Neural signal accurate capturing and quantifying method for low-light fluorescence microscopic imaging

Abstract

The invention discloses a neural signal accurate capturing and quantifying method for low-light fluorescence microscopic imaging. According to the invention, through the full-chain modular collaborative design of motion correction, deconvolution, baseline correction, image denoising, axial drift compensation and quantization enhancement, a biological feature constrained deep learning fusion module is innovatively constructed. The module adopts a hybrid architecture of cascading an eight-layer convolutional neural network and a four-layer bidirectional gating loop network, adopts an end-to-end self-supervised training strategy in a self-encoder form, performs reconstruction learning by taking an original image as input and a label, and realizes cross-scene adaptation in combination with biological feature constraints constructed from real data. And a dynamic noise estimation and self-adaptive parameter adjustment strategy is introduced, so that the dependence on manual annotation data is remarkably reduced, and accurate quantification of weak fluorescence signals such as neurotransmitters and calcium signals is realized.

Description

Technical Field

[0001] This invention relates to the field of low-light fluorescence microscopy image processing technology, specifically to a method for precise capture and quantification of neural signals for low-light fluorescence microscopy imaging, which can be widely applied in neuroscience, biomedical imaging, and other fields. This method enables efficient processing and precise quantification of weak fluorescence signals in living, fixed, or freely moving experimental animals, while also being compatible with the analysis needs of calcium signals and various neurotransmitter signals. It supports synaptic-level neural signal transmission pattern analysis, neural circuit function analysis, and related disease mechanism research, and is particularly suitable for complex imaging scenarios involving low light, long time durations, and multiple interfering factors. Background Technology

[0002] In neuroscience research, exploring the correlation between neural activity and behavior in the brains of living animals relies on low-light fluorescence microscopy. The fluorescence signals related to physiological processes such as gene-encoded neurotransmitter release and calcium signal changes are typically only 1-3 times the background noise, and are easily affected by various factors in experimental settings. Existing imaging signal processing methods (such as gray-scale matching-based registration algorithms and fixed-parameter deconvolution methods) have significant limitations, restricting research accuracy and efficiency. These limitations mainly include the following key issues:

[0003] Poor handling of motion interference: When experimental animals move freely, the inter-frame translation of images is often corrected by existing methods that use registration algorithms based on gray-level matching. These algorithms are not optimized for the low signal-to-noise ratio characteristics of low-light scenes and are prone to translation estimation errors due to noise, introducing new artifacts.

[0004] It is difficult to balance resolution and noise: Spatial deconvolution methods mostly use fixed parameters and lack targeted calibration and noise control mechanisms. They cannot dynamically adjust the deconvolution strategy according to the optical characteristics of the imaging system, which can easily lead to noise amplification or overfitting, and cannot achieve both resolution improvement and noise suppression.

[0005] Poor baseline stability: Low-light fluorescence signals are easily affected by light source drift and dark current fluctuations. Traditional baseline correction methods often use a single frame or fixed window to calculate the baseline, which cannot effectively smooth temporal fluctuations, leading to subsequent signal quantization deviations.

[0006] Difficulty in preserving weak signals: Traditional denoising methods struggle to distinguish weak biological signals from noise, fail to incorporate the inherent temporal characteristics and spatial distribution patterns of biological signals, and are prone to losing transient weak signals such as neurotransmitter release. Some processing methods result in distorted signal trajectories. Some deep learning-based methods rely excessively on manually labeled data and are only applicable to single types of neural signals, lacking the ability to process multiple types of neurotransmitter signals.

[0007] Insufficient axial drift correction: During long-term imaging, the microscope focal length is prone to axial drift. Existing methods mostly rely on manual intervention or simple interpolation, which cannot achieve automated, high-precision pixel-level correction, resulting in blurred imaging of deep tissues.

[0008] Limited effectiveness of weak signal enhancement: Traditional enhancement methods often use global brightness adjustment, which cannot highlight the temporal differences and spatial reliability of signals, and is not conducive to subsequent synaptic signal analysis and neural transmission mode research.

[0009] In addition, existing methods mostly rely on fixed parameters or manual intervention, lack adaptive learning capabilities, and are difficult to maintain stable performance under different imaging conditions (such as different microscope systems, animal models, or noise levels), which limits their large-scale application and generalization ability.

[0010] Therefore, there is an urgent need for an integrated processing method that takes into account motion correction accuracy, resolution improvement, noise suppression, baseline stabilization, weak signal preservation and signal enhancement, and can adapt to multiple types of neural signals, eliminate the need for manual annotation, and be compatible with complex imaging scenarios, so as to meet the research needs of low-light fluorescence in vivo imaging. Summary of the Invention

[0011] To address the shortcomings of existing technologies, this invention provides a method for the precise capture and quantification of neural signals in low-light fluorescence microscopy. Through modular design and technology integration, it solves the aforementioned technical pain points and achieves efficient processing and precise quantification of low-light fluorescence signals.

[0012] The technical solution adopted by this invention to solve its technical problem is as follows:

[0013] Step 1: Obtain a sequence of live low-light fluorescence micrographs;

[0014] Step 2, Motion Correction: The Fourier phase correlation registration method is used to convert the image to the frequency domain and calculate the cross power spectrum to estimate the translation transformation between adjacent image frames in the live low-light fluorescence microscopy image sequence, which is used to eliminate motion artifacts in the imaging process.

[0015] Step 3, Deconvolution: Based on the calibrated point spread function, the motion-corrected image sequence is processed through iterative deconvolution and regularization mechanisms to obtain a high-resolution deconvolution image sequence;

[0016] Step 4, Baseline Correction: A dual baseline region dynamic selection and pixel-level exponential fitting model is adopted. The baseline adjustment factor is calculated through nonlinear least squares fitting to smooth the temporal fluctuations of the deconvolution image sequence and generate a stable baseline-corrected image sequence.

[0017] Step 5, Image Denoising: Denoising the baseline-corrected image sequence using the constructed deep learning fusion module;

[0018] Step 6, Quantization and Enhancement: Using the baseline of the denoised image sequence as a reference, calculate the relative change value of the signal, remove abnormal artifacts, and use a spatiotemporal coding and channel-specific enhancement collaborative strategy to amplify the weak fluorescence signal, and finally output the quantized data.

[0019] Furthermore, when the total imaging time exceeds 1 minute or the observed object exhibits physiological fluctuations, axial drift compensation is enabled. Its core function is to correct axial defocus blur during long-term imaging. Specifically, the process is as follows: The denoised image sequence and the baseline-corrected image sequence are loaded, and the reference frame with the highest intensity is identified; the pixel-level difference map between the reference frame and the average frame is calculated, and an effective pixel mask is generated using a dynamic threshold; vectorized three-dimensional matrix operations are employed to apply a weighted average of the effective pixels to the baseline data, generating a pixel-level baseline correction curve; intensity normalization and offset compensation are performed using the baseline correction curve, and the axial drift-compensated image sequence is output.

[0020] Furthermore, in step 2, motion correction for low signal-to-noise ratio adaptation: its core function is to eliminate inter-frame motion artifacts caused by animal activity and deeply adapt to the low signal-to-noise ratio