All-optical three-dimensional scanning confocal fluorescence microscopic imaging device and implementation method thereof

Abstract

The invention discloses an all-optical three-dimensional scanning confocal fluorescent microscopic imaging device and an implementation method thereof. According to the invention, through a deep learning driven adaptive regulation and control method, CNN is adopted to process spatial distribution data and dynamically regulate and control a phase hologram and the light intensity and phase compensation of a laser, so that the focal point of exciting light is subjected to aberration-free axial displacement, a bidirectional parallel optical scanning track of a two-dimensional scanning system is optimized, and all-optical three-dimensional scanning is realized; lSTM and TCN are combined to obtain a time sequence dependency relationship, a laser light source, an adjustable diaphragm, an electric focus-adjustable lens and a photoelectric detector are integrally controlled, efficient synchronization and automatic operation and high-speed axial focusing adjustment are realized, errors and time sequence mismatch are eliminated, optical characteristics of different samples and environmental interference are automatically adapted, and high-quality imaging is kept. The robustness and the applicable scene range of the system are improved; the method is used for model biological embryo real-time tracking, intracellular signal molecule dynamic visualization, cell membrane protein migration and aggregation observation and intracellular organelle interaction tracking.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of microscopic imaging, in particular to a full light three-dimensional scanning confocal fluorescence microscopic imaging device and an implementation method thereof. BACKGROUND

[0002] As an important branch of modern optics, optical fluorescence microscopy is continuously innovated with the breakthrough progress of optical engineering and exhibits irreplaceable application value in biomedical research. Optical fluorescence microscopy works based on the principle of fluorescence labeling and optical detection, and through specific wavelength excitation light to promote the fluorescence molecules in the sample to emit fluorescence signals. This technology has become an indispensable tool for modern biomedical research, providing key technical support for revealing the basic laws of life activities. Contemporary biological imaging puts forward three core requirements for optical microscope systems: large field of view coverage, high spatiotemporal resolution, and three-dimensional volume imaging function. These technical requirements essentially come from the fundamental needs of living organism research - only with high-fidelity optical imaging means, can the precise observation and recording of subcellular structures and their dynamic biological processes be realized. Especially when studying fast transient life phenomena, these imaging parameter indicators often restrict each other, which constitutes the technical bottleneck that needs to be broken through in the field of biological optical imaging. In the existing microscopic technology system, wide-field fluorescence microscopy technology, with its single two-dimensional imaging characteristics, can obtain a series of two-dimensional images by cooperating with three-dimensional scanning of the sample. However, this technology has inherent physical limitations, that is, the fluorescence signal of the focal plane and the out-of-focus plane cannot be effectively distinguished during the imaging process, resulting in that the collected images contain a large amount of background noise. This technical defect directly leads to the inability to realize three-dimensional volume imaging, and cannot be applied to researches that require three-dimensional structure analysis. These limitations make it difficult for traditional wide-field fluorescence microscopy technology to meet the stringent requirements of modern biological research on imaging quality.

[0003] Confocal microscopy is an important breakthrough in modern fluorescence microscopy. It has many technical advantages through precise optical design. Firstly, it uses point scanning imaging. The excitation light source is focused by the objective lens to form a diffraction-limited spot for point-by-point scanning of the sample. The fluorescence signal excited at each point is collected by the same objective lens. The coaxial structure ensures the stability of in-situ excitation and imaging. The tightly focused light source forms a highly localized excitation focus in the sample. The focal point of the laser focused in the sample and the focal point of the fluorescence signal incident on the detector are in conjugate positions. The pinhole is close to the surface of the photodetector. The collected fluorescence is separated by a dichroic mirror and filtered by a conjugate pinhole to remove out-of-focus signals, leaving only focal point information. This significantly improves the signal-to-noise ratio and image contrast of the system. Secondly, the confocal microscope has excellent axial resolution due to its pinhole filtering method. It can achieve sub-micron optical slicing, making it particularly suitable for biological research scenarios that require high-precision two-dimensional imaging. In addition, the confocal system also has good optical tomography capability, which can clearly distinguish the structural information at different depths in the sample and achieve three-dimensional imaging of thick samples. Therefore, the confocal system is an excellent three-dimensional imaging tool for cells and other mesoscopic structures. However, the three-dimensional imaging speed of the confocal microscope is also limited by its point scanning principle. It can only focus on a single point at a time. To achieve three-dimensional reconstruction, the system must acquire three-dimensional information of the sample through point-by-point scanning, which significantly increases the data acquisition time. There are two main technical paths for confocal three-dimensional imaging: the first is the Z-axis scanning method based on mechanical displacement. It drives the sample stage or objective lens to switch the focal plane through a precision stepper motor. This method can ensure the imaging quality of the focal plane, but it significantly reduces the frame rate, which may introduce motion artifacts in fast dynamic process observation. More importantly, it limits the sample preparation to a fixed displacement stage, completely excluding the possibility of in vivo imaging. The second method changes the focal length of the lens to achieve fast axial scanning. Although this scheme improves the imaging speed, it introduces difficult-to-correct spherical aberrations and is limited by the lens focusing range, making it impossible to meet the needs of large-scale volume imaging. In addition, both schemes face the challenges of photobleaching and phototoxicity when implementing fast three-dimensional imaging, which to some extent limits their application in long-term observation of living organisms.The core of these problems is that the existing solutions generally lack adaptive sensing ability to dynamic changes of imaging scenes and intelligent regulation mechanism of multi-component cooperation: the mechanical displacement solution cannot adaptively adjust the scanning path and focal plane switching rhythm according to the real-time motion state of the sample, the imaging area characteristics and the like, and cannot adapt to in-vivo imaging at the hardware level; the focal length regulation solution cannot intelligently compensate for aberration for the optical characteristics of different axial depths, and the fixed focusing logic cannot adapt to the dynamic aberration changes of large-scale imaging; more importantly, both types of solutions have not established an intelligent correlation between imaging precision requirements, light dose allocation and scanning parameter matching, and can only output light dose and scanning parameters in a fixed mode, resulting in a difficulty in balancing imaging efficiency and biological sample activity protection, and failing to meet the needs of high-precision application scenarios such as long-time dynamic observation of living bodies and imaging of large-scale complex samples. SUMMARY

[0004] In view of the core bottlenecks of high-speed confocal microscopic imaging in imaging speed, aberration control, scene adaptability and biological sample activity protection, the present application provides a full-light three-dimensional scanning confocal fluorescence microscopic imaging device and an implementation method thereof, uses a depth adaptive regulation device, dynamically optimizes light field modulation parameters, accurately compensates for aberration, realizes axial large-range scanning without aberration under high numerical aperture, relies on intelligent collaborative control of high-speed resonant mirror bidirectional parallel scanning imaging technology, shortens sampling time, improves time resolution, accurately controls light dose output to reduce phototoxicity, and realizes full-light three-dimensional high-speed aberration-free confocal microscopic imaging.

[0005] One object of the present application is to provide a full-light three-dimensional scanning confocal fluorescence microscopic imaging device.

