, when the sample thickness exceeds several hundred micrometers, the efficiency of 3D imaging decreases significantly due to the geometrical increase in the number of scanning layers and the need for massive data processing. Light sheet microscopy employs an excitation-detection orthogonal optical path design, using a thin-layer laser beam (typically 1~10 μm thick) to excite the sample from the side and collect Raman signals in the vertical direction. Layer-by-layer imaging is achieved through axial scanning light sheets or precision displacement sample stages. However, its axial resolution is limited by the optical diffraction limit (typically >2 μm), and the imaging quality deteriorates sharply due to light scattering effects when imaging thick tissues. Raman projection tomography is based on the Radon transform principle and inverts three-dimensional structures through multi-angle projection data. However, its imaging speed is limited by the accuracy of mechanical rotation and it cannot be applied to the observation of non-fixed living samples. It is difficult to meet the strict requirements of life sciences for in-situ, dynamic, subcellular three-dimensional chemical imaging of complex biological systems.

[0004] Light field imaging technology has opened up new avenues for Raman 3D imaging visualization. It allows for the simultaneous acquisition of spatial light field information (including multi-angle and multi-depth information) from a single exposure, significantly improving imaging speed and data acquisition efficiency, and providing more comprehensive and accurate foundational data for subsequent 3D reconstruction. However, compared to traditional Raman imaging, light field imaging methods suffer from insufficient pixel sampling and limited sub-numerical aperture when simultaneously capturing images from different light field perspectives. Furthermore, during 3D reconstruction, traditional methods suffer from slow scanning speeds, low resolution, and artifacts, leading to a significant decrease in the accuracy and quality of the resulting 3D model. This prevents the model from accurately reflecting the true 3D structure of the sample, impacting subsequent analysis and judgment. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention provides a rapid hyperspectral Raman imaging system and method. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a rapid hyperspectral light field Raman volume imaging system, comprising: an excitation source module, a signal acquisition module, and a control and calculation module. The excitation source module provides a high-quality, low-noise, and stable uniform planar illumination source for the light field imaging system. The signal acquisition module, located after the excitation source module, illuminates the sample with the uniform planar illumination source to excite the sample to generate a Raman scattering signal. After filtering out Rayleigh scattering light from the Raman scattering signal, it selects and receives the Raman scattering signal of the desired wavelength through an adjustable spectral filtering submodule. The signal acquisition module also encodes information across the entire depth range of the sample using a multi-dimensional information encoding unit, transmits and acquires the encoded information to obtain a light field image. The control and calculation module is electrically connected to both the excitation source module and the signal acquisition module, and performs light field reconstruction on the light field image based on a light field reconstruction algorithm to obtain a three-dimensional Raman scattering signal image of the sample.

[0006] In one embodiment of the present invention, the excitation source module includes a continuous-wave laser and a light source generation submodule; the continuous-wave laser is used to generate continuous-wave laser light; the light source generation submodule is disposed after the continuous-wave laser and includes a 4F system composed of two conjugate lenses, a lens, a dichroic mirror, and an objective lens arranged sequentially along the optical path, used to uniformly illuminate the light source according to the output of the continuous-wave laser; the laser light output by the continuous-wave laser is expanded and collimated by the 4F system, reflected and focused by the lens and the dichroic mirror onto the rear pupil plane of the objective lens, and then generated into a uniformly propagating beam by the objective lens.

[0007] In one embodiment of the present invention, the signal acquisition module includes: an objective lens, a dichroic mirror, a notch filter, an adjustable spectral filtering submodule, a lens, a multi-dimensional information encoding unit, a relay system, and a CCD detector arranged sequentially along the optical path. The objective lens and the lens are used together to collect the Raman scattering signal excited by the sample irradiated by the excitation source module. The dichroic mirror is used to accurately separate the laser excitation light and the Raman scattering signal light emitted by the sample according to the wavelength of the light, achieving efficient acquisition of the Raman scattering signal. The notch filter is used to filter out Rayleigh scattering light from the Raman scattering signal. The adjustable spectral filtering submodule includes an adjustable long-pass filter, an adjustable short-pass filter, and an electric rotary stage for selecting and receiving Raman scattering signals of the desired wavelength. The cutoff frequencies of the adjustable long-pass filter and the adjustable short-pass filter change with the angle between the incident light and the filter. The system dynamically selects the desired wavelength of Raman scattering signal to remove background noise interference. The multi-dimensional information encoding unit encodes information across the entire sample depth range, specifically the selected wavelength of Raman scattering signal. The encoded information across the entire sample depth range is transmitted to the CCD detector via the relay system to obtain a light field image. The excitation source module illuminates the sample to generate a Raman scattering signal. The objective lens collects the Raman scattering signal, which is then filtered by a dichroic mirror and a notch filter to remove the laser excitation light. The tunable spectral filtering submodule then selects the desired wavelength of Raman scattering signal. The desired wavelength Raman scattering signal is focused onto the multi-dimensional information encoding unit by a lens, thus encoding the signal across the entire sample depth range. Finally, the encoded information is transmitted to the CCD detector via the relay system.

[0008] In one embodiment of the present invention, the control and calculation module includes: a computer control unit and a computer processing unit, wherein the computer control unit controls the continuous wave laser, the tunable spectral filter submodule and the CCD detector respectively; The computer processing unit includes: a light field super-resolution and denoising functional block, a light field 3D reconstruction functional block, and an isotropic resolution recovery functional block. The light field super-resolution and denoising functional block is used to perform denoising and super-resolution processing on the light field image to obtain a high signal-to-noise ratio light field image. The light field 3D reconstruction functional block is used to perform 3D image reconstruction on the high signal-to-noise ratio light field image to obtain 3D volume reconstruction data. The isotropic recovery functional block is used to perform isotropic resolution recovery processing on the 3D volume reconstruction data based on the trained light field isotropic recovery network to obtain a 3D Raman scattering signal image of the sample.

[0009] The present invention also provides a fast hyperspectral light field Raman imaging method, using the above-mentioned fast hyperspectral light field Raman imaging system, comprising: Step 1: Illuminate the sample using a uniform planar illumination light source to excite the sample to generate a Raman scattering signal; Step 2: Use a notch filter to filter out Rayleigh scattering light in the Raman scattering signal, and use an adjustable spectral filtering submodule to select the Raman scattering signal of the desired wavelength to receive; Step 3: Encode the Raman scattering signal of the required wavelength using a multi-dimensional information encoding unit to obtain multiple encoded beams with multi-dimensional information of sample space and angle; use a CCD detector to collect the beam information of the multiple encoded beams to generate an optical field image. Step 4: Perform three-dimensional image restoration processing on the light field image based on the light field reconstruction algorithm to obtain the three-dimensional Raman scattering signal image of the sample; wherein, the three-dimensional image restoration processing includes: image denoising and super-resolution processing, three-dimensional image reconstruction processing and isotropic resolution restoration processing.

