A method and system for adaptive illumination imaging of live cells

Abstract

The present application relates to the technical field of live cell super-resolution optical imaging, and particularly relates to a live cell adaptive illumination imaging method and system. The method comprises the following steps: building a live cell adaptive illumination imaging system, including an illumination imaging module, an image sensing module, a decision control module and an imaging acquisition module; performing initialization and calibration; through image sensing feature extraction and photochemical kinetics model calculation, establishing an illumination-imaging-light damage coupling model, and extracting a seven-dimensional state vector of the cell; executing deep reinforcement learning reasoning, and outputting an optimal illumination control instruction; driving the high-speed stripe generation unit and the nanometer precision phase control unit in the illumination imaging module to work cooperatively, completing structured light illumination and image acquisition; performing super-resolution image reconstruction and iterative optimization output, and completing long-term low-light toxicity live cell imaging. The advantage lies in realizing real-time closed-loop feedback of illumination parameters, significantly improving photon utilization, reducing light toxicity, and improving response speed and long-term stable imaging.

Description

Technical Field

[0001] This invention relates to the field of live-cell super-resolution optical imaging technology, and in particular to a live-cell adaptive illumination imaging method and system. Background Technology

[0002] Dynamic analysis of subcellular structures in living cells is a core frontier for revealing the essence of life and elucidating disease mechanisms. In research on malignant tumors such as leukemia, excessive fragmentation caused by the imbalance of mitochondrial fission-fusion homeostasis has been proven to be a key cellular event driving chemotherapy resistance and relapse. However, its dynamic regulatory network and signaling causal chain remain unclear, with the fundamental bottleneck being the lack of reliable techniques for long-term, high-fidelity dynamic observation of mitochondria.

[0003] Currently, super-resolution live-cell dynamic imaging is mainly based on structured illumination microscopy (SIM), stimulated emission depletion microscopy (STED), and single-molecule localization microscopy (SMLM). Among these, STED suffers from the physical contradiction that "high resolution inevitably comes with high phototoxicity," making it difficult to avoid cumulative damage to photosensitive cells. SMLM is trapped in the fundamental dilemma of "mutual exclusion between ultra-high spatial resolution and high temporal resolution," failing to capture crucial millisecond-level dynamic processes. SIM, with its good balance between imaging speed and phototoxicity, has become one of the mainstream technologies for long-term, live-cell dynamic imaging.

[0004] Existing structured illumination microscopy (SIM) systems generally employ a preset illumination mode, where parameters such as the direction, frequency, and phase of the illumination fringes are pre-set and remain fixed during imaging. This fixed illumination mode cannot match the rapid and random morphological changes and motion states of cells, resulting in low photon utilization efficiency, numerous ineffective illuminations, and significant light damage, severely limiting the observation time of live cells. Furthermore, due to the lack of a mechanism to adjust illumination parameters based on real-time sample status feedback, the system struggles to continuously acquire high signal-to-noise ratio dynamic images under low phototoxicity conditions. At the structured illumination level, existing SIM technologies all employ a "preset illumination" mode, meaning that illumination parameters (fringe direction, frequency, and phase) are fixed before imaging and cannot be adjusted in real-time according to dynamic cell changes. Its fundamental flaw lies in the fact that a fixed illumination wavefront cannot match the rapid and random morphological and motion changes of cells, leading to low photon utilization efficiency. Numerous ineffective illuminations not only fail to effectively excite valid signals but also cause unnecessary light damage, severely limiting the observation time. Meanwhile, due to the lack of an intelligent decision-making mechanism that modulates illumination parameters in real time based on sample status, the imaging system struggles to continuously acquire dynamic image sequences with high signal-to-noise ratios under low phototoxicity conditions, creating a vicious cycle of mutual constraints between "illumination mode - optical damage - imaging quality." At the imaging system level, existing SIM technologies mostly employ structured light generation schemes based on spatial light modulators (SLMs). Although SLMs can flexibly generate various patterns, their refresh rates are limited (typically tens of hertz, with inter-frame intervals ranging from milliseconds to tens of milliseconds), making it difficult to meet the demands of millisecond-level dynamic processes in live cells for high-speed switching of illumination modes. Furthermore, the low light energy utilization of SLMs further exacerbates the light dose burden. In addition, while existing structured light illumination systems based on interferometric imaging (such as those using Mach-Zehnder interferometers) can produce high-quality interference fringes, their phase stability is susceptible to environmental vibrations and temperature drift, and they lack a real-time control mechanism linked to sample dynamics, remaining within the scope of open-loop control.

[0005] Specifically, existing technologies disclose a super-resolution microscopy imaging method and system based on active structured light illumination. The core of this method lies in employing spatiotemporal joint intensity modulation technology, which allows for arbitrary spatial intensity control of the excitation light on the illuminated sample without altering the coherence of the incident light, achieving high-modulation active structured light illumination super-resolution imaging. However, in this scheme, the control of all illumination parameters (such as the intensity distribution of the flat-top illumination, the spatial mapping of adaptive light intensity, and the shape and position of the custom region) is preset before imaging and remains constant during the imaging process. Although it achieves the preset capability of "active control," it essentially still belongs to the "preset illumination" technical paradigm and fails to achieve closed-loop feedback modulation of illumination parameters based on real-time dynamic changes in cells. This limitation is essentially the same as the preset illumination paradigm of typical SIM systems.

[0006] To address the aforementioned issues, existing technical solutions primarily focus on hardware acceleration (such as using faster cameras or SLM) or post-processing algorithm optimization (such as image denoising and super-resolution reconstruction). However, none of these solutions fundamentally change the preset illumination, failing to address the core contradictions of mismatch between illumination parameters and cell dynamics, and mismatch between hardware response speed and dynamic processes. In summary, existing imaging systems remain in an open-loop control state in the "perception-decision-execution" chain. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides a live-cell adaptive illumination imaging method and system.