[0006] The full-light three-dimensional scanning confocal fluorescence microscopic imaging device of the present application comprises: an excitation light source, a depth adaptive regulation device, a dichroic mirror, an adjustable diaphragm, a two-dimensional scanning system, a telecentric scanning correction system, an objective lens, an electrically adjustable focus lens, a pinhole, a photodetector and an intelligent control system; wherein the two-dimensional scanning system is connected to a data acquisition card, and the photodetector is connected to an image acquisition card; the excitation light source, the depth adaptive regulation device, the adjustable diaphragm, the data acquisition card, the electrically adjustable focus lens and the image acquisition card are respectively connected to the intelligent control system.

[0007] The excitation light source emits a polarized Gaussian light beam as excitation light, which enters the depth adaptive regulation device; the spatial light modulator of the depth adaptive regulation device adjusts the phase of the excitation light by loading a phase hologram calculated by a feedback type aberration-free axial scanning algorithm and adjusted in real time by an adaptive regulation method driven by deep learning, to realize the movement of the focal point without aberration in the axial direction behind the objective lens; the first-order diffracted light is selected by the adjustable diaphragm after the dichroic mirror to the two-dimensional scanning system, for two-dimensional plane optical scanning, and focused by the objective lens to the surface or internal observation area of the sample.

[0008] The fluorescence signal generated on the surface or inside of the sample is collected through the same objective and transmitted back to the two-dimensional scanning system in the direction from which the excitation light comes, is descanned by the two-dimensional scanning system, is separated from the excitation light through a dichroic mirror, is adjusted in focal length by an electrically adjustable focusing lens according to the change in the axial position of the excitation light, so that the fluorescence signals at different axial positions in the sample can be focused on the detection surface of the photodetector, and is received by the photodetector after the out-of-focus signals are filtered out through a pinhole located in front of the photodetector.

[0009] Through the intelligent control system, an adaptive regulation and control method driven by deep learning is adopted to obtain the accurate relationship between the phase hologram and the axial position, and the light intensity of the laser is optimized according to the axial position of the focal point, and the inherent aberration of the optical system is eliminated, the phase hologram loaded by the depth adaptive regulation and control device, the light intensity of the laser and the phase compensation of the optical system are dynamically regulated and controlled, so that the excitation light is axially displaced without aberration at the focal point behind the high numerical aperture objective, and the bidirectional parallel optical scanning track of the two-dimensional scanning system is optimized in real time, and full-optical three-dimensional scanning is realized.

[0010] The sample is placed on a sample holder.

[0011] The beam expansion and collimation device is further included and is placed behind the excitation light source to expand and collimate the excitation light; the beam expansion and collimation device adopts a double-lens to expand and collimate; the double-lens is accurately configured by a pair of lenses with different focal lengths; a short focal length lens with a focal length of f1 is used as an input lens, a long focal length lens with a focal length of f2 is used as an output lens, the expansion ratio M = f2 / f1, and the diameter of the expanded light beam should match the liquid crystal panel of the spatial light modulator.

[0012] The depth adaptive regulation and control device includes a half-wave plate, a prism and a kilohertz pure phase silicon-based liquid crystal spatial light modulator, which together realize precise regulation and control of the light field; the half-wave plate modulates the polarization state of the incident excitation light, and by accurately rotating the angle of the half-wave plate, linearly polarized light is adjusted to a polarization state that matches the best working polarization direction of the spatial light modulator; the prism is used to cooperate with the spatial light modulator to complete the light field regulation of the expanded and collimated and polarization state modulated Gaussian light; the Gaussian light is reflected by one side of the prism into the liquid crystal screen of the spatial light modulator, the spatial light modulator changes the axial position of the excitation light focused on the objective by loading a phase hologram, thereby realizing light field regulation and completing high-speed axial optical scanning.

[0013] A 4f relay system is further included, which is arranged between the dichroic mirror and the two-dimensional scanning system. A strict 4f relay structure realizes a beam shrinking function, including two coaxially arranged input lenses and output lenses, which relay the plane modulated by the spatial light modulator to the two-dimensional scanning system. An adjustable diaphragm is located at the back focal plane of the input lens and the front focal plane of the output lens, which is used to screen the first-order diffracted light with complete phase modulation information and the highest diffraction efficiency as the excitation light.

[0014] A telecentric scanning correction system is further included, which is arranged between the two-dimensional scanning system and the objective lens. The system includes a scanning lens and a first tube lens. The light passing through the two-dimensional scanning system first passes through the scanning lens and then passes through the first tube lens, ensuring that the light is relayed from the two-dimensional scanning system to the back focal plane of the objective lens. The system can maintain a constant optical magnification during scanning, ensuring that the imaging quality is not affected by changes in the position or distance of the scanned object.

[0015] A bright-field imaging system is further included. The bright-field imaging system includes a white light source, a beam splitter, and a camera. The white light source is located in front of the objective lens, and the white light source emits white illumination light that uniformly irradiates the sample surface. The light signal is accurately transmitted to the first tube lens after being collected by the objective lens and reaches the detection surface of the camera through the beam splitter. The camera collects the reflected light signal from the sample surface to obtain the two-dimensional plane information of the sample surface. During this process, the system strictly follows the 4f optical transfer principle to ensure the faithful transmission of sample information from the front focal plane of the objective lens to the back focal plane of the first tube lens. Finally, the light signal is received by a high-sensitivity scientific camera to realize bright-field imaging of the two-dimensional structure of the sample.

[0016] The photodetector includes a focusing lens, a pinhole, and a photomultiplier tube. The pinhole allows only the in-focus light signal at the focal plane of the objective lens to pass through the pinhole and enter the photomultiplier tube to participate in imaging, while the out-of-focus fluorescent signal is blocked outside the detection pinhole. The photomultiplier tube receives the weak fluorescent signal from the focal point of the objective lens, and finally reconstructs the three-dimensional structure of the sample through three-dimensional point-by-point scanning.

[0017] The two-dimensional scanning system comprises a resonant mirror and a high-precision galvanometer mirror; wherein the resonant mirror is used to realize high-speed scanning in the x direction, and the working frequency is stabilized at the order of kilohertz (a typical value is 8 kHz), and high-efficiency cosine trajectory scanning is realized through mechanical resonance characteristics; the high-precision galvanometer mirror is used to realize programmable positioning scanning in the y direction. The two-axis scanning mechanism ensures the orthogonality of the light path and the flatness of the scanning plane through precise mechanical calibration. The two-dimensional scanning system is connected to the intelligent control system through a data acquisition card, an analog voltage signal is output to the resonant mirror through the data acquisition card, the resonant mirror is controlled to swing in a cosine waveform according to the set scanning angle at a fixed high frequency, and the fixed-range high-speed scanning in the x direction is completed; a triangular wave analog voltage signal is output to the galvanometer mirror through the data acquisition card, and the galvanometer mirror is controlled to complete the fixed-range scanning in the y direction. While the resonant mirror swings at a frequency of the order of kilohertz, a square wave signal synchronized with the mechanical swing is output in real time, and the rising edge and the falling edge correspond to the two end point positions of the scanning trajectory, respectively. After these synchronization signals are captured by the data acquisition card, they are used as row trigger signals to control the data acquisition card to send specified voltage signals to the galvanometer mirror according to the triangular wave form, so as to accurately trigger the stepping action of the galvanometer mirror in the y direction, thereby realizing strict time sequence cooperation and completing the row-by-row x scanning in the y direction.