[0010] In one embodiment of the present invention, the step of performing three-dimensional image restoration processing on the light field image based on the light field reconstruction algorithm to obtain a three-dimensional Raman scattering signal image of the sample includes: S1: Use the trained light field super-resolution and denoising sub-network to perform denoising and super-resolution processing on the light field image to obtain a high signal-to-noise ratio light field image. The light field super-resolution and denoising sub-network includes a spatial Transformer network model and a channel Transformer network model. The spatial Transformer network model is used to model the global dependencies between local regions within a single-view image, and the channel Transformer network model is used to model the global dependencies between cross-view images. S2: Using the trained light field 3D reconstruction network, the high signal-to-noise ratio light field image is reconstructed into a 3D image to obtain 3D volume reconstruction data; S3: Based on the trained isotropic optical field recovery network, isotropic resolution recovery processing is performed on the three-dimensional volume reconstruction data to obtain the three-dimensional Raman scattering signal image of the sample.

[0011] In one embodiment of the present invention, the trained optical field super-resolution and denoising sub-network is obtained by training using a blind super-resolution method, and the training process includes: S1.1: Using the physical model of light field imaging, a training dataset is generated, which includes paired dealiased light field images and noisy light field images; S1.2: Using the noisy light field images in the training dataset, train the noisy light field reconstruction network to extract the degenerate embedding, and obtain the trained noisy light field reconstruction network; S1.3: Input the noisy light field images of the i-th batch into the (i-1)-th light field super-resolution and denoising sub-network, and perform feature extraction and image enhancement on the input noisy light field images by alternately executing the spatial Transformer network model and the channel Transformer network model, and fusing the upsampling module, to obtain the super-resolution result of the i-th batch; where i is a positive integer; S1.4: Input the super-resolution results of the i-th batch into the trained noisy light field reconstruction network to obtain the corresponding degenerate embedding results, calculate the self-supervised loss between the super-resolution results of the i-th batch and the corresponding noisy light field image, and extract the high-frequency components from the noisy light field image of the i-th batch through discrete wavelet transform, and calculate the high-frequency loss between the high-frequency components and the super-resolution results. S1.5: Based on the calculated self-supervised loss and high-frequency loss of the i-th batch, update the model parameters of the initial optical field super-resolution and denoising sub-network through the backpropagation algorithm to obtain the i-th optical field super-resolution and denoising sub-network. S1.6: Determine whether the i-th optical field super-resolution and denoising sub-network has converged; if not, let i = i + 1 in S1.3 and re-execute S1.3 to S1.5; if yes, use the current optical field super-resolution and denoising sub-network as the trained optical field super-resolution and denoising sub-network.

[0012] In one embodiment of the present invention, the trained light field 3D reconstruction network is obtained by training in the following manner: S2.1: Using a light field imaging physical model, generate a pre-training dataset; the pre-training dataset includes multiple spatially enhanced light field images; S2.2: Perform percentile normalization on the spatial information enhanced light field image and the target high-resolution image of the jth batch, and obtain the corresponding synthetic light field image and target three-dimensional volume data by scaling, flipping and rotation transformation. S2.3: The synthetic light field images of the jth batch are rearranged and placed in multiple independent channels, and the corresponding channel information is obtained. The jth group of three-dimensional representations is constructed using the channel information. S2.4: Input the j-th group of three-dimensional representations into the (j-1)-th light field three-dimensional reconstruction network, and output the corresponding three-dimensional volume reconstruction data; S2.5: Calculate the difference between the three-dimensional volume reconstruction data corresponding to the j-th group of three-dimensional representations and the target three-dimensional volume data to obtain the j-th loss value; S2.6: Determine whether the (j-1)th light field 3D reconstruction network has converged; if not, update the parameters of each layer in the (j-1)th light field 3D reconstruction network using the j-th loss value to obtain the j-th light field 3D reconstruction network, let j=j+1, and return to execute S2.2; if yes, use the current light field 3D reconstruction network as the trained light field 3D reconstruction network.

[0013] In one embodiment of the present invention, the trained isotropic optical field restoration network includes a trained degradation generator and a trained isotropic resolution restoration generator; the trained isotropic optical field restoration network is obtained through training in the following manner: S3.1: Obtain a historical 3D volume reconstruction dataset, and segment multiple 3D volume reconstruction data in the historical 3D volume reconstruction dataset along the horizontal plane and the axial plane to obtain a set of horizontal reconstruction data and a set of axial reconstruction data respectively; the image resolution in the horizontal reconstruction data is greater than the image resolution in the axial reconstruction data; S3.2: The first and second phases are executed alternately until the network converges, resulting in a trained isotropic optical field recovery network. The first stage: In the p-th iteration, the lateral reconstruction data of the p-th batch is input into the (p-1)-th iteration degradation generator to generate the p-th batch degradation image. The p-th loss value is obtained by using the difference between the p-th batch degradation image and the p-th batch axial reconstruction data. It is then determined whether the (p-1)-th iteration degradation generator has converged. If not, the (p+1)-th iteration is executed. If so, the current degradation generator is used as the trained degradation generator. The second stage: In the q-th iteration, the axial reconstruction data of the q-th batch is input into the isotropic restoration generator of the (q-1)-th iteration to generate the restored image of the q-th batch. The restored image of the q-th batch is input into the trained degradation generator to generate the cyclic axial image of the q-th batch. The cyclic consistency loss between the cyclic axial image of the q-th batch and the axial reconstruction data of the q-th batch is calculated to obtain the q-th loss value. It is then determined whether the isotropic restoration generator of the (q-1)-th iteration has converged. If not, the (q+1)-th iteration is executed. If so, the current isotropic restoration generator is used as the trained isotropic restoration generator.