[0008] The primary objective of this invention is to provide a live-cell adaptive illumination imaging method, which specifically includes the following steps: S1. Construct a live-cell adaptive illumination imaging system and perform system initialization and calibration; the live-cell adaptive illumination imaging system includes an illumination imaging module, an image perception module, a decision control module, and an imaging acquisition module; S2. By extracting image-perceptual features and calculating photochemical dynamics models, an illumination-imaging-photodamage coupling model is established to extract a seven-dimensional cell state vector; S3. Based on the decision control module, perform deep reinforcement learning inference on the seven-dimensional state vector to output the optimal lighting control command; S4. According to the optimal illumination control command, drive the high-speed stripe generation unit and the nano-precision phase modulation unit in the illumination imaging module to work together to complete structured light illumination and image acquisition; S5. Perform super-resolution image reconstruction and iterative optimization on the acquired images to complete long-term low-phototoxicity live cell imaging.

[0009] Preferably, the illumination imaging module includes a laser, a collimator, a beam expander, a high-speed fringe generation unit, a nanometer-precision phase modulation unit, and a long-term active stabilization unit. The high-speed stripe generation unit includes a non-polarizing beam splitter prism, a first high-speed scanning mirror, a first reflecting mirror, a second reflecting mirror, a second high-speed scanning mirror, and a beam combiner; the nanometer-precision phase control unit includes a high-precision spatial light delay line, a capacitive displacement sensor, and a digital PID controller, with the high-precision spatial light delay line disposed in one of the optical paths of the high-speed stripe generation unit; After the laser beam is collimated by the collimator, it is expanded by the incident beam expander. The beam is split into two paths by the non-polarizing beam splitter. One path passes through the first high-speed scanning galvanometer and the first reflecting mirror and is then incident on the beam combiner. The other path passes through the second high-speed scanning galvanometer and the second reflecting mirror and is then incident on the beam combiner. The two beams are combined by the beam combiner to form interference structured light and then output to the subsequent optical path.

[0010] Preferably, the first high-speed scanning galvanometer and the second high-speed scanning galvanometer are synchronously and collaboratively controlled, and the phase difference between the two beams is adjusted in real time by a nanometer-precision phase control unit, so that the spatial frequency and direction adjustment response time of the structured light stripes is no higher than 1ms.

[0011] Preferably, the long-term active stabilization unit is used to monitor the intensity fluctuation and phase drift of the interference structured light in real time, and forms a closed-loop feedback through a digital PID controller to make the phase drift of the interference fringes less than λ / 50.

[0012] Preferably, step S2 specifically includes: inputting the live cell fluorescence image acquired in real time by the imaging acquisition module into the perception branch of the image perception module, performing real-time feature extraction and analysis on the input cell image; simultaneously, combining the pre-constructed photochemical dynamics model, recursively calculating the cell damage state under the current illumination conditions based on real-time illumination parameters and historical illumination records, and finally fusing the dynamic features obtained from image perception with the damage state calculated by the model to construct a unified seven-dimensional state vector, and outputting it to the decision control module; The seven-dimensional state vector is represented as: ; in, Indicates the average velocity of the cell. Indicates the spatial coherence length. Indicates the cumulative light dose. C f ( t () indicates the concentration of the fluorophore. C R ( t Let be the concentration of reactive oxygen species at time t. This represents the cell viability predicted by the model. It indicates the trend of changes in cell viability.

[0013] Preferably, the perceptual branch is a lightweight convolutional neural network based on MobileNetV3-Small; The photochemical kinetic model includes: Fluorescent probe bleaching kinetic model: ; Kinetic model of reactive oxygen species generation and scavenging: ; Cell viability decay kinetic model: ; in, C f ( t () represents the concentration of the fluorophore. t Indicates time, C R ( t Let be the concentration of reactive oxygen species at time t.V For relative cell viability, I ( t ) represents the light intensity, and σ represents the photon absorption cross section. For quantum yield, k g The rate of reactive oxygen species generation. k c This represents the cellular rate constant for scavenging reactive oxygen species. k d The damage coefficient is denoted as .

[0014] Preferably, the decision control module is configured with a decision branch, which adopts a three-layer fully connected network structure with the number of neurons set to 128, 64 and 3 respectively, and is used to receive the seven-dimensional state vector and output the normalized optimal lighting parameter combination.

[0015] Preferably, the decision control module sets a multi-objective weighted reward function, incorporating imaging information gain, optical damage suppression, illumination stability, and system stability into a unified optimization objective to achieve synergistic optimization of imaging quality and cell protection; the instantaneous reward value expression at time t is as follows: ; in, α , β , c , d These are the weighting coefficients; The instantaneous reward value at time t; the first term The first term is the information gain term, which incentivizes image quality improvement; the second term... As a penalty for photodamage, it actively avoids situations where cell viability falls below a safe threshold. The situation; the third item This is a light intensity stability constraint term to suppress drastic fluctuations in illumination light intensity; The lighting pattern output for the decision branch. The feature distribution of the current image, For the cell viability predicted by the model, The rate of change in cell viability. Let be the illumination intensity at time t.

[0016] Preferably, in step S4, the imaging acquisition module uses a scientific-grade CMOS camera, and the camera is strictly time-synchronized with the first high-speed scanning galvanometer, the second high-speed scanning galvanometer, and the high-precision spatial optical delay line.

[0017] The second objective of this invention is to provide a live-cell adaptive illumination imaging system for executing the aforementioned live-cell adaptive illumination imaging method, comprising: an illumination imaging module, an image perception module, a decision control module, and an imaging acquisition module; The illumination imaging module is used to output adaptively modulated interference structured light; The image perception module is used to extract cell features in real time and construct an illumination-imaging-optical damage coupling model, outputting a seven-dimensional state vector; The decision control module is used to perform deep reinforcement learning inference based on a seven-dimensional state vector and output the optimal lighting control command. The imaging acquisition module is used to complete the acquisition and super-resolution reconstruction of live cell fluorescence images according to the optimal illumination control command.