[0018] The axial focusing adjustment is realized by a phase-type spatial light modulator, and the phase hologram calculated by the feedback-type non-diffraction axial scanning algorithm and adjusted in real time by the adaptive regulation method driven by deep learning is loaded on the spatial light modulator to control the axial focusing position of the laser along the z direction, realize non-mechanical focus regulation, and ensure the space-time consistency of three-dimensional data acquisition through synchronous cooperation with the above two-dimensional scanning. Using a kilohertz spatial light modulator, the focus position can be quickly switched without moving the sample or the objective lens. After the two-dimensional scanning system completes the two-dimensional scanning, a frame trigger signal is output to the spatial light modulator by the data acquisition card, and the spatial light modulator loads and refreshes a phase hologram after receiving the signal, which is used to change the axial focusing position of the laser along the z direction, i.e. switch to the next layer for scanning.

[0019] The high-fidelity acquisition depends on the control of the electrically tunable focus lens through the data acquisition card. After the focus is switched along the z direction each time, a trigger signal is output to the electrically tunable focus lens to accurately adjust the refractive power, so that the focus is always located on the detection surface of the photodetector, and distortion caused by defocusing is avoided. The above time sequence logic significantly improves the frame frequency of the system by using the high-speed characteristics of the resonant mirror; the delay caused by the motion inertia is avoided by the non-mechanical z direction adjustment; and the accurate registration of three-dimensional data sampling is ensured by the strict synchronization mechanism. These characteristics make the application particularly suitable for dynamic observation of living samples, and provide a powerful tool for biomedical research.

[0020] Another object of the present application is to provide an implementation method of a full-optical three-dimensional scanning confocal fluorescence microscopic imaging device.

[0021] The implementation method of the full-optical three-dimensional scanning confocal fluorescence microscopic imaging device of the present application comprises the following steps:

[0022] 1) The excitation light source emits a polarized Gaussian beam as excitation light into the depth adaptive control device;

[0023] 2) The spatial light modulator of the depth adaptive control device loads the phase hologram calculated by the feedback-free aberration axial scanning algorithm and adjusted in real time by the adaptive control method driven by deep learning to control the phase of the excitation light, achieving aberration-free axial movement of the focal point behind the objective lens;

[0024] 3) The first-order diffracted light is filtered out by the adjustable diaphragm after the dichroic mirror to the two-dimensional scanning system for two-dimensional plane optical scanning, and focused by the objective lens to the surface or internal observation area of the sample;

[0025] 4) The fluorescence signal generated on the surface or inside of the sample is collected by the same objective lens and transmitted back to the two-dimensional scanning system in the direction of the excitation light, and is descanned by the two-dimensional scanning system and separated from the excitation light by the dichroic mirror;

[0026] 5) The electrically adjustable focus lens adjusts the focal length according to the change of the axial position of the excitation light, so that the fluorescence signals at different axial positions in the sample can be focused on the detection surface of the photodetector;

[0027] 6) After passing through the pinhole located in front of the photodetector, the out-of-focus signals are filtered out and received by the photodetector;

[0028] 7) Through the intelligent control system, the adaptive control method driven by deep learning is used to obtain the relationship between the accurate phase hologram and the axial position, and to optimize the light intensity of the laser according to the axial position of the focal point, as well as to eliminate the inherent aberration of the optical system, dynamically control the phase distribution of the depth adaptive control device, the light intensity and the phase compensation of the laser, so that the excitation light can be aberration-free axially displaced behind the focal point of the high numerical aperture objective lens, and the bidirectional parallel optical scanning trajectory of the two-dimensional scanning system is optimized in real time, realizing full-optical three-dimensional scanning.

[0029] The deep learning driven adaptive control method is based on a neural network architecture constrained by a physical model, and realizes the autonomous optimization and real-time adjustment of the optical scanning system through end-to-end training. The neural network architecture adopts a multi-scale feature fusion convolutional neural network (CNN), and uses a convolutional neural network (CNN) to process spatial distribution data, optimize the light intensity of the laser and the phase hologram loaded by the feedback type non-axial scanning algorithm calculated by the spatial light modulator (SLM), and uses a multi-input and multi-output convolutional neural network. The inputs are three types of spatial data, the CNN extracts the local distortion features of the phase hologram, the sidelobe pattern of the light intensity distribution and the spatial distribution law of the phase error through the convolution layer, and outputs three types of static correction quantities after dimension compression by the pooling layer. The first type of input is the phase hologram loaded by the spatial light modulator (SLM), and the corresponding first type of output is the optimized spatial residual phase hologram, which is used to correct the nonlinear response of the SLM pixels caused by the residual error of the objective conditions such as devices or environmental conditions, to obtain the accurate relationship between the phase hologram and the axial position; the second type of input is the focal region light intensity distribution image, which is used to improve the focal light intensity consistency at different axial positions, and the corresponding second type of output is the light intensity distribution equalization parameter, which optimizes the corresponding light intensity of the laser according to the axial position of the focal point; the third type of input is the phase residual image, i.e. the difference distribution between the actual phase and the theoretical phase, which eliminates the inherent aberration of the optical system, and the corresponding third type of output is the static phase compensation matrix, which compensates the phase for the inherent aberration of different optical systems. The first type of data set of the convolutional neural network is obtained by loading the theoretical phase hologram covering the full phase range and different spatial frequency patterns to the SLM, and inversely solving the phase distribution by the camera probe. The phase distribution is considered as the actual phase, the difference between the actual phase and the theoretical phase is calculated, i.e. the SLM pixel nonlinear residual error, and the residual error compensation quantity is output, i.e. the phase hologram optimized for spatial residual error, and the data set covers the full pixel area of the SLM and different working temperatures.

[0030] The second type of data set of the convolutional neural network is obtained by controlling the SLM to load different phase holograms to realize axial scanning, using a camera to record the focal region light intensity distribution under different axial positions and different sample scattering characteristics, taking the actually detected light field intensity of the objective lens focal plane without axial displacement as the reference to calculate the light intensity deviation, and outputting the light intensity distribution equalization parameter by experimentally calibrating the mapping relationship between the light intensity deviation and the laser power adjustment quantity. The data set includes different sample types and scanning depths.

[0031] Different optical systems have different inherent phase differences; the third type of data set of the convolutional neural network is obtained by inputting the phase residual map to the SLM, detecting the actual three-dimensional intensity distribution at different axial depths and field positions with the camera, calculating the difference between the actual three-dimensional intensity distribution and the ideal light field distribution, further inversely solving the corresponding phase distribution, and outputting a static phase compensation matrix, the data set covering the full field of view and scanning depth of the objective lens; the test is performed in the actual optical system, so that the phase compensation corresponding to the inherent phase difference of different optical systems is performed.

[0032] The first to third types of training data sets are used to train the convolutional neural network respectively, and a trained convolutional neural network is obtained.

[0033] Further, the neural network architecture further includes a long short-term memory network (LSTM) and a time convolution network (TCN) to obtain a time sequence dependency relationship, and a laser light source, an adjustable diaphragm, an electrically adjustable focus lens, and a photoelectric detector are integrated to control the efficient synchronization and automatic operation of multiple devices and high-speed axial focusing adjustment. When the long short-term memory network (LSTM) is used to model the time sequence dependency relationship, the three types of static correction quantities output by the convolutional neural network CNN are arranged in time sequence respectively and used as the input of the LSTM, and a historical position error sequence reflecting a dynamic drift trend is also included; the LSTM captures the time sequence dependency relationship through a gating mechanism, such as the linear evolution of the phase drift caused by temperature change over time and the accumulation law of the position error in continuous scanning, and outputs dynamic correction parameters: a time sequence smoothed phase adjustment sequence, a real-time phase compensation increment, and a position prediction correction value.