[0014] In one embodiment of the present invention, multiple 3D volume reconstruction data in the historical 3D volume reconstruction dataset are sliced ​​along the Z-axis to obtain a transverse image in the XY plane; multiple 3D volume reconstruction data in the historical 3D volume reconstruction dataset are sliced ​​along the X-axis or Y-axis to obtain an axial image in the YZ plane or XZ plane.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: The rapid hyperspectral Raman imaging system of this invention outputs a uniform planar illumination light source through an excitation source module. The signal acquisition module, while filtering Rayleigh scattering light and selecting the desired wavelength Raman signal, simultaneously uses a multi-dimensional information encoding unit to encode and acquire full depth-of-field information. The control and calculation module rapidly completes 3D image reconstruction based on a light field reconstruction algorithm. This system integrates optical excitation, signal processing, and intelligent computing, achieving efficient data acquisition and reconstruction, and significantly improving imaging speed and data accuracy.

[0016] This invention presents a rapid hyperspectral Raman imaging method. It employs a uniform planar illumination source to excite the sample and generate Raman scattering signals. A notch filter and an adjustable spectral filtering submodule effectively filter out Rayleigh scattering light and select the target wavelength. A multi-dimensional information encoding unit encodes the signal into a light field, and a CCD detector acquires the encoded beam information to generate a light field image. Finally, a light field reconstruction algorithm is used to restore the three-dimensional image, significantly improving the efficiency and image quality of Raman imaging. This method not only overcomes the inefficiency and signal interference problems caused by layer-by-layer scanning in traditional imaging techniques but also effectively suppresses image blurring and artifact generation by integrating denoising, super-resolution, and isotropic resolution restoration processing. It achieves high signal-to-noise ratio and high-resolution three-dimensional Raman imaging, suitable for in-situ and dynamic observation of complex samples.

[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a fast hyperspectral light field Raman volume imaging system provided in an embodiment of the present invention; Figure 2 This is a flowchart of the rapid hyperspectral light field Raman imaging method provided in the embodiments of the present invention; Figure 3 This is a flowchart of the light field reconstruction algorithm provided in the embodiments of the present invention; Figure 4 This is a flowchart of S1 in the light field reconstruction algorithm provided in this embodiment of the invention; Figure 5 This is a flowchart of S2 in the light field reconstruction algorithm provided in this embodiment of the invention; Figure 6 This is a flowchart of S3 in the light field reconstruction algorithm provided in this embodiment of the invention.

[0019] Figure reference numerals: 1-Excitation source module; 11-Continuous wave laser; 12-Light source generation sub-module; 2-Signal acquisition module; 21-Dichroic mirror; 22-Objective lens; 23-Notch filter; 24-Tunable spectral filtering sub-module; 25-Lens; 26-Multi-dimensional information encoding unit; 27-Relay system; 28-CCD detector; 3-Control and calculation module. Detailed Implementation

[0020] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of a rapid hyperspectral light field Raman imaging system and method proposed according to the present invention is provided in conjunction with the accompanying drawings and specific embodiments.

[0021] The foregoing and other technical contents, features, and effects of the present invention will be clearly presented in the following detailed description of specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and concrete understanding can be gained of the technical means and effects adopted by the present invention to achieve its intended purpose. However, the accompanying drawings are for reference and illustration only and are not intended to limit the technical solutions of the present invention.

[0022] Example 1 Traditional Raman imaging primarily relies on point scanning, line scanning, and wide-field illumination, focusing on two-dimensional chemical information analysis. However, it has inherent limitations in acquiring three-dimensional structural information when analyzing complex samples. Three-dimensional reconstruction technology can reveal the internal chemical distribution and structural characteristics of samples, playing an irreplaceable role in many fields. However, traditional 3D reconstruction methods, such as those employing a layer-by-layer scanning strategy to acquire information at different depths, require considerable time for each layer. As the number of scanning layers increases, photon scattering intensifies, severely interfering with the acquisition of the sample's Raman signal, leading to blurred images and increased noise. When these interfered two-dimensional images are used for 3D reconstruction, the accuracy and quality of the resulting 3D model significantly decrease, failing to accurately reflect the sample's true three-dimensional structure and affecting subsequent analysis and judgment.

[0023] Therefore, the present invention provides a fast hyperspectral light field Raman volume imaging system, such as... Figure 1 As shown, Figure 1 This is a schematic diagram of a fast hyperspectral light field Raman imaging system provided in an embodiment of the present invention.

[0024] This embodiment provides a fast hyperspectral light field Raman volume imaging system, including: an excitation source module 1, a signal acquisition module 2, and a control and calculation module 3. The excitation source module 1 provides a high-quality, low-noise, and stable uniform planar illumination source for the light field imaging system. The signal acquisition module 2, located after the excitation source module 1, illuminates the sample with the uniform planar illumination source to excite the sample to generate a Raman scattering signal. After filtering out Rayleigh scattering light from the Raman scattering signal, it selects and receives the Raman scattering signal of the desired wavelength through an adjustable spectral filtering submodule 24. The signal acquisition module 2 encodes information within the entire sample depth range through a multi-dimensional information encoding unit 26, transmits and acquires the encoded information to obtain a light field image. The control and calculation module 3 is electrically connected to the excitation source module 1 and the signal acquisition module 2, respectively, and performs light field reconstruction on the light field image based on a light field reconstruction algorithm to obtain a three-dimensional Raman scattering signal image of the sample.

[0025] It is worth noting that traditional light sheet illumination requires a specialized optical system to generate the light sheet, including complex components such as cylindrical lenses. This results in high equipment complexity and increased cost, and also places more stringent requirements on equipment calibration and debugging, increasing the difficulty of use and maintenance. The Raman wide-field illumination system used in this invention has relatively simple equipment, lower cost, and is easy to set up and maintain. Combined with a light field reconstruction algorithm, it can achieve fast and accurate light field Raman imaging. In an optional embodiment, the excitation source module 1 includes a continuous-wave laser 11 and a light source generation submodule 12. The continuous-wave laser 11 generates continuous-wave laser light. The light source generation submodule 12 includes a 4F system (composed of two conjugate lenses), a lens 25, a dichroic mirror 21, and an objective lens 22 arranged sequentially along the optical path. It is positioned after the continuous-wave laser 11 and is used to uniformly illuminate the light source based on the output of the continuous-wave laser 11. The laser light output from the continuous-wave laser 11 is expanded and collimated by the 4F system, then reflected and focused by the lens 25 and the dichroic mirror 21 onto the rear pupil plane of the objective lens 22, and then generates a uniformly propagating beam through the objective lens 22.