[0018] Compared with the prior art, the present invention can achieve the following beneficial effects: (1) Although existing solutions achieve active preset control of lighting intensity (such as flat-top lighting and area customization), all parameters are set before imaging and remain constant during imaging. Essentially, they are still "open-loop control" and cannot be dynamically adjusted according to the sample state. This invention introduces a lightweight deep learning network into structured light lighting control to construct a fully closed-loop feedback mechanism of "acquisition-sensing-decision-modulation". This enables real-time collaborative modulation of lighting parameters and cell dynamic changes, giving the system the intelligent ability to adaptively adjust the lighting strategy according to the real-time state of the sample.

[0019] (2) Traditional fixed illumination mode is mismatched with the rapid random movement of cells, resulting in a lot of ineffective illumination and high phototoxicity. This invention achieves precise illumination on demand based on real-time cell status, which greatly improves photon utilization efficiency; while maintaining the same imaging signal-to-noise ratio, it significantly reduces the cumulative light dose, reduces fluorescence bleaching and phototoxic damage, and effectively extends the long-term observation time of live cells, making it particularly suitable for continuous observation of photosensitive cells.

[0020] (3) Traditional structured light relies heavily on spatial light modulators (SLMs), which have low refresh rates and poor light energy utilization. This invention adopts an interference structured light generation scheme that combines a high-speed scanning galvanometer with a nanometer-precision spatial light delay line. The structured light mode switching speed is improved by 1-2 orders of magnitude, which can meet the high-fidelity capture requirements of millisecond-level cell dynamic events (such as mitochondrial division-fusion). At the same time, it has high phase adjustment accuracy and fast response speed, breaking through the speed bottleneck of traditional hardware.

[0021] (4) By using active vibration isolation, low temperature drift mirror frame and digital PID closed-loop feedback to form a long-term active stabilization unit, environmental vibration and temperature drift are suppressed in real time, ensuring that the interference fringes are not distorted, shifted or attenuated during continuous imaging, and the phase drift is less than λ / 50, which significantly improves the long-term working stability of the system and meets the stringent requirements of long-term live cell super-resolution imaging for hardware stability.

[0022] (5) This invention integrates multi-dimensional information such as cell movement, light field distribution, system stability, and photochemical damage to construct an illumination-imaging-photodamage coupling model and form a seven-dimensional state vector. It combines deep reinforcement learning to complete intelligent decision-making. The decision network is optimized by lightweighting and embedding, with low inference latency and strict temporal synchronization. While ensuring the quality of super-resolution imaging and high signal-to-noise ratio, it maximizes cell activity and achieves dynamic imaging of live cell substructures with high dynamics, long time range, and low phototoxicity. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the illumination imaging module in a live-cell adaptive illumination imaging system provided according to an embodiment of the present invention.

[0024] Figure 2 This is a flowchart of a live-cell adaptive illumination imaging method provided according to an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of a perception-decision dual-branch network structure provided in an embodiment of the present invention.

[0026] Figure label: 1. Laser; 2. Collimator; 3. Beam expander; 4. Non-polarizing beam splitter; 5. First high-speed scanning galvanometer; 6. First reflecting mirror; 7. Second reflecting mirror; 8. Second high-speed scanning galvanometer; 9. High-precision spatial optical delay line; 10. Bundle combiner; 11. Focusing lens; 12. Microscope objective. Detailed Implementation

[0027] In the following description, embodiments of the invention will be described with reference to the accompanying drawings. In the description below, the same modules are denoted by the same reference numerals. Where the same reference numerals are used, their names and functions are also the same. Therefore, their detailed description will not be repeated.

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof.

[0029] This invention provides a live-cell adaptive illumination imaging method, the flowchart of which is shown below. Figure 2 As shown; specifically, it includes the following steps: S1. See also Figure 2 A live-cell adaptive illumination imaging system was constructed, and the system was initialized and calibrated. The live-cell adaptive illumination imaging system includes an illumination imaging module, an image sensing module, a decision control module, and an image acquisition module; The illumination imaging module includes a laser 1, a collimator 2, a beam expander 3, a high-speed fringe generation unit, a nanometer-precision phase control unit, a focusing lens 11, a microscope objective 12, and a long-term active stabilization unit; all of the above optical components are integrated and mounted on the optical platform of the long-term active stabilization unit to achieve rigid connection and stable support of the overall structure; The high-speed stripe generation unit includes a non-polarizing beam splitter prism 4, a first high-speed scanning galvanometer 5, a first reflecting mirror 6, a second reflecting mirror 7, a second high-speed scanning galvanometer 8, and a beam combiner 10. The beam emitted by laser 1 is collimated by collimator 2 and then expanded by beam expander 3. The incident end of non-polarizing beam splitter 4 is connected to beam expander 3. Non-polarizing beam splitter 4 splits the beam into two paths. One path passes through the first high-speed scanning galvanometer 5 and the first reflecting mirror 6 and is then incident on beam combiner 10. The other path passes through the second high-speed scanning galvanometer 8 and the second reflecting mirror 7 and is then incident on beam combiner 10. The two beams are combined by beam combiner 10 to form interference structured light and then output to the focusing lens 11 in the subsequent stage. The beam passes through focusing lens 11 and microscope objective 12 in sequence and finally illuminates the sample. The first high-speed scanning galvanometer 5 and the second high-speed scanning galvanometer 8 adopt an integrated high-speed two-dimensional scanning galvanometer. By precisely controlling the angle between the two coherent beams and the incident angle, the fringe direction and spatial frequency can be independently programmed. In a specific embodiment, the system performance indicators are: fringe contrast >90% and direction and frequency switching speed 15μs.