[0034] The data set of the long short-term memory network (LSTM) is obtained by collecting two types of time sequence data as input at the scanning frame rate of the full-light three-dimensional scanning confocal fluorescence microscopic imaging device: one is the three types of static correction quantity sequences output by the CNN, including a phase hologram for optimizing spatial residual error, an intensity distribution equalization parameter, and a static phase compensation matrix, each sequence covering a complete scanning period; and the other is the time sequence data of the focal point position error recorded by the camera, the environmental temperature and the optical path vibration parameters are synchronously collected to ensure time alignment with the static correction quantity sequence; the time sequence sorting correction of the static correction quantity sequence output by the convolutional neural network is performed by using the time sequence data of the focal point position error, and three types of dynamic correction parameters are output: a time sequence smoothed phase adjustment sequence obtained by taking the inverse compensation amount of the phase mutation of adjacent frames, a real-time phase compensation increment calculated by measuring the phase residual, and a position prediction correction value obtained by predicting the position error and taking the inverse value. The input and output data are time-aligned and normalized, the data set is expanded by adding dynamic interference, and the training sample is formed for LSTM network training.

[0035] The time convolution network (TCN) models the time sequence dependency between devices, optimizes the control instruction sequence, and ensures accurate synchronization of each component; the dynamic correction parameters output by the LSTM, the power adjustment instructions of the laser light source, the aperture control signals of the adjustable diaphragm, the scanning trajectory instructions of the two-dimensional scanning system, the focal length parameters of the electrically adjustable focus lens, and the sampling frequency of the photodetector are collectively used as inputs of the time convolution network; the TCN models the time sequence dependency between multiple devices (such as the hysteresis characteristic of laser power change and light intensity response, and the synchronization accuracy of focus lens action and focal point position) through causal convolution and dilated convolution, optimizes the generation of the collaborative control instruction sequence, and outputs the multi-device collaborative control instruction sequence, thereby ensuring the accurate synchronization of the power of the laser light source, the aperture of the adjustable diaphragm, the scanning trajectory of the two-dimensional scanning system, the focal length of the electrically adjustable focus lens, and the photodetector and SLM phase modulation, and realizing efficient collaborative control of the laser light source, the adjustable diaphragm, the two-dimensional scanning system, the electrically adjustable focus lens, and the photodetector.

[0036] The data set of the time convolution network synchronously collects two types of time sequence data as inputs at a frame rate: one is the original control sequence of the laser light source power instruction, the adjustable diaphragm aperture signal, the electrically adjustable focus lens focal length parameter, and the photodetector sampling frequency; the other is the dynamic correction parameter output by the LSTM, which outputs the optimal control instruction sequence executable by each device, and forms a training sample for training of the time convolution network.

[0037] The network training adopts a physical constraint loss function to ensure that the prediction result conforms to the optical transmission law: L = αL MSE + βL Physics + γL Smoothness , wherein L MSE is a prediction error loss, L Physics is a Helmholtz equation constraint term, and L Smoothness is a phase smoothing regularization term. The adaptive regulation adopts a reinforcement learning framework, takes the scanning quality such as focal point sharpness, aberration compensation, and scanning efficiency as a reward signal, and dynamically adjusts the control parameters through a policy gradient method to adapt to different sample characteristics and environmental disturbances.

[0038] The feedback type aberration-free axial scanning algorithm calculates the phase hologram as follows: according to the vector light field diffraction theory, the process of focusing the incident light field through the objective lens is mainly a diffraction process; the intensity distribution of the incident light field on the spatial light modulator is I0, and the intensity distribution of the target light field expected to be obtained is I t ; the spatial light modulator is a pure phase type, and only the phase distribution needed to be modulated on the incident light field is obtained, and the corresponding target light field can be obtained. The feedback type aberration-free axial scanning algorithm of the present application is as follows: firstly, a random phase distribution Φ r is added to the incident light field with a known intensity distribution Li0 (I0,Φ r ) undergoes a first diffraction calculation to obtain a first calculated target light field U t1 (I t1 ,Φ t1 ), I t1 and Φ t1 are the intensity distribution and phase distribution of the first calculated target light field, respectively; after replacing the intensity distribution I t1 of the first calculated target light field with the desired intensity distribution I t of the target light field, an inverse diffraction calculation is performed to obtain a corresponding first inverse diffraction calculated incident light field L i1 (I i1 ,Φ i1 ), I i1 and Φ i1 are the intensity distribution and phase distribution of the first inverse diffraction calculated incident light field, respectively; after replacing the intensity distribution I i1 of the first calculated incident light field with the intensity distribution I0 of the incident light field, a second diffraction calculation is performed to obtain a second diffraction calculated target light field U t2 (I t2 ,Φ t2 ), I t2 and Φ t2 are the intensity distribution and phase distribution of the second diffraction calculated target light field, respectively; after replacing the intensity distribution I t2 of the second diffraction calculated target light field with the intensity distribution I t of the target light field, an inverse diffraction calculation is performed to obtain a corresponding second inverse diffraction calculated incident light field L i2 (I i2 ,Φ i2 ), I i2 and Φ i2 are the intensity distribution and phase distribution of the second inverse diffraction calculated incident light field, respectively; after repeating the above steps for k times, a kth diffraction calculated target light field U tk (I tk ,Φ tk ), I tk and Φ tk are the intensity distribution and phase distribution of the kth diffraction calculated target light field, and a kth inverse diffraction calculated incident light field L ik (I ik ,Φ ik ), I ik and Φ ik are the intensity distribution and phase distribution of the kth inverse diffraction calculated incident light field, k is a natural number ≥ 1, and the intensity distribution of the target light field is used as the evaluation standard for the feedback mechanism, and when the intensity distribution Itk the intensity distribution of the target light field I t the phase distribution Φ ik of the incident light field corresponding to the kth diffraction calculation is updated to a phase distribution Φ

[0039] Advantages of the present application:

[0040] The intelligent control system of the present application adopts a deep learning driven adaptive regulation method, and cooperates with a high numerical aperture objective lens to realize full-optical three-dimensional scanning confocal fluorescence microscopic imaging based on high-speed spatial light field regulation, effectively filters out out-of-focus light signals, significantly improves the image signal-to-noise ratio, axial resolution and system adaptability, and can realize three-dimensional reconstruction and high-contrast observation of dynamic processes of a sample; through dynamic learning of the spatial and time sequence characteristics of the nonlinear response of the SLM pixels and the system aberration by CNN and LSTM, the phase hologram is optimized in real time, the phase modulation precision and the light intensity of the excitation light are improved, the dynamic aberration caused by environmental disturbance is effectively offset, the signal-to-noise ratio is ensured, the light intensity of the high-sensitivity area is reduced and the excitation power of the weak signal area is improved, the photobleaching is reduced to prolong the observation time of the living sample; through TCN modeling of the time sequence dependence relationship between devices, the control instruction sequence is optimized to ensure accurate synchronization of each component, and the SLM phase modulation, resonant mirror scanning, adjustable focus lens focusing and photodetector sampling are performed synchronously, errors and time sequence mismatches are eliminated, different sample optical characteristics and environmental disturbances are automatically adapted, high-quality imaging is maintained, and the system robustness and applicable scene range are further improved; the resonant mirror and the spatial light modulator used in the present application are both kilohertz-level instruments, which can complete high-speed three-dimensional scanning imaging; in cooperation with the data acquisition card and the image acquisition card, integrated control of the two-dimensional scanning system, the spatial light modulator, the electrically adjustable focus lens and the photodetector and other electrically controlled instruments is completed, and high-speed, high-resolution and large three-dimensional volume biological imaging is realized. The present application uses an end-to-end feedback type aberration-free axial scanning algorithm to realize the technology of aberration-free axial large-range scanning under a high numerical aperture, which can successfully eliminate the phase difference and realize complete high-quality output compared with the method of realizing axial scanning by direct zooming; in the process of three-dimensional scanning of the sample, no mechanical movement is performed, and three-dimensional scanning is completed only by moving the focal point, which completely avoids the influence of motion artifacts on the final imaging effect, avoids the problem of slow imaging speed caused by mechanical movement, and greatly reduces the damage of photobleaching and phototoxicity to the sample. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a schematic diagram of an embodiment of the full-optical three-dimensional scanning confocal fluorescence microscopic imaging device of the present application;

[0042] Figure 2This is a timing control diagram of an embodiment of the all-optical three-dimensional scanning confocal fluorescence microscopy imaging device of the present invention;

[0043] Figure 3 The following is a result diagram of an embodiment of the all-optical three-dimensional scanning confocal fluorescence microscopy imaging device of the present invention, wherein (a) is a field of view diagram and (b) is a resolution diagram in each direction. Detailed Implementation

[0044] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0045] like Figure 1 As shown, the optical three-dimensional scanning confocal fluorescence microscopy imaging device of this embodiment includes: an excitation light source 1, a beam expander and collimator 2, a depth adaptive control device 3, a dichroic mirror 4, a 4f relay system 5, an adjustable aperture 5-1, a two-dimensional scanning system 6, a telecentric scanning correction system, a beam splitter 8, an objective lens 10, a sample holder 11, a white light source 12, a camera 13, a first reflecting mirror 14, a motorized adjustable focusing lens 15, a second reflecting mirror 16, a second tube mirror 17, a filter 18, a pinhole 19, a photodetector 20, and an intelligent control system; wherein, the two-dimensional scanning system 6 is connected to a data acquisition card, and the photodetector is connected to an image acquisition card; the excitation light source 1, the depth adaptive control device 3, the adjustable aperture 5-1, the data acquisition card, the motorized adjustable focusing lens 15, and the image acquisition card are respectively connected to the intelligent control system;

[0046] A 488 nm continuous output semiconductor laser is used as the excitation source 1 to emit a polarized Gaussian beam as the excitation light, which enters the beam expander and collimator 2.

[0047] The beam expander and collimator 2 expands and collimates the excitation light using a dual-lens system. It consists of a pair of lenses with different focal lengths precisely configured. A short focal length lens with a focal length of f1 serves as the input lens, and a long focal length lens with a focal length of f2 serves as the output lens. The beam expansion ratio is M = f2 / f1. The diameter of the expanded beam is precisely extended to 12 millimeters, which is matched with the liquid crystal panel of the spatial light modulator and transmitted to the depth adaptive control device 3.

[0048] The depth adaptive regulation device 3 includes a half-wave plate 3-1, a prism 3-2, and a kilohertz pure phase type silicon-based liquid crystal spatial light modulator 3-3, which together realize precise regulation of the light field; the half-wave plate modulates the polarization state of the incident excitation light, and by precisely rotating the angle of the half-wave plate, linearly polarized light is adjusted to a polarization state that matches the optimal working polarization direction of the spatial light modulator; a prism with an apex angle of 96° is used to fold the light path, so that the incident light beam forms a 6° angle with the modulator surface, and the prism is used to cooperate with the 1000Hz spatial light modulator to complete the light field regulation of the Gaussian light after beam expansion and collimation and polarization state modulation; the Gaussian light is reflected by one side of the prism into the spatial light modulator, and the spatial light modulator changes the axial focusing position of the excitation light after the objective lens 10 by loading a phase hologram, thereby realizing light field regulation and completing high-speed axial optical scanning; the phase of the excitation light is regulated by loading a phase hologram calculated by a feedback type non-diffractive axial scanning algorithm and driven by a deep learning adaptive regulation method, so as to move the focal point axially without diffraction after the objective lens 10;

[0049] After passing through the dichroic mirror 4, it reaches the 4f relay system 5, which includes two coaxially arranged input and output lenses, and relays the plane modulated by the spatial light modulator to the two-dimensional scanning system 6;

[0050] The adjustable diaphragm 5-1 is located at the rear focal plane of the input lens and the front focal plane of the output lens, and is used to select the first-order diffracted light with the highest diffraction efficiency and complete phase modulation information as the excitation light to the two-dimensional scanning system 6;

[0051] The two-dimensional scanning system 6 includes a resonant mirror and a high-precision galvanometer mirror; wherein, as Figure 2As shown, the x direction uses a resonant galvanometer to achieve high-speed scanning, and a constant voltage is used, and the voltage determines the scanning angle, and the working frequency is 7910 Hz, and high-efficiency cosine trajectory scanning is achieved through mechanical resonance characteristics; the y direction uses a high-precision galvanometer, and the galvanometer driving signal is a 90% duty cycle triangular wave, and the working frequency is set to 1000 Hz; an analog voltage signal is output to the resonant galvanometer through the data acquisition card, the resonant galvanometer is controlled to swing in a cosine waveform according to the set scanning angle at a frequency of 8 kHz, and high-speed scanning in a fixed range in the x direction is completed; a triangular wave analog voltage signal is output to the galvanometer through the data acquisition card, and the galvanometer is controlled to complete scanning in a fixed range in the y direction; while the resonant galvanometer is swinging at a frequency of kHz, a square wave signal synchronized with the mechanical swinging is output in real time, and the rising and falling edges correspond to the two end point positions of the scanning trajectory respectively; after these synchronization signals are captured by the data acquisition card, they are used as row trigger signals to control the data acquisition card to send specified voltage signals to the galvanometer according to the triangular wave waveform, so as to accurately trigger the stepping action of the galvanometer in the y direction, thereby realizing strict timing cooperation and completing the row-by-row x scanning in the y direction; taking a typical 128*128 sampling point scanning as an example, the system performs an "S" shaped scanning trajectory: after outputting a voltage signal to the resonant galvanometer, the resonant galvanometer in the x direction quickly completes the forward scanning of the first row from (1, 1) to (1, 128), and the edge signal generated at the end point position indicates the data acquisition card to send a specified voltage signal to the galvanometer, so that the galvanometer swings an angle, triggers the scanning position to move a step in the y direction, and switches to the next row for scanning; then the resonant galvanometer in the x direction immediately performs reverse scanning from (2, 128) to (2, 1), and generates an edge signal at the end point position to complete the second row scanning sampling; in this way, the two-dimensional scanning system 6 cooperates to complete a 128*128 sampling point two-dimensional plane high-speed "S" shaped scanning under the control of the data acquisition card; after the two-dimensional scanning system 6 completes the two-dimensional scanning, a frame trigger signal will be output to the spatial light modulator by the data acquisition card, and the spatial light modulator receives the signal to load and refresh a phase hologram, which is used to change the axial focusing position of the laser along the z direction, that is, switch to the next layer for scanning to complete the z direction scanning; the embodiment completes a typical 128*128*100 sampling point three-dimensional scanning only in 800 ms, and the system scanning frame rate can reach 125 Hz, and the intelligent control system optimizes the bidirectional parallel optical scanning trajectory of the two-dimensional scanning system 6 in real time, compared with the traditional confocal microscope, the number of orders is improved, and mechanical movement is not required;