[0026] In an optional implementation, the signal acquisition module 2 includes: an objective lens 22, a dichroic mirror 21, a notch filter 23, an adjustable spectral filtering submodule 24, a lens 25, a multi-dimensional information encoding unit 26, a relay system 27, and a CCD detector 28 arranged sequentially along the optical path. The objective lens 22 and lens 25 are used together to collect the Raman scattering signal excited by the sample irradiated by the excitation source module 1. The dichroic mirror 21 is used to accurately separate the laser excitation light and the Raman scattering signal light emitted by the sample according to the wavelength of the light, achieving efficient acquisition of the Raman scattering signal. The notch filter 23 is used to filter out Rayleigh scattering light in the Raman scattering signal. The adjustable spectral filtering submodule 24 includes an adjustable long-pass filter, an adjustable short-pass filter, and an electric rotary stage for selecting and receiving the Raman scattering signal of the desired wavelength. The multi-dimensional... Information encoding unit 26, such as a microlens array, is used to encode information within the entire sample depth range. The information is the Raman scattering signal of the desired wavelength to be received. The encoded information within the entire sample depth range is transmitted to CCD detector 28 via relay system 27, such as a macro lens, to obtain a light field image. Excitation source module 1 illuminates the sample to generate Raman scattering signals. Objective lens 22 collects the Raman scattering signals, which are then filtered by dichroic mirror 21 and notch filter 23 to remove the laser excitation light. The Raman scattering signal of the desired wavelength is then selected and received via tunable spectral filtering submodule 24. The Raman scattering signal of the desired wavelength is focused onto multi-dimensional information encoding unit 26 by lens 25, thereby achieving the encoding of the signal within the entire sample depth range. Finally, the encoded information is transmitted to CCD detector 28 via relay system 27.

[0027] For example, the tunable spectral filtering submodule 24 includes an tunable long-pass filter, an tunable short-pass filter, and mechanical devices such as an electric rotary table, used to selectively receive Raman scattering signals of the desired wavelength to remove background noise interference. The cutoff frequencies of both the tunable long-pass filter and the tunable short-pass filter change with the angle between the incident light and the filter, dynamically selecting the Raman scattering signal of the desired wavelength to remove background noise interference; the cutoff wavelength is largest when the incident light is perpendicular, and gradually decreases as the angle between the filter normal and the light increases, reaching its minimum at 60°; at this time, the tunable spectral filtering submodule 24 can dynamically adjust the cutoff wavelength by changing the angle between the incident light and the filter to selectively receive specific signals.

[0028] Based on this, by collecting angle-wavelength data of two filters between 0° and 60° and comparing it with the actual calculated cutoff wavelength, a fitting relationship between the cutoff wavelength and the rotation angle is obtained. After determining the target wavelength, the target cutoff wavelength of the filter can be calculated, and the rotation angle of each of the two filters can be calculated based on the fitting relationship between the cutoff wavelength and the rotation angle. Finally, according to the rotation angle of each of the two filters, the rotating platform is controlled to rotate accordingly to collect the Raman scattering signal of the required wavelength. Wavelength selective reception is achieved by dynamically adjusting the filter angle through the adjustable spectral filtering submodule 24, avoiding the inconvenience of replacing filters required by traditional fixed filters, and significantly improving the system flexibility and multi-wavelength imaging efficiency.

[0029] In an optional implementation, the control and calculation module 3 includes a computer control unit and a computer processing unit. The computer control unit controls the continuous wave laser 11, the tunable spectral filter submodule 24, and the CCD detector 28, respectively. The computer processing unit includes a light field super-resolution and denoising function block, a light field three-dimensional reconstruction function block, and an isotropic recovery function block. The light field super-resolution and denoising function block is used to perform denoising and super-resolution processing on the light field image to obtain a high signal-to-noise ratio light field image. The light field three-dimensional reconstruction function block is used to perform three-dimensional image reconstruction on the high signal-to-noise ratio light field image to obtain three-dimensional volume reconstruction data. The isotropic recovery function block is used to perform isotropic resolution recovery processing on the three-dimensional volume reconstruction data according to the trained light field isotropic recovery network to obtain a three-dimensional Raman scattering signal image of the sample.

[0030] The rapid hyperspectral Raman imaging system of this invention outputs a uniform planar illumination light source through an excitation source module. The signal acquisition module, while filtering noise and selecting the desired wavelength Raman signal, simultaneously uses a multi-dimensional information encoding unit to encode and acquire full depth-of-field information. The control and calculation module rapidly completes 3D image reconstruction based on a light field reconstruction algorithm. This system integrates optical excitation, signal processing, and intelligent computing, achieving efficient data acquisition and reconstruction, and significantly improving imaging speed and data accuracy.

[0031] Example 2 This embodiment provides a fast hyperspectral light field Raman imaging method, using the fast hyperspectral light field Raman imaging system of Embodiment 1, such as... Figure 2 As shown, Figure 2 This is a flowchart of the rapid hyperspectral light field Raman imaging method provided in the embodiments of the present invention.

[0032] In this embodiment, a fast hyperspectral light field Raman volume imaging method includes: Step 1: Illuminate the sample using a uniform planar illumination light source to excite the sample to generate Raman scattering signals; Step 2: Use a notch filter to filter out Rayleigh scattering light in the Raman scattering signal, and use an adjustable spectral filtering submodule to select the Raman scattering signal of the desired wavelength to receive; Step 3: Encode the Raman scattering signal of the required wavelength using a multi-dimensional information encoding unit to obtain multiple encoded beams with multi-dimensional information of sample space and angle; use a CCD detector to collect beam information of multiple encoded beams to generate an optical field image. Step 4: Perform three-dimensional image restoration processing on the light field image based on the light field reconstruction algorithm to obtain the three-dimensional Raman scattering signal image of the sample; wherein, the three-dimensional image restoration processing includes: image denoising and super-resolution processing, three-dimensional image reconstruction processing and isotropic resolution restoration processing.

[0033] It is worth noting that the Raman signals excited by light-sheet illumination techniques based on tomographic stitching imaging are relatively weak. To obtain sufficient signal strength to ensure image quality, it is often necessary to extend the exposure time or use multi-frame averaging, which greatly reduces imaging efficiency. In contrast, this invention uses wide-field illumination to simultaneously excite the Raman scattering signal of the sample across the entire field of view. A single exposure can acquire the Raman signal of the sample across the entire depth of field, which has a significant advantage in imaging speed and can meet the rapid detection needs of fields such as biomedicine.