[0030] The nanometer-precision phase control unit includes a high-precision spatial light delay line 9, a capacitive displacement sensor, and a digital PID controller. The high-precision spatial light delay line 9 is located in one of the optical paths of the high-speed stripe generation unit. The detection end of the capacitive displacement sensor is connected to the displacement execution end of the high-precision spatial light delay line 9 to acquire the displacement signal of the spatial light delay line in real time. The signal output end of the capacitive displacement sensor is connected to the signal input end of the digital PID controller. The control output end of the digital PID controller is connected to the piezoelectric drive end of the high-precision spatial light delay line 9. In a specific embodiment, the high-precision spatial light delay line 9 adopts a closed-loop piezoelectric ceramic drive (e.g., ChipTomorrow, P66.X15S), and through capacitive displacement sensing and digital PID control, achieves a repeatability accuracy better than 6nm (corresponding to λ / 80@488nm) and a response time <1ms.

[0031] The long-term active stabilization unit includes an active vibration isolation platform, a low-temperature drift precision optical frame, and a feedback control module. The laser 1, collimator 2, beam expander 3, high-speed fringe generation unit, nanometer-precision phase adjustment unit, and focusing lens 11 are all fixedly mounted on the active vibration isolation platform. The low-temperature drift precision optical frame is used to clamp and fix each optical component, so that the relative position of each component remains stable. The feedback control module collects vibration and temperature signals in real time and outputs control commands to the active vibration isolation platform and the low-temperature drift precision optical frame. Among them, the active vibration isolation platform is used to suppress environmental vibration interference, and the low-temperature drift precision optical frame (temperature drift <1μrad / ℃) is used to reduce the influence of temperature drift. Real-time feedback control makes the phase stability of the interference fringes better than λ / 50 (i.e., the phase drift is less than λ / 50), ensuring that the phase fluctuations introduced by the hardware in continuous imaging do not affect the reconstruction quality of the super-resolution image. In a specific embodiment, the active vibration isolation platform adopts the TA-600 model from Lien Sheng Optics.

[0032] Perception-Decision Two-Branch Network ( Figure 3 The following are configured in different modules: The image perception module is configured with a perception branch, which is a lightweight convolutional neural network based on MobileNetV3-Small, used to quickly extract dynamic features such as cell motion speed and spatial coherence length from real-time cell images of 256×256ROI; The decision control module is configured with a decision branch, which adopts a three-layer fully connected network structure with the number of neurons set to 128, 64 and 3 respectively, used to receive the seven-dimensional state vector and output the normalized optimal combination of illumination parameters. The image perception module is a lightweight real-time cell state perception unit used to extract cell dynamic features from the real-time cell images output by the imaging acquisition module, including the average cell movement speed and spatial coherence length. Combined with the cumulative light dose, concentration of photosensitive reaction intermediate products, cell viability and viability change trend, a seven-dimensional state vector is formed and output to the decision control module to provide real-time state input for adaptive lighting decision-making. The decision control module is an embedded computing platform with a deep reinforcement learning policy network deployed on it. The decision control module is configured to model adaptive lighting optimization as a Markov decision process, defining a state space, action space and reward function. It adopts a "perception-decision" dual-branch network structure. Based on the seven-dimensional state vector input from the image perception module, it outputs the optimal combination of lighting parameters in real time, including structured light illumination direction, spatial frequency, phase, illumination intensity and exposure time, and sends control commands to the lighting imaging module for execution, realizing closed-loop control of "perception-decision-execution". Specifically, in the decision control module, the state space adopts the output seven-dimensional state vector, which can comprehensively characterize cell dynamics, the degree of light damage accumulation, cell viability, and imaging quality, providing complete observation input for the policy network. The action space is defined as an adjustable combination of illumination parameters, including structured light illumination direction, spatial frequency, phase adjustment, illumination intensity, and camera exposure time. The action commands output by the decision control module can directly drive the illumination imaging module to perform hardware adjustments. The reward function adopts a multi-objective weighted mechanism, which simultaneously maximizes imaging information gain, minimizes light damage, stabilizes illumination intensity, and stabilizes system operation. This allows intelligent decision-making to improve imaging quality while maximizing the protection of live cell viability and avoiding cell damage and signal attenuation caused by photobleaching and phototoxicity. The decision control module, after structured pruning, INT8 low-precision quantization, and TensorRT runtime optimization, achieves an end-to-end inference latency of less than 8ms, meeting the requirements for real-time dynamic control of live cells. In a specific embodiment, to meet the real-time requirements of live cell dynamic imaging, the trained policy network execution model is optimized and embeddedly deployed: L1-norm-based structured channel pruning is adopted to reduce the network's floating-point computation; post-training INT8 low-precision quantization is adopted to significantly reduce memory usage and data bandwidth requirements while ensuring that the accuracy loss is less than 1%; TensorRT SDK is used to perform operator fusion, automatic kernel optimization, and hardware scheduling adaptation on the computation graph, ultimately achieving an end-to-end inference latency of less than 8ms on the embedded platform, meeting the requirements for high-speed, closed-loop adaptive illumination control.

[0033] The imaging acquisition module is used to acquire live cell fluorescence images after adaptive structured light illumination, and transmits the real-time image sequence to the image sensing module for state feature extraction. At the same time, the original image is cached for subsequent super-resolution image reconstruction. The imaging acquisition module maintains strict timing synchronization with the first high-speed scanning galvanometer 5, the second high-speed scanning galvanometer 8, and the high-precision spatial light delay line 9 to ensure that the phase and direction switching of the structured light is completely synchronized with the image exposure, thereby improving the imaging signal-to-noise ratio and reconstruction quality. The imaging acquisition module is a high-performance scientific-grade CMOS camera that is triggered synchronously with the illumination imaging module.