[0052] The telecentric scanning correction system comprises a scanning lens 7 and a first tube lens 9. The light passing through the two-dimensional scanning system 6 first passes through the scanning lens 7 and then passes through the first tube lens 9, ensuring that the light is relayed from the two-dimensional scanning system 6 to the back focal plane of the objective lens 10. The telecentric scanning correction system enables the system to maintain a highly consistent spot size throughout the entire scanning field of view, thereby ensuring the uniformity of the imaging resolution. This design not only ensures high-quality scanning imaging, but also maximizes the reduction of the need for post-processing, thereby improving the overall efficiency of the system. If the telecentric scanning correction system is not used, the scanning beam will produce significant trapezoidal distortion and field curvature, resulting in positional deviation in the edge region. Moreover, it is unable to maintain a constant magnification, and the magnification will fluctuate when the scanning angle increases, causing measurement errors. Off-axis aberrations will worsen the spot quality, increase the spot size, and cause uneven energy distribution, thereby reducing the imaging resolution.

[0053] The excitation light passing through the telecentric scanning correction system is focused by the objective lens 10 to the sample.

[0054] The white light source 12 emits white illumination light, which uniformly irradiates the sample surface. The optical signal is accurately transmitted to the first tube lens 9 after being collected by the objective lens 10 and reaches the detection surface of the camera 13 through the beam splitter 8. The camera 13 collects the reflected light signal from the sample surface to obtain the two-dimensional plane information of the sample surface.

[0055] The sample is placed on the sample holder 11. The fluorescent signal generated on or inside the sample surface is collected by the same objective lens 10 and is transmitted in the opposite direction of the excitation light to the two-dimensional scanning system 6, where it is descanned and separated from the excitation light by the dichroic mirror 4. The electrically adjustable focus lens 15 is controlled by the intelligent control system according to the change in the axial position of the excitation light. The electrically adjustable focus lens is controlled with high precision and synchronization to ensure that the focal point fluorescent signal of different axial depths of the sample is relayed from the front focal plane of the objective lens to the back focal plane of the second tube lens 17. That is, the fluorescent signal at different axial positions of the sample can be focused on the detection surface of the photodetector 20. The electrically adjustable focus lens 15 is placed between the first mirror 14 and the second mirror 16, which are parallel to each other, so that the electrically adjustable focus lens 15 is placed horizontally. The stray light is removed by the optical filter 18, and only the signal light is retained. The out-of-focus signal is filtered out by the pinhole 19 located in front of the photodetector 20, and then received by the photodetector 20.

[0056] The photoelectric detector 20 comprises a focusing lens, a pinhole 19 and a photomultiplier tube. The pinhole 19 allows only the in-focus light rays at the focal plane of the objective lens 10 to pass through the pinhole 19 into the photomultiplier tube to participate in imaging, while the out-of-focus fluorescent signals are blocked outside the detection pinhole 19. The photomultiplier tube receives the weak fluorescent signals from the focal point of the objective lens 10. Finally, the three-dimensional structure of the sample is reconstructed through three-dimensional point-by-point scanning.

[0057] The deep learning driven adaptive control method is based on a neural network architecture constrained by a physical model, and realizes the autonomous optimization and real-time adjustment of the optical scanning system through end-to-end training. The neural network architecture adopts a multi-scale feature fusion convolutional neural network (CNN). The spatial distribution data is processed by a convolutional neural network (CNN), and the light intensity of the laser and the phase hologram loaded by the spatial light modulator (SLM) calculated by the feedback type non-diffractive axial scanning algorithm are optimized. A multi-input and multi-output convolutional neural network is adopted, and the inputs are three types of spatial data. The CNN extracts the local distortion features of the phase hologram, the sidelobe patterns of the light intensity distribution and the spatial distribution rules of the phase error through the convolution layer, and outputs three types of static correction quantities after dimension compression by the pooling layer. The first type of input is the phase hologram loaded by the spatial light modulator (SLM), and the corresponding first type of output is the phase hologram optimized for spatial residual error, which is used to correct the nonlinear response of the SLM pixels caused by the residual error due to the objective conditions of the device or the environment, to obtain the accurate relationship between the phase hologram and the axial position. The second type of input is the focal region light intensity distribution image, which is used to improve the focal point light intensity consistency at different axial positions, and the corresponding second type of output is the light intensity distribution equalization parameter, which optimizes the corresponding light intensity of the laser according to the axial position of the focal point. The third type of input is the phase residual image, i.e. the difference distribution between the actual phase and the theoretical phase, which eliminates the inherent aberration of the optical system, and the corresponding third type of output is the static phase compensation matrix, which compensates the phase for the inherent aberration of different optical systems. The first data set of the convolutional neural network is obtained by loading the theoretical phase hologram covering the full phase range and different spatial frequency patterns to the SLM, and inversely solving the phase distribution from the light field intensity distribution detected by the camera 13. The phase distribution is considered as the actual phase, the difference between the actual phase and the theoretical phase, i.e. the SLM pixel nonlinear residual error, is calculated, and the residual error reverse compensation quantity, i.e. the phase hologram optimized for spatial residual error, is output. The data set covers the full pixel area of the SLM and different working temperatures.

[0058] The second type of data set of the convolutional neural network is realized by controlling the SLM to load different phase holograms for axial scanning, and the camera 13 is used to record the light intensity distribution of the focal region under different axial positions and different sample scattering characteristics, so as to calculate the light intensity deviation based on the actually detected light field intensity of the objective lens 10 focal plane without axial displacement. The mapping relationship between the light intensity deviation and the laser power adjustment amount is calibrated through experiments, and the light intensity distribution equalization parameter is output. The data set includes different sample types and scanning depths.

[0059] Different optical systems have different inherent phase differences; the third type of data set of the convolutional neural network is realized by inputting the phase residual map to the SLM, and the camera 13 is used to detect the actual three-dimensional intensity distribution at different axial depths and field positions, and calculate the difference between the ideal light field distribution. The difference is further inversely calculated to obtain the corresponding phase distribution, and the output is a static phase compensation matrix. The data set covers the full field of view and scanning depth of the objective lens 10; and the actual optical system is tested, so as to compensate for the inherent phase difference of different optical systems.

[0060] The first to third types of training data sets are used to train the convolutional neural network respectively, and the trained convolutional neural network is obtained.