[0034] Furthermore, since most existing deep learning-based light field reconstruction methods rely on a single network, they struggle to effectively overcome the multiple degradation problems caused by noise, dimensionality reduction (insufficient angular resolution), and insufficient spatial resolution in light field imaging. Therefore, this invention innovatively designs a fast, high-resolution light field reconstruction algorithm based on multi-stage data prior guidance. This algorithm is used to perform 3D image restoration processing on light field images to obtain a 3D Raman scattering signal image of the sample. The 3D image restoration processing includes: image denoising and super-resolution processing, 3D image reconstruction processing, and isotropic resolution restoration processing. After 3D image restoration processing, the lost 3D imaging accuracy due to noise, dimensionality reduction, and resolution limitations during the imaging process can be restored, ensuring that the reconstructed image highly restores the true structure of the sample.

[0035] like Figure 3 As shown, Figure 3 This is a flowchart of the light field reconstruction algorithm provided in an embodiment of the present invention.

[0036] In this embodiment, a three-dimensional image restoration process is performed on the light field image based on the light field reconstruction algorithm to obtain a three-dimensional Raman scattering signal image of the sample, including: S1: The trained light field super-resolution and denoising sub-network is used to denoise and super-resolution the light field image to obtain a high signal-to-noise ratio light field image. The light field super-resolution and denoising sub-network includes a spatial Transformer network model and a channel Transformer network model. The spatial Transformer network model is used to model the global dependencies between local regions within a single-view image, and the channel Transformer network model is used to model the global dependencies between cross-view images.

[0037] like Figure 4 As shown, Figure 4 This is a flowchart of S1 in the light field reconstruction algorithm provided in the embodiment of the present invention.

[0038] In this embodiment, the light field super-resolution and denoising sub-network addresses the interdependence between images from different viewpoints in the light field image and the noise present during the actual image acquisition process. The trained light field super-resolution and denoising sub-network is obtained through a blind super-resolution method, and the training process includes: S1.1: Using the physical model of light field imaging, a training dataset is generated, which includes paired dealiased light field images and noisy light field images; S1.2: Using noisy light field images in the training dataset, train a noisy light field reconstruction network to extract degenerate embeddings, and obtain a trained noisy light field reconstruction network; S1.3: Input the noisy light field images of the i-th batch into the (i-1)-th light field super-resolution and denoising sub-network. By alternately executing the spatial Transformer network model and the channel Transformer network model, and fusing the upsampling module, feature extraction and image enhancement are performed on the input noisy light field images to obtain the super-resolution results of the i-th batch; where i is a positive integer. S1.4: Input the super-resolution results of the i-th batch into the trained noisy light field reconstruction network to obtain the corresponding degenerate embedding results. Calculate the self-supervised loss between the super-resolution results of the i-th batch and the corresponding noisy light field image. Extract high-frequency components from the noisy light field image of the i-th batch through discrete wavelet transform and calculate the high-frequency loss between the high-frequency components and the super-resolution results. S1.5: Based on the calculated self-supervised loss and high-frequency loss of the i-th batch, update the model parameters of the initial optical field super-resolution and denoising sub-network through the backpropagation algorithm to obtain the i-th optical field super-resolution and denoising sub-network. S1.6: Determine whether the i-th optical field super-resolution and denoising sub-network has converged; if not, let i = i + 1 in S1.3 and re-execute S1.3 to S1.5; if yes, use the current optical field super-resolution and denoising sub-network as the trained optical field super-resolution and denoising sub-network.

[0039] The principle behind this invention is that, although current convolutional neural network-based methods have achieved significant results in super-resolution of light field images, these methods still have limitations. They cannot fully simulate the nonlocal properties of 4D light field data and are only suitable for overcoming degradation caused by downsampling, but are difficult to effectively overcome degradation caused by noise. Therefore, this invention proposes a method for training light field super-resolution and denoising sub-networks based on a blind super-resolution approach. First, using a physical model of light field imaging, combined with sub-aperture displacement characteristics and the noise and background distribution of the actual acquired images, de-aliased light field images and noisy light field images are generated respectively to guide the training of the light field super-resolution and denoising sub-networks.

[0040] Furthermore, addressing the interdependencies between light field images, this invention designs a network model comprising a spatial Transformer and a channel Transformer. The spatial Transformer models the global dependencies between local regions within a single-view image, while the channel Transformer models the global dependencies between images across different viewpoints. This significantly improves the super-resolution and denoising capabilities of light field images, especially maintaining detail recovery even under low signal-to-noise ratio conditions. By alternately executing spatial and channel Transformer operations and fusing an upsampling module, a light field image containing richer spatial information is generated. However, actual captured light field images typically contain significant noise, which a single super-resolution network struggles to overcome. Therefore, by combining the aforementioned network model with a blind super-resolution method, direct super-resolution of light field images with low signal-to-noise ratios is achieved.

[0041] Specifically, the blind super-resolution method integrates supervised training and self-supervised learning, comprising two stages: The first stage trains a noisy light field reconstruction network specifically designed to extract degraded embeddings from noisy light field images. These degraded embeddings are then applied to the target dealiased light field image to guide the noisy light field reconstruction process. The second stage first obtains the super-resolution result from the pre-trained super-resolution network and inputs this result into the fixed noisy light field reconstruction network to generate the corresponding degraded embedding. Subsequently, the degraded result is compared with the original noisy light field image, and the self-supervised loss is calculated and used only to update specific parameters in the super-resolution network. Furthermore, it was observed that the super-resolution network recovers low-frequency information well, but its high-frequency detail recovery is unstable. Using discrete wavelet transform to directly extract high-frequency components from the noisy light field image can specifically guide the model to focus on high-frequency information recovery without modifying its good modeling ability for low-frequency information. This method can significantly improve the generalization and detail recovery capabilities of super-resolution network models on unseen real-world datasets.

[0042] S2: Using a trained light field 3D reconstruction network, perform 3D image reconstruction on a high signal-to-noise ratio light field image to obtain 3D volume reconstruction data.

[0043] like Figure 5As shown, Figure 5 This is a flowchart of S2 in the light field reconstruction algorithm provided in the embodiment of the present invention.