[0034] After the system is built, calibration is performed: turn on laser 1, and form stable parallel light through collimator 2 and beam expander 3; adjust the non-polarizing beam splitter prism 4 to split the beam into two paths; adjust the first high-speed scanning mirror 5 and the second high-speed scanning mirror 8 to make the two beams overlap in beam combiner 10 to form high-contrast interference fringes; use the nanometer-precision phase control unit to accurately calibrate the optical path difference, obtain the correspondence between phase and displacement, and complete the system initialization.

[0035] S2. Establish an illumination-imaging-optical damage coupling model and extract the seven-dimensional state vector of the cell: This step aims to establish a complete lighting-imaging-photodamage coupling relationship through image-sensory feature extraction and photochemical dynamics model calculation. It unifies and fuses real-time cell image information with cell damage state information to construct a seven-dimensional state vector that comprehensively represents the current cell state, providing standardized state input for the subsequent intelligent lighting decision-making of the decision control module. The specific implementation process and operation are as follows: The live cell fluorescence images acquired in real time by the imaging acquisition module are input into the perception branch of the image perception module (using a lightweight convolutional neural network based on MobileNetV3-Small) to perform real-time feature extraction and analysis on the input cell images. At the same time, combined with a pre-built photochemical dynamics model, the damage state of cells under the current lighting conditions is recursively calculated based on real-time lighting parameters and historical lighting records. Finally, the dynamic features obtained from image perception and the damage state calculated by the model are fused to construct a unified seven-dimensional state vector, which is then output to the decision control module as the basis for intelligent lighting optimization decisions.

[0036] The illumination-imaging-photodamage coupling model is a multi-physics coupling model that jointly models the illumination light field parameters, imaging signal characteristics, and cell photodamage state. Based on photochemical dynamics, this model uses cell dynamic characteristics, fluorescence signal changes, and cumulative photodamage as output variables to establish a quantitative mapping relationship between the illumination process, imaging signal, and photodamage degree, thereby achieving a unified characterization of the real-time state and damage level of cells.

[0037] The specific steps are as follows: 1. The perceptual branch performs inference on the 256×256 ROI image and outputs dynamic feature parameters of the samples, including the average cell velocity. Coherence length in space It is used to characterize the current motion state and spatial distribution characteristics of cells.

[0038] 2. Based on the principles of photochemical kinetics, cell damage-related state parameters are updated in real time recursively using a three-tiered kinetic model: Fluorescent probe bleaching kinetic model: ; Kinetic model of reactive oxygen species generation and scavenging: ; Cell viability decay kinetic model: ; in, C f ( t () represents the concentration of the fluorophore. t Indicates time, C R ( tLet be the concentration of reactive oxygen species at time t. V For relative cell viability, I ( t ) represents the light intensity, and σ represents the photon absorption cross section. For quantum yield, k g The rate of reactive oxygen species generation. k c This is the cellular rate constant for scavenging reactive oxygen species (representing the cell's scavenging capacity). k d Damage coefficient; The above model parameters ( s , , k g , k c , k d Independent experimental calibration: Fluorescence decay curves were fitted under different constant light intensities. , s The dynamic curves of the reactive oxygen species fluorescent probe were fitted by measuring the dynamic curves after pulsed light irradiation. k g , k c ; Labeling by long-term live-cell imaging and detection of cell viability markers k d ; Based on the above model recursively calculated, the following damage state parameters can be obtained: (1) Cumulative light dose : Due to light intensity I ( t Integrating over time yields the result; (2) Concentration of intermediate products in photosensitized reaction, including fluorophore concentration C f ( t and reactive oxygen species concentration C R ( t ); (3) Cell viability predicted by the model ; (4) Trends in cell viability : Calculated from the rate of change over consecutive time intervals.

[0039] 3. By integrating dynamic feature parameters and damage state parameters, a seven-dimensional state vector is constructed and output to the decision control module as the basis for intelligent lighting optimization decisions; and Cumulative light dose recursively updated with the damage model Concentration of intermediate products in photosensitized reaction C f ( t )and C R ( t Cell viability predicted by the model and its changing trends Together they form a seven-dimensional state vector: .

[0040] S3. Based on the seven-dimensional state vector, deep reinforcement learning inference is performed to output the optimal lighting control command: The decision control module receives the seven-dimensional state vector output by S2, uses a pre-constructed Markov decision process as the basis for decision-making, completes real-time forward inference calculation through decision branches, outputs the lighting parameter combination that best matches the current cell motion state, light damage accumulation level, and imaging signal-to-noise ratio, and converts it into hardware-executable control commands, providing accurate input for the high-speed regulation of the lighting imaging module. The specific implementation process is as follows: The decision control module receives the seven-dimensional state vector generated by the fusion in S2 and inputs it into the deep reinforcement learning policy network that has completed training, pruning, quantization and hardware adaptation optimization. The deep reinforcement learning policy network uses the state space, action space and multi-objective weighted reward function as the calculation criteria. The reward function simultaneously motivates the maximization of imaging information gain, the minimization of cell photodamage penalty, the smoothing of illumination intensity changes and the optimization of system operation stability, so as to ensure that the decision output takes into account the dual objectives of high imaging quality and low cell damage. The decision branch uses a three-layer fully connected network structure (with 128, 64, and 3 neurons respectively) to perform forward reasoning on the state vector, and the output includes the structured lighting pattern. Illumination light intensity amplitude I t and camera exposure time t t The optimal combination of lighting parameters, such as light direction, spatial frequency, and phase adjustment; During inference, the decision control module maintains the optimized hardware operating state. Through structured pruning, INT8 low-precision quantization, and TensorRT kernel scheduling, it ensures that the end-to-end inference latency is consistently below 8ms, meeting the real-time control requirements of millisecond-level dynamic changes in living cells. After inference is completed, the decision control module encapsulates the optimal illumination parameter combination into standard timing synchronization control commands and sends them to the high-speed stripe generation unit and nanometer-precision phase adjustment unit in the illumination imaging module. This allows the control commands to directly drive the high-speed scanning galvanometer and high-precision spatial light delay line to perform adjustment actions.