[0061] Further, the neural network architecture further includes a long short-term memory network (LSTM) and a time convolution network (TCN) to obtain a time sequence dependency relationship, and integrates control of a laser light source, an adjustable diaphragm, an electrically adjustable focus lens 15, and a photoelectric detector to control efficient synchronization and automatic operation of multiple devices and high-speed axial focusing adjustment. When the long short-term memory network (LSTM) is used to model the time sequence dependency relationship, the three types of static correction amounts output by the convolutional neural network CNN are arranged in time sequence respectively and used as the input of the LSTM, and a historical position error sequence reflecting a dynamic drift trend is also included. The LSTM captures the time sequence dependency relationship through a gating mechanism, such as the linear evolution of the phase shift caused by temperature change over time and the accumulation law of the position error in continuous scanning, and outputs dynamic correction parameters: a time sequence smoothed phase adjustment sequence corrects the dynamic phase deviation to suppress the light intensity mutation of adjacent positions, a real-time phase compensation increment ensures the time sequence consistency of the light intensity, that is, tracks the dynamic error, and a position prediction correction value offsets the time sequence cumulative deviation.

[0062] The dataset of long short-term memory network (LSTM) collects two kinds of time series data as input by scanning the frame rate of the system: one is the three-class static correction sequence output by CNN, including the phase hologram of optimized spatial residual error, the light intensity distribution equalization parameter, and the static phase compensation matrix, each sequence covering the entire scanning period; the second is the focal point position error time series data recorded by camera 13, which synchronously collects environmental temperature and light path vibration parameters to ensure time alignment with the static correction sequence; the static correction sequence output by the convolutional neural network is corrected in time sequence by using the focal point position error time series data, and the output is three kinds of dynamic correction parameters: the time series smoothed phase adjustment sequence obtained by taking the adjacent frame phase mutation reverse compensation amount, the real-time phase compensation increment calculated by the measured phase residual error, and the position prediction correction value obtained by predicting the position error and taking the inverse value. The input and output data are time-aligned and normalized, the dataset is expanded by adding dynamic interference, and the training samples are formed for LSTM network training.

[0063] The time convolution network (TCN) models the time sequence dependency relationship between devices, optimizes the control instruction sequence, and ensures accurate synchronization of each component; the dynamic correction parameters output by the LSTM and the power adjustment instructions of the laser light source, the aperture control signals of the adjustable diaphragm, the focal length parameters of the electrically adjustable focus lens, and the sampling frequency of the photodetector are collectively used as the input of the time convolution network; the TCN models the time sequence dependency relationship between multiple devices (such as the hysteresis characteristic of laser power change and light intensity response, and the synchronization accuracy of focus lens action and focal point position) through causal convolution and dilated convolution, optimizes the generation of collaborative control instruction sequence, and outputs multiple device collaborative control instruction sequence, to ensure the accurate synchronization of the power of the laser light source, the aperture of the adjustable diaphragm, the scanning trajectory of the two-dimensional scanning system, the focal length of the electrically adjustable focus lens, and the phase modulation of the photodetector and SLM, to realize efficient collaborative control of the laser light source, the adjustable diaphragm, the two-dimensional scanning system, the electrically adjustable focus lens, and the photodetector. The time sequence control diagram is shown in Figure 2 .

[0064] The dataset of the time convolution network synchronously collects two kinds of time series data as input by frame rate: one is the original control sequence of the laser light source power instruction, the adjustable diaphragm aperture signal, the electrically adjustable focus lens focal length parameter, and the photodetector sampling frequency; the second is the dynamic correction parameter output by the LSTM, which outputs the optimal control instruction sequence executable by each device, to form a training sample for training the time convolution network.

[0065] The network training adopts a physical constraint loss function to ensure that the prediction result conforms to the optical transmission law: L = aL MSE + βL Physics + γL Smoothness , wherein, L MSE is the prediction error loss, L PhysicsL is the constraint term of Helmholtz equation Smoothness is the phase smoothing regularization term. Wherein the adaptive control adopts a reinforcement learning framework, taking scanning quality such as focal sharpness, aberration compensation, scanning efficiency, etc. as reward signals, and dynamically adjusting the control parameters through policy gradient method to adapt to different sample characteristics and environmental disturbances.

[0066] As shown in Figure 3 the field of view range and three-dimensional resolution of the present application can be calibrated by imaging fluorescent microspheres, (a) is the result image of imaging 5 μm fluorescent microspheres using the full light three-dimensional scanning confocal fluorescence microscopic imaging device based on high-speed spatial light field regulation, and the XY direction imaging field of view range is 186.18 μm x 186.18 μm; (b) is the light intensity full width at half maximum image obtained by imaging 100 nm microspheres using the method of the present application, and the X, Y, Z directions are 202 nm, 221 nm and 554 nm respectively. The experimental results show that under the condition of maintaining sufficient field of view coverage and near-diffraction-limited resolution, the technical scheme of the present application realizes significant three-dimensional imaging speed improvement compared with the traditional confocal microscopic system, and expands the three-dimensional imaging range of the full light scanning confocal microscopic technology. This breakthrough mainly benefits from the cooperative working mode of the innovative non-mechanical axial scanning scheme and the high-speed resonant mirror, which enables the system to meet the time sequence requirements of dynamic observation of living samples while maintaining the subcellular level spatial resolution.

[0067] Finally, it should be noted that the purpose of the disclosed embodiments is to help further understand the present application, but those skilled in the art can understand that various replacements and modifications are possible without departing from the spirit and scope of the present application and the appended claims. Therefore, the present application should not be limited to the disclosed content, and the scope of protection claimed by the present application is subject to the scope defined by the claims.

Claims

1. A full-light three-dimensional scanning confocal fluorescence microscopy imaging device, characterized in that, The fluorescence microscopy imaging device includes: an excitation source, a depth adaptive control device, a dichroic mirror, an adjustable aperture, a two-dimensional scanning system, a telecentric scanning correction system, an objective lens, a motorized adjustable focusing lens, a pinhole camera, a photodetector, and an intelligent control system; wherein, the two-dimensional scanning system is connected to a data acquisition card, and the photodetector is connected to an image acquisition card; the excitation source, the depth adaptive control device, the adjustable aperture, the data acquisition card, the motorized adjustable focusing lens, and the image acquisition card are respectively connected to the intelligent control system; The excitation source emits a polarized Gaussian beam as excitation light, which enters the depth adaptive control device. The phase of the excitation light is controlled by loading a phase hologram. After passing through a dichroic mirror, the first-order diffracted light is filtered out by an adjustable aperture and sent to the two-dimensional scanning system for two-dimensional planar optical scanning. The light is then focused onto the sample by the objective lens. The sample generates a fluorescence signal, which is collected by the same objective lens and transmitted back to the two-dimensional scanning system in the opposite direction of the excitation light. The two-dimensional scanning system performs a descan and splits the excitation light beam by a dichroic mirror. The electric adjustable focusing lens adjusts the focal length according to the change of the excitation light along the axial position, so that the fluorescence signals at different axial positions in the sample can be focused on the detection surface of the photodetector. After passing through the pinhole, the fluorescence is received by the photodetector. The intelligent control system adopts a deep learning-driven adaptive control method to optimize the phase hologram calculated by the feedback aberration-free axial scanning algorithm to obtain an accurate relationship between the phase hologram and the axial position. It also optimizes the laser intensity based on the axial position of the focal point and performs phase compensation to achieve aberration-free axial displacement of the excitation light at the focal point behind the objective lens. Furthermore, it optimizes the bidirectional parallel optical scanning trajectory of the two-dimensional scanning system in real time to achieve all-optical three-dimensional scanning.