[0044] In this embodiment, targeting the different scale structural features present in biological samples, a fast and high-resolution 3D reconstruction of the super-resolution and denoised images is achieved by combining a multi-scale convolutional residual structure with a U-net network structure. The trained light field 3D reconstruction network is obtained through the following training method: S2.1: Generate a pre-training dataset using a light field imaging physical model; the pre-training dataset includes multiple spatially enhanced light field images; S2.2: Perform percentile normalization on the spatial information enhanced light field image and the target high-resolution image of the jth batch, and obtain the corresponding synthetic light field image and target three-dimensional volume data by scaling, flipping and rotation transformation. S2.3: Rearrange the synthetic light field images of the jth batch and place them in multiple independent channels, obtain the corresponding channel information, and use the channel information to construct the jth group of three-dimensional representations; S2.4: Input the j-th group of three-dimensional representations into the (j-1)-th light field three-dimensional reconstruction network, and output the corresponding three-dimensional volume reconstruction data; S2.5: Calculate the difference between the 3D volume reconstruction data corresponding to the j-th group of 3D representations and the target 3D volume data to obtain the j-th loss value; S2.6: Determine whether the (j-1)th light field 3D reconstruction network has converged; if not, update the parameters of each layer in the (j-1)th light field 3D reconstruction network using the j-th loss value to obtain the j-th light field 3D reconstruction network, let j=j+1, and return to execute S2.2; if yes, use the current light field 3D reconstruction network as the trained light field 3D reconstruction network.

[0045] The principle is as follows: First, a light field image containing more spatial information, generated based on a physical model of light field imaging, is used as the network input, and high-resolution 3D volume data is used as the target image output. After percentile normalization of the input and target images, the data is further expanded through scaling, flipping, and rotation transformations of the light field image and the real image. In this process, each synthesized light field image is rearranged into different views, which are placed in different channels, thereby converting the channel information into the depth information of the 3D image to construct the initial 3D representation.

[0046] By iteratively minimizing the difference between the network output and the target image, the parameters of each layer of the network are gradually optimized. Specifically, the encoding module in the U-net network is replaced with a multi-scale convolutional attention mechanism (MSCA) module to coordinate features at different scales and enhance the reconstruction capability from super-resolution light field images to high-resolution 3D volume data. Each MSCA module consists of batch normalization, an attention mechanism module, and a feed-forward neural network (FFN). The attention module consists of a 1×1 convolution, a Gaussian Error Linear Unit (GELU) activation function, and the MSCA module; the FFN consists of a 1×1 convolution, a 3×3 depthwise convolution, and the GELU activation function. In particular, multi-branch depthwise separable convolutions can fuse local details and multi-scale information, and integrate information from multiple feature channels through 1×1 convolutions.

[0047] S3: Based on the trained optical field isotropic resolution recovery network, isotropic resolution recovery processing is performed on the 3D volume reconstruction data to obtain the 3D Raman scattering signal image of the sample.

[0048] like Figure 6 As shown, Figure 6 This is a flowchart of S3 in the light field reconstruction algorithm provided in this embodiment of the invention.

[0049] In this embodiment, the output of the light field 3D reconstruction network (i.e., the reconstructed 3D volume data) is input into the light field isotropic recovery network for training, ultimately obtaining high-resolution 3D volume data with near-isotropic resolution. The trained light field isotropic recovery network includes a trained degradation generator and a trained isotropic recovery generator; the trained light field isotropic recovery network is obtained through the following training method: S3.1: Obtain the historical 3D volume reconstruction dataset, and segment the multiple 3D volume reconstruction data in the historical 3D volume reconstruction dataset along the horizontal plane and the axial plane to obtain the set of horizontal reconstruction data and the set of axial reconstruction data respectively; the image resolution in the horizontal reconstruction data is greater than the image resolution in the axial reconstruction data; S3.2: The first and second phases are executed alternately until the network converges, resulting in a trained isotropic optical field recovery network. Phase 1: In the p-th iteration, the lateral reconstruction data of the p-th batch is input into the degradation generator of the (p-1)-th iteration to generate the degradation image of the p-th batch. The p-th loss value is obtained by using the difference between the degradation image of the p-th batch and the axial reconstruction data of the p-th batch. It is then determined whether the degradation generator of the (p-1)-th iteration has converged. If not, the p+1-th iteration is executed. If so, the current degradation generator is used as the trained degradation generator. Phase 2: In the q-th iteration, the axial reconstruction data of the q-th batch is input into the isotropic restoration generator of the (q-1)-th iteration to generate the restored image of the q-th batch. The restored image of the q-th batch is input into the trained degradation generator to generate the cyclic axial image of the q-th batch. The cyclic consistency loss between the cyclic axial image of the q-th batch and the axial reconstruction data of the q-th batch is calculated to obtain the q-th loss value. It is then determined whether the isotropic restoration generator of the (q-1)-th iteration has converged. If not, the (q+1)-th iteration is executed. If so, the current isotropic restoration generator is used as the trained isotropic restoration generator.

[0050] In one optional implementation, multiple 3D volume reconstruction data from the historical 3D volume reconstruction dataset are sliced ​​along the Z-axis to obtain a lateral image in the XY plane; multiple 3D volume reconstruction data from the historical 3D volume reconstruction dataset are sliced ​​along the X-axis or Y-axis to obtain an axial image in the YZ plane or XZ plane.

[0051] The principle is as follows: First, the high-resolution 3D volume data is divided into multiple horizontal and axial planar images. Utilizing the inherent anisotropic properties of the optical point spread function, the high-resolution horizontal image is considered a natural "gold standard" for improving the axial resolution of the original data. Based on this, without the need for additional paired training data, unpaired horizontal and axial images extracted from the same anisotropic 3D volume can be directly used. Unsupervised training learns the mapping relationship from the low-resolution axial image to the high-resolution horizontal image, thus achieving effective training of a single image stack. The entire optimization process employs a two-stage CycleGAN strategy: Stage 1 trains the CycleGAN network, learning the degenerate mapping from the high-resolution horizontal image to the low-resolution axial image, generating the corresponding blurred axial image; unsupervised CycleGAN structure is used for isotropic restoration, significantly reducing the difficulty of obtaining training data and improving the model's generalization ability in practical applications; Stage 2 combines the blurred axial image generated in Stage 1 with the pixel-aligned high-resolution horizontal image, and by minimizing the pixel-by-pixel reconstruction loss, trains a sub-network specifically for improving axial resolution. Finally, the networks in these two stages are interconnected and optimized collaboratively to improve performance in a loop, thereby achieving the restoration of isotropic high-resolution 3D images.