[0041] The decision control module sets up a multi-objective weighted reward function, incorporating imaging information gain, optical damage suppression, illumination stability, and system stability into a unified optimization objective to achieve synergistic optimization of imaging quality and cell protection; the instantaneous reward value expression at time t is as follows: ; in, α , β , c , d These are the weighting coefficients; The instantaneous reward value at time t; the first term The first term is the information gain term, which incentivizes image quality improvement; the second term... As a penalty for photodamage, it actively avoids situations where cell viability falls below a safe threshold. The situation; the third item This is a light intensity stability constraint term to suppress drastic fluctuations in illumination light intensity; The lighting pattern output for the decision branch. The feature distribution of the current image, For the cell viability predicted by the model, The rate of change in cell viability. Let be the illumination intensity at time t.

[0042] In this invention, the information gain model and the photodamage model are dynamically coupled, under cell activity constraints ( V ( T )≥ V min By maximizing the cumulative information gain, adaptive illumination regulation is formalized into an optimal control problem, achieving a joint optimality between imaging utility and cell safety. .

[0043] S4. The illumination and imaging module performs high-speed closed-loop control, simultaneously completing structured light illumination and image acquisition: based on the optimal illumination control command (including illumination mode) issued by the decision control module. Illumination light intensity amplitude I t and camera exposure time t t The parameters drive the high-speed stripe generation unit and the nanometer-precision phase modulation unit in the illumination imaging module to work together to achieve high-speed switching of structured light direction and spatial frequency and precise phase adjustment. Under the continuous guarantee of the long-term active stabilization unit, it outputs highly stable, high-contrast (>90%) interference structured light projected onto the sample surface; at the same time, the imaging acquisition module completes strictly time-synchronized image acquisition and split-output. The specific implementation process is as follows: The high-speed stripe generation unit receives the illumination mode in the control command. The parameters drive the first high-speed scanning galvanometer 5 and the second high-speed scanning galvanometer 8 to perform rapid angle deflection. By precisely controlling the incident angle and incident direction of the two coherent beams, the switching of the structured light direction and spatial frequency is completed within 15μs, ensuring that the structured light stripe contrast is greater than 90% and meeting the modulation light field quality required for super-resolution imaging. The nanometer-precision phase control unit adjusts the phase according to the target phase in the control command (belonging to the illumination mode). (Part of it), driving the high-precision spatial light delay line 9 to perform nanometer-level optical path difference adjustment; the capacitive displacement sensor collects the position signal of the delay line displacement execution end in real time and feeds it back to the digital PID controller. The digital PID controller outputs the closed-loop adjustment signal to the piezoelectric drive end in real time according to the feedback deviation, so that the repeatability of the high-precision spatial light delay line is better than 6nm, the corresponding phase adjustment accuracy is better than λ / 80@488nm, and the response time is less than 1ms. After high-speed modulation and precise phase control, the two coherent beams are superimposed and interfered within the beam combiner 10 to form structured light that is optimally matched to the current cell state. The structured light is then sequentially incident on the focusing lens 11 and the microscope objective 12 to control the illumination intensity in the command. I t To achieve the desired output intensity, the light is ultimately projected onto the surface of the live cell sample in a low-dose, high-signal-to-noise-ratio illumination mode, thus completing the adaptive illumination output.

[0044] Throughout the entire control and illumination process, the long-term active stabilization unit operates continuously: the active vibration isolation platform suppresses environmental vibration interference, the low-temperature drift precision optical frame suppresses optical path offset introduced by temperature drift, and the feedback control module outputs compensation commands in real time, ensuring that the phase stability of the interference fringes is better than λ / 50, guaranteeing that the structured light is not distorted, shifted, or attenuated during continuous long-term imaging; the imaging acquisition module uses a high-performance scientific-grade CMOS camera, based on the exposure time specified in the control commands. t t Set the acquisition parameters and maintain strict timing synchronization with the first high-speed scanning galvanometer 5, the second high-speed scanning galvanometer 8, and the high-precision spatial light delay line; once the structured light mode and phase adjustment are completed and stabilized, the camera immediately triggers exposure to acquire live cell fluorescence images under the current optimal illumination conditions.

[0045] The acquired real-time images are transmitted in two paths: the first path is transmitted in real-time to the image sensing module for dynamic feature extraction, recursive updating of optical damage states, and generation of a seven-dimensional state vector in step S2, before entering the next round of closed-loop control; the second path is stored in the image buffer unit of the imaging acquisition module for subsequent super-resolution image reconstruction. During the acquisition process, the imaging acquisition module maintains parameters such as exposure time, gain, readout rate, and illumination mode. Illumination light intensity amplitude It A perfect match maximizes the image signal-to-noise ratio and reduces the interference of noise on feature extraction and image reconstruction.