2. The fluorescence microscopy imaging device as described in claim 1, characterized in that, It also includes a beam expander and collimator, which uses a dual-lens system for beam expansion and collimation. The diameter of the expanded beam is matched with the spatial light modulator of the depth adaptive control device.

3. The fluorescence microscopy imaging device as described in claim 1, characterized in that, The depth adaptive control device includes a half-wave plate, a prism, and a spatial light modulator. The half-wave plate modulates the polarization state of the incident excitation light. By precisely rotating the angle of the half-wave plate, the linearly polarized light is adjusted to a polarization state that matches the optimal working polarization direction of the spatial light modulator. The prism, in conjunction with the spatial light modulator, completes the optical field control of the Gaussian light. The Gaussian light is reflected by one side of the prism and incident on the spatial light modulator. The spatial light modulator changes the axial position of the excitation light focused behind the objective lens by loading a phase hologram.

4. The fluorescence microscopy imaging device as described in claim 1, characterized in that, It also includes a telecentric scanning correction system, which is disposed between the two-dimensional scanning system and the objective lens; it includes a scanning lens and a first tube mirror, and the light passing through the two-dimensional scanning system first passes through the scanning lens and then through the first tube mirror, ensuring that the light is relayed from the two-dimensional scanning system to the back focal plane of the objective lens.

5. The fluorescence microscopy imaging device as described in claim 1, characterized in that, The two-dimensional scanning system includes a resonant galvanometer and a high-precision galvanometer galvanometer. The resonant galvanometer is used in the x-direction to achieve high-speed scanning with a stable operating frequency in the kilohertz range. It achieves high-efficiency cosine trajectory scanning through mechanical resonance characteristics. The high-precision galvanometer galvanometer is used in the y-direction to achieve programmable positioning scanning.

6. A method for implementing the all-optical three-dimensional scanning confocal fluorescence microscopy imaging device as described in claim 1, characterized in that, The implementation method includes the following steps: 1) The excitation source emits a polarized Gaussian beam as the excitation light, which enters the depth adaptive control device; 2) The spatial light modulator of the depth adaptive control device loads a phase hologram calculated by a feedback aberration-free axial scanning algorithm and adjusted in real time by an adaptive control method driven by deep learning to control the phase of the excitation light, thereby achieving aberration-free axial focus movement behind the objective lens. 3) After passing through a dichroic mirror, the first-order diffracted light is filtered through an adjustable aperture to a two-dimensional scanning system for two-dimensional planar optical scanning. The objective lens focuses the light onto the surface or internal area of ​​the sample to be observed. 4) Fluorescent signals are generated on or inside the sample surface. They are collected by the same objective lens and transmitted back to the two-dimensional scanning system in the opposite direction of the excitation light. The two-dimensional scanning system performs the descanning and splits the excitation light beam by a dichroic mirror. 5) The focal length is adjusted by an electrically adjustable lens according to the change of the position of the excitation light along the axis, so that the fluorescence signals at different axial positions in the sample can be focused on the detection surface of the photodetector. 6) After passing through the pinhole located in front of the photodetector, the out-of-focus signal is filtered out and then received by the photodetector; 7) Through the intelligent control system, the relationship between the accurate phase hologram and the axial position is obtained by adopting the deep learning-driven adaptive control method. The laser intensity is optimized according to the axial position of the focal point, and the inherent aberrations of the optical system are eliminated. The phase distribution of the depth adaptive control device, the laser intensity and phase compensation are dynamically controlled, so that the excitation light has an aberration-free axial displacement at the focal point of the high numerical aperture objective lens. The bidirectional parallel optical scanning trajectory of the two-dimensional scanning system is optimized in real time to realize all-optical three-dimensional scanning.

7. The implementation method as described in claim 6, characterized in that, The deep learning-driven adaptive control method is based on a neural network architecture constrained by a physical model. The neural network architecture employs a multi-input multi-output (MIMO) convolutional neural network (CNN). The input consists of three types of spatial data. The CNN extracts local distortion features of the phase hologram, sidelobe patterns of the light intensity distribution, and the spatial distribution of phase error through convolutional layers. After dimensionality compression via pooling layers, the output consists of three types of static corrections. The first type of input is the phase hologram loaded by the spatial light modulator (SLM), and the corresponding first type of output is a phase hologram with optimized spatial residuals. This corrects the nonlinear response of SLM pixels caused by residuals due to objective conditions of the device or environment, obtaining an accurate relationship between the phase hologram and the axial position. The second type of input is the light intensity distribution image of the focal region, used to improve the consistency of focal light intensity at different axial positions. The corresponding second type of output is a light intensity distribution equalization parameter, optimizing the corresponding light intensity of the laser based on the axial position of the focal point. The third type of input is the phase residual map, i.e., the distribution of the difference between the actual phase and the theoretical phase, eliminating inherent aberrations of the optical system. The corresponding third type of output is a static phase compensation matrix, performing corresponding phase compensation for the inherent phase aberrations of different optical systems.

8. The implementation method as described in claim 7, characterized in that, The neural network architecture also includes a Long Short-Term Memory (LSTM) network and a Temporal Convolutional Network (TCN) to obtain temporal dependencies. It integrates control of the laser source, adjustable aperture, two-dimensional scanning system, motorized adjustable focus lens, and photodetector to control the efficient synchronization and automated operation of multiple devices and high-speed axial focusing adjustment.

9. The implementation method as described in claim 8, characterized in that, The dataset for the Long Short-Term Memory (LSTM) network is obtained by collecting two types of temporal data at the scanning frame rate of an all-optical 3D scanning confocal fluorescence microscopy device: one is the three types of static correction sequences output by the CNN, including phase holograms for optimizing spatial residuals, light intensity distribution equalization parameters, and static phase compensation matrices, with each sequence covering the entire scanning cycle; the other is the focal position error temporal data recorded by the camera, with ambient temperature and optical path vibration parameters collected simultaneously to ensure time alignment with the static correction sequence. The focal position error temporal data is used to temporally sort and correct the static correction sequence output by the convolutional neural network, and the output is three types of dynamic correction parameters: a temporally smoothed phase adjustment sequence obtained by taking the inverse compensation amount of phase abrupt changes in adjacent frames, a real-time phase compensation increment calculated by the measured phase residuals, and a position prediction correction value obtained by predicting the position error and taking the inverse value.

10. The implementation method as described in claim 9, characterized in that, The Temporal Convolutional Network (TCN) models temporal dependencies and optimizes control command sequences. It uses the dynamic correction parameters output from the LSTM, along with the power adjustment commands of the laser source, the aperture control signal of the adjustable aperture, the scanning trajectory commands of the 2D scanning system, the focal length parameters of the motorized adjustable lens, and the sampling frequency of the photodetector, as inputs to the TCN. Through causal convolution and dilated convolution, the TCN models the temporal dependencies between multiple devices, optimizes the generation of collaborative control command sequences, and outputs a multi-device collaborative control command sequence. This ensures precise synchronization of the laser source power, the aperture of the adjustable aperture, the scanning trajectory of the 2D scanning system, the focal length of the motorized adjustable lens, and the phase modulation of the photodetector and SLM, achieving efficient collaborative control of the laser source, adjustable aperture, 2D scanning system, motorized adjustable lens, and photodetector.

Images