[0052] It is understood that the imaging method provided in Embodiment 2 of the present invention is implemented using the imaging system of Embodiment 1, and therefore has similar beneficial effects to the system embodiment of Embodiment 1. For technical details not disclosed in the method embodiments of the present invention, please refer to the description of the system embodiment for clarification.

[0053] This invention presents a rapid hyperspectral Raman imaging method. It employs a uniform planar illumination source to excite the sample and generate Raman scattering signals. A notch filter and an adjustable spectral filtering submodule effectively filter out Rayleigh scattering light and select the target wavelength. A multi-dimensional information encoding unit encodes the signal into a light field, and a CCD detector acquires the encoded beam information to generate a light field image. Finally, a light field reconstruction algorithm is used to restore the three-dimensional image, significantly improving the efficiency and image quality of Raman imaging. This method not only overcomes the inefficiency and signal interference problems caused by layer-by-layer scanning in traditional imaging techniques but also effectively suppresses image blurring and artifact generation by integrating denoising, super-resolution, and isotropic resolution restoration processing. It achieves high signal-to-noise ratio and high-resolution three-dimensional Raman imaging, suitable for in-situ and dynamic observation of complex samples.

[0054] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0055] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A fast hyperspectral light field Raman tomography system, characterized in that, include: The module consists of an excitation source module, a signal acquisition module, and a control and calculation module. The excitation source module is used to provide a high-quality, low-noise, stable, uniform planar illumination source for the light field imaging system; The signal acquisition module, located after the excitation source module, is used to irradiate the sample with the uniform plane illumination light source to excite the sample to generate a Raman scattering signal. After filtering out Rayleigh scattering light from the Raman scattering signal, the module selects and receives the Raman scattering signal of the desired wavelength through an adjustable spectral filtering submodule. The signal acquisition module is also used to encode information within the entire depth range of the sample through a multi-dimensional information encoding unit, and to transmit and acquire the encoded information to obtain a light field image. The control and calculation module is electrically connected to the excitation source module and the signal acquisition module, respectively, and is used to reconstruct the light field image based on the light field reconstruction algorithm to obtain the three-dimensional Raman scattering signal image of the sample.

2. The rapid hyperspectral light field Raman tomography system of claim 1, wherein, The excitation source module includes a continuous wave laser and a light source generation submodule; The continuous wave laser is used to generate continuous wave laser; the light source generation submodule is located after the continuous wave laser and includes a 4F system composed of two conjugate lenses, a lens, a dichroic mirror and an objective lens arranged sequentially along the optical path, for use to output a uniform plane irradiation light source according to the continuous wave laser. The laser output from the continuous wave laser is expanded and collimated by the 4F system, then reflected and focused onto the rear pupil plane of the objective lens by the lens and the dichroic mirror, and then generated into a uniformly propagating beam by the objective lens.

3. The rapid hyperspectral light field Raman volume imaging system according to claim 2, characterized in that, The signal acquisition module includes: an objective lens, a dichroic mirror, a notch filter, an adjustable spectral filtering submodule, a lens, a multi-dimensional information encoding unit, a relay system, and a CCD detector arranged sequentially along the optical path. The objective lens and the lens are used together to collect the Raman scattering signal excited by the sample irradiated by the excitation source module; The dichroic mirror is used to accurately separate the laser excitation light and the Raman scattering signal light emitted by the sample according to the wavelength of the light, so as to achieve efficient acquisition of the Raman scattering signal. The notch filter is used to filter out Rayleigh scattered light in the Raman scattering signal; The adjustable spectral filtering submodule includes an adjustable long-pass filter, an adjustable short-pass filter, and an electric rotary stage, used to select and receive Raman scattering signals of the desired wavelength; the cutoff frequencies of the adjustable long-pass filter and the adjustable short-pass filter change with the angle between the incident light and the filter, so as to dynamically select and receive Raman scattering signals of the desired wavelength to remove background noise interference. The multi-dimensional information encoding unit is used to encode information within the entire sample depth range, wherein the information is a Raman scattering signal of the desired wavelength selected for reception; the encoded information within the entire sample depth range is transmitted to the CCD detector through the relay system to obtain a light field image. The excitation source module illuminates the sample to generate a Raman scattering signal. The objective lens collects the Raman scattering signal, which is then filtered out by the dichroic mirror and the notch filter to remove the laser excitation light. The tunable spectral filtering submodule then selects and receives the Raman scattering signal of the desired wavelength. The Raman scattering signal of the desired wavelength is focused onto the multi-dimensional information encoding unit by a lens, thereby encoding the signal within the entire depth range of the sample. Finally, the encoded information is transmitted to the CCD detector via the relay system.

4. The rapid hyperspectral light field Raman volume imaging system according to claim 3, characterized in that, The control and calculation module includes: a computer control unit and a computer processing unit, wherein the computer control unit controls the continuous wave laser, the tunable spectral filter submodule and the CCD detector respectively; The computer processing unit includes: a light field super-resolution and denoising functional block, a light field 3D reconstruction functional block, and an isotropic resolution restoration functional block. The light field super-resolution and denoising functional block is used to perform denoising and super-resolution processing on the light field image to obtain a high signal-to-noise ratio light field image. The light field 3D reconstruction functional block is used to perform 3D image reconstruction on the high signal-to-noise ratio light field image to obtain 3D volume reconstruction data. The isotropic resolution restoration functional block is used to perform isotropic resolution restoration processing on the 3D volume reconstruction data based on the trained light field isotropic resolution restoration network to obtain a 3D Raman scattering signal image of the sample.

5. A rapid hyperspectral Raman imaging method, characterized in that, Using the rapid hyperspectral light field Raman volume imaging system according to any one of claims 1 to 4, comprising: Step 1: Illuminate the sample using a uniform planar illumination light source to excite the sample to generate a Raman scattering signal; Step 2: Use a notch filter to filter out Rayleigh scattering light in the Raman scattering signal, and use an adjustable spectral filtering submodule to select the Raman scattering signal of the desired wavelength to receive; Step 3: Encode the Raman scattering signal of the required wavelength using a multi-dimensional information encoding unit to obtain multiple encoded beams with multi-dimensional information of sample space and angle; use a CCD detector to collect the beam information of the multiple encoded beams to generate an optical field image. Step 4: Perform three-dimensional image restoration processing on the light field image based on the light field reconstruction algorithm to obtain the three-dimensional Raman scattering signal image of the sample; wherein, the three-dimensional image restoration processing includes: image denoising and super-resolution processing, three-dimensional image reconstruction processing and isotropic resolution restoration processing.