[0046] S5. Perform super-resolution image reconstruction and iterative optimization on the acquired images to achieve long-term, low-phototoxicity live-cell imaging: Based on multiple frames of adaptive structured light illumination images cached by the imaging acquisition module, and combined with structured light modulation parameters and phase calibration information, perform frequency domain demodulation and super-resolution image reconstruction. The reconstruction results are then fed back to the image perception module to update cell dynamic features and photodamage state parameters, completing a closed-loop iteration of "perception-decision-illumination-acquisition-reconstruction." This provides high-quality image input for the next round of adaptive illumination control, achieving long-term, high-stability, and low-phototoxicity continuous live-cell imaging. The specific implementation process is as follows: The imaging acquisition module inputs multiple frames of raw fluorescence images acquired under different structured light illumination modes into the image reconstruction unit. Based on the illumination parameters recorded in S4, the module performs frequency domain decoupling on the multiple frames of raw images, separating high-frequency information, low-frequency structural information, and environmental noise components of the sample, and eliminating image distortion and phase errors caused by optical path vibration and temperature drift. Based on real-time phase feedback data from high-precision spatial light delay lines, the module performs sub-pixel-level alignment on the multiple frames of images, correcting the positional errors caused by structured light phase shift, and ensuring the spatial positioning accuracy of super-resolution imaging. The corrected multi-frame images are fused and reconstructed to output a single-frame super-resolution imaging result, which can clearly present the dynamic morphology of subcellular structures such as mitochondria. After reconstruction, the super-resolution image is sent back to the image perception module as input data for a new round of cell dynamic feature extraction, updating the seven-dimensional state vector parameters such as average cell movement velocity, spatial coherence length, and cumulative photodamage; in a specific embodiment, the spatial resolution is better than 120nm. Based on the updated state vector, the decision control module quickly executes a new round of forward inference, outputs a combination of illumination parameters that better suits the current cell state, and drives the illumination imaging module to complete the next round of adaptive regulation. Through the iterative cycle from S1 to S5, the system continuously improves the imaging signal-to-noise ratio and temporal resolution while ensuring that cell viability is not significantly damaged, enabling long-term, high-fidelity, and low-phototoxicity continuous observation of the dynamic process of living cells, and finally outputting a high-quality dynamic imaging sequence of living cells.

[0047] Example 1 This embodiment provides a live-cell adaptive illumination imaging method, which specifically includes the following steps: S1. Construct a live-cell adaptive illumination imaging system and perform system initialization and calibration; S2. Establish an illumination-imaging-optical damage coupling model and extract a seven-dimensional cell state vector: using a 256×256 pixel real-time ROI image as input, the perception branch MobileNetV3-Small outputs cell dynamic feature parameters, including average cell velocity. Coherence length in space ;Will and Cumulative light dose recursively updated with the damage model Concentration of intermediate products in photosensitized reactions (fluorophore concentration) C f ( t and reactive oxygen species concentration C R ( t Cell viability predicted by the model and its changing trends Together they form a seven-dimensional state vector: ; S3. Based on the seven-dimensional state vector, perform deep reinforcement learning inference and output the optimal lighting control command: The decision branch is a 3-layer fully connected network with 128, 64, and 3 neurons respectively. It receives the complete state vector and outputs normalized continuous actions, i.e., three-dimensional lighting parameters. A t ={ i t , I t , t t}; The expression for the instantaneous reward value at time t is as follows: ; in, α , β , c , d These are the weighting coefficients; The instantaneous reward value at time t; the first term The first term is the information gain term, which incentivizes image quality improvement; the second term... As a penalty for photodamage, it actively avoids situations where cell viability falls below a safe threshold. The situation; the third item This is a light intensity stability constraint term to suppress drastic fluctuations in illumination light intensity; The lighting pattern output for the decision branch. The feature distribution of the current image, For the cell viability predicted by the model, The rate of change in cell viability. Let be the illumination intensity at time t; To meet real-time requirements, the following deployment optimization schemes are adopted: (1) Structured model compression: Channel pruning based on L1-norm, aiming to reduce floating-point operations by 30%; (2) Low-precision inference: Dynamic range INT8 quantization after training is adopted, reducing memory usage and bandwidth requirements to 1 / 4 with a precision loss of <1%; (3) Runtime deep optimization: Operator fusion, automatic kernel tuning and specific kernel scheduling for the Jetson AGX Orin platform are implemented using NVIDIA TensorRT SDK.

[0048] Through the above optimizations, the end-to-end inference latency on the Jetson AGX Orin platform is consistently below 8ms, laying the core computing power foundation for high frame rate and low latency adaptive closed-loop optical control. S4. The illumination and imaging module performs high-speed closed-loop control, simultaneously completing structured light illumination and image acquisition: S5. Super-resolution image reconstruction and iterative optimization output enable long-term low-phototoxicity live cell imaging.

[0049] Example 2 This embodiment uses K562 leukemia cells as a model to verify the effectiveness of the live-cell adaptive illumination imaging method of the present invention.

[0050] The experiment employed a randomized controlled design: the experimental group used the live-cell adaptive illumination imaging system constructed in this invention; the control group used the traditional SIM fixed parameter mode (light intensity 100mW / cm²). 2 (Exposure time 10ms). Each group collected 30 independent fields of view, performed 3 biological replicates, and observed continuously for 60 minutes.

[0051] The evaluation system covers three dimensions: (1) Imaging performance: average signal-to-noise ratio of time-series images, and the identifiability of key events; (2) Probe durability: fluorescence half-life and photobleaching rate; (3) Biocompatibility: immediate cell survival rate, 24-hour proliferation activity and morphological integrity.

[0052] All quantitative data were statistically analyzed using a two-tailed t-test (α=0.05).

[0053] Experimental results show that the method of the present invention maintains the super-resolution imaging capability of ≤120nm, while reducing phototoxicity by more than 20% compared with the traditional SIM, maintaining cell viability of >80%, and significantly extending the duration of long-term observation of live cells; at the same time, it achieves an end-to-end inference delay of less than 8ms to ensure the real-time performance of adaptive illumination.

[0054] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0055] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A live-cell adaptive illumination imaging method, characterized in that: Specifically, the steps include the following: S1. Construct a live-cell adaptive illumination imaging system and perform system initialization and calibration; the live-cell adaptive illumination imaging system includes an illumination imaging module, an image perception module, a decision control module, and an imaging acquisition module; S2. By extracting features from image perception and calculating photochemical dynamics, an illumination-imaging-photodamage coupling model is established to extract a seven-dimensional state vector of cells. Specifically, this includes: inputting the real-time acquired live cell fluorescence images from the imaging acquisition module into the perception branch of the image perception module, performing real-time feature extraction and analysis on the input cell images; simultaneously, combining the pre-constructed photochemical dynamics model, recursively calculating the cell damage state under the current illumination conditions based on real-time illumination parameters and historical illumination records, and finally fusing the dynamic features obtained from image perception with the damage state calculated by the model to construct a unified seven-dimensional state vector, which is then output to the decision control module. The seven-dimensional state vector is represented as: ; in, Indicates the average velocity of the cell. Indicates the spatial coherence length. Indicates the cumulative light dose. C f ( t () indicates the concentration of the fluorophore. C R ( t Let be the concentration of reactive oxygen species at time t. This represents the cell viability predicted by the model. Indicates the trend of changes in cell viability; S3. Based on the decision control module, perform deep reinforcement learning inference on the seven-dimensional state vector to output the optimal lighting control command; S4. According to the optimal illumination control command, drive the high-speed stripe generation unit and the nano-precision phase modulation unit in the illumination imaging module to work together to complete structured light illumination and image acquisition; S5. Perform super-resolution image reconstruction and iterative optimization on the acquired images to complete long-term low-phototoxicity live cell imaging.

2. The live-cell adaptive illumination imaging method according to claim 1, characterized in that: The illumination imaging module includes a laser, a collimator, a beam expander, a high-speed stripe generation unit, a nanometer-precision phase modulation unit, and a long-term active stabilization unit. The high-speed stripe generation unit includes a non-polarizing beam splitter prism, a first high-speed scanning mirror, a first reflecting mirror, a second reflecting mirror, a second high-speed scanning mirror, and a beam combiner; the nanometer-precision phase control unit includes a high-precision spatial light delay line, a capacitive displacement sensor, and a digital PID controller, with the high-precision spatial light delay line disposed in one of the optical paths of the high-speed stripe generation unit; After the laser beam is collimated by the collimator, it is expanded by the incident beam expander. The beam is split into two paths by the non-polarizing beam splitter. One path passes through the first high-speed scanning galvanometer and the first reflecting mirror and is then incident on the beam combiner. The other path passes through the second high-speed scanning galvanometer and the second reflecting mirror and is then incident on the beam combiner. The two beams are combined by the beam combiner to form interference structured light and then output to the subsequent optical path.

3. The live-cell adaptive illumination imaging method according to claim 2, characterized in that: The first high-speed scanning galvanometer and the second high-speed scanning galvanometer are synchronously and collaboratively controlled. The phase difference between the two beams is adjusted in real time by a nanometer-precision phase control unit, and the spatial frequency and direction adjustment response time of the structured light stripes is no higher than 1ms.

4. The live-cell adaptive illumination imaging method according to claim 2, characterized in that: The long-term active stabilization unit is used to monitor the intensity fluctuations and phase drift of the interference structured light in real time, and forms a closed-loop feedback through a digital PID controller to make the phase drift of the interference fringes less than λ / 50.

5. The live-cell adaptive illumination imaging method according to claim 1, characterized in that: The perception branch is a lightweight convolutional neural network based on MobileNetV3-Small; The photochemical kinetic model includes: Fluorescent probe bleaching kinetic model: ; Kinetic model of reactive oxygen species generation and scavenging: ; Cell viability decay kinetic model: ; in, C f ( t () represents the concentration of the fluorophore. t Indicates time, C R ( t Let be the concentration of reactive oxygen species at time t. V For relative cell viability, I ( t ) represents the light intensity, and σ represents the photon absorption cross section. For quantum yield, k g The rate of reactive oxygen species generation. k c This represents the cellular rate constant for scavenging reactive oxygen species. k d The damage coefficient is denoted as .

6. The live-cell adaptive illumination imaging method according to claim 1, characterized in that: The decision control module is configured with a decision branch, which adopts a three-layer fully connected network structure with the number of neurons set to 128, 64 and 3 respectively. The decision branch is used to receive the seven-dimensional state vector and output the normalized optimal lighting parameter combination.

7. The live-cell adaptive illumination imaging method according to claim 6, characterized in that: The decision control module sets up a multi-objective weighted reward function, incorporating imaging information gain, optical damage suppression, illumination stability, and system stability into a unified optimization objective to achieve synergistic optimization of imaging quality and cell protection; the instantaneous reward value expression at time t is as follows: ； in, α , β , γ , δ These are the weighting coefficients; The instantaneous reward value at time t; the first term The first term is the information gain term, which incentivizes image quality improvement; the second term... As a penalty for photodamage, it actively avoids situations where cell viability falls below a safe threshold. The situation; the third item This is a light intensity stability constraint term to suppress drastic fluctuations in illumination light intensity; The lighting pattern output for the decision branch. The feature distribution of the current image, For the cell viability predicted by the model, The rate of change in cell viability. Let be the illumination intensity at time t.

8. The live-cell adaptive illumination imaging method according to claim 1, characterized in that: In step S4, the imaging acquisition module uses a scientific-grade CMOS camera, and the camera is strictly time-synchronized with the first high-speed scanning galvanometer, the second high-speed scanning galvanometer, and the high-precision spatial optical delay line.

9. A live-cell adaptive illumination imaging system, characterized in that: The method for performing the live-cell adaptive illumination imaging method according to claim 1 includes: an illumination imaging module, an image perception module, a decision control module, and an imaging acquisition module; The illumination imaging module is used to output adaptively modulated interference structured light; The image perception module is used to extract cell features in real time and construct an illumination-imaging-optical damage coupling model, outputting a seven-dimensional state vector; The decision control module is used to perform deep reinforcement learning inference based on a seven-dimensional state vector and output the optimal lighting control command. The imaging acquisition module is used to complete the acquisition and super-resolution reconstruction of live cell fluorescence images according to the optimal illumination control command.

Images