6. The rapid hyperspectral Raman imaging method according to claim 5, characterized in that, The step of performing three-dimensional image restoration processing on the light field image based on the light field reconstruction algorithm to obtain the three-dimensional Raman scattering signal image of the sample includes: S1: Use the trained light field super-resolution and denoising sub-network to perform denoising and super-resolution processing on the light field image to obtain a high signal-to-noise ratio light field image. The light field super-resolution and denoising sub-network includes a spatial Transformer network model and a channel Transformer network model. The spatial Transformer network model is used to model the global dependencies between local regions within a single-view image, and the channel Transformer network model is used to model the global dependencies between cross-view images. S2: Using the trained light field 3D reconstruction network, the high signal-to-noise ratio light field image is reconstructed into a 3D image to obtain 3D volume reconstruction data; S3: Based on the trained optical field isotropic resolution recovery network, the three-dimensional volume reconstruction data is subjected to isotropic resolution recovery processing to obtain the three-dimensional Raman scattering signal image of the sample.

7. The rapid hyperspectral Raman imaging method according to claim 6, characterized in that, The trained optical field super-resolution and denoising sub-network was obtained through blind super-resolution training, and the training process included: S1.1: Using the physical model of light field imaging, a training dataset is generated, which includes paired dealiased light field images and noisy light field images; S1.2: Using the noisy light field images in the training dataset, train the noisy light field reconstruction network to extract the degenerate embedding, and obtain the trained noisy light field reconstruction network; S1.3: Input the noisy light field images of the i-th batch into the (i-1)-th light field super-resolution and denoising sub-network, and perform feature extraction and image enhancement on the input noisy light field images by alternately executing the spatial Transformer network model and the channel Transformer network model, and fusing the upsampling module, to obtain the super-resolution result of the i-th batch; where i is a positive integer; S1.4: Input the super-resolution results of the i-th batch into the trained noisy light field reconstruction network to obtain the corresponding degenerate embedding results, calculate the self-supervised loss between the super-resolution results of the i-th batch and the corresponding noisy light field image, and extract the high-frequency components from the noisy light field image of the i-th batch through discrete wavelet transform, and calculate the high-frequency loss between the high-frequency components and the super-resolution results. S1.5: Based on the calculated self-supervised loss and high-frequency loss of the i-th batch, update the model parameters of the initial optical field super-resolution and denoising sub-network through the backpropagation algorithm to obtain the i-th optical field super-resolution and denoising sub-network. S1.6: Determine whether the i-th optical field super-resolution and denoising sub-network has converged; if not, let i = i + 1 in S1.3 and re-execute S1.3 to S1.5; if yes, use the current optical field super-resolution and denoising sub-network as the trained optical field super-resolution and denoising sub-network.

8. The rapid hyperspectral Raman imaging method according to claim 6, characterized in that, The trained light field 3D reconstruction network was obtained through the following training method: S2.1: Using a light field imaging physical model, generate a pre-training dataset; the pre-training dataset includes multiple spatially enhanced light field images; S2.2: Perform percentile normalization on the spatial information enhanced light field image and the target high-resolution image of the jth batch, and obtain the corresponding synthetic light field image and target three-dimensional volume data by scaling, flipping and rotation transformation. S2.3: The synthetic light field images of the jth batch are rearranged and placed in multiple independent channels, and the corresponding channel information is obtained. The jth group of three-dimensional representations is constructed using the channel information. S2.4: Input the j-th group of three-dimensional representations into the (j-1)-th light field three-dimensional reconstruction network, and output the corresponding three-dimensional volume reconstruction data; S2.5: Calculate the difference between the three-dimensional volume reconstruction data corresponding to the j-th group of three-dimensional representations and the target three-dimensional volume data to obtain the j-th loss value; S2.6: Determine whether the (j-1)th light field 3D reconstruction network has converged; if not, update the parameters of each layer in the (j-1)th light field 3D reconstruction network using the j-th loss value to obtain the j-th light field 3D reconstruction network, let j=j+1, and return to execute S2.2; if yes, use the current light field 3D reconstruction network as the trained light field 3D reconstruction network.

9. The rapid hyperspectral Raman imaging method according to claim 6, characterized in that, The trained isotropic resolution restoration network includes a trained degradation generator and a trained isotropic restoration generator; the trained isotropic resolution restoration network is obtained through the following training method: S3.1: Obtain a historical 3D volume reconstruction dataset, and segment multiple 3D volume reconstruction data in the historical 3D volume reconstruction dataset along the horizontal plane and the axial plane to obtain a set of horizontal reconstruction data and a set of axial reconstruction data respectively; the image resolution in the horizontal reconstruction data is greater than the image resolution in the axial reconstruction data; S3.2: The first and second stages are executed alternately until the network converges, resulting in a trained isotropic resolution recovery network for the optical field. The first stage: In the p-th iteration, the lateral reconstruction data of the p-th batch is input into the (p-1)-th iteration degradation generator to generate the p-th batch degradation image. The p-th loss value is obtained by using the difference between the p-th batch degradation image and the p-th batch axial reconstruction data. It is then determined whether the (p-1)-th iteration degradation generator has converged. If not, the (p+1)-th iteration is executed. If so, the current degradation generator is used as the trained degradation generator. The second stage: In the q-th iteration, the axial reconstruction data of the q-th batch is input into the isotropic resolution restoration generator of the (q-1)-th iteration to generate the restored image of the q-th batch. The restored image of the q-th batch is input into the trained degradation generator to generate the cyclic axial image of the q-th batch. The cyclic consistency loss between the cyclic axial image of the q-th batch and the axial reconstruction data of the q-th batch is calculated to obtain the q-th loss value. It is then determined whether the isotropic restoration generator of the (q-1)-th iteration has converged. If not, the (q+1)-th iteration is executed. If so, the current isotropic restoration generator is used as the trained isotropic restoration generator.

10. The rapid hyperspectral Raman imaging method according to claim 9, characterized in that, Slice multiple 3D volume reconstruction data in the historical 3D volume reconstruction dataset along the Z-axis to obtain a horizontal image in the XY plane; Slice multiple 3D volume reconstruction data in the historical 3D volume reconstruction dataset along the X-axis or Y-axis direction to obtain axial images of the YZ plane or XZ plane.

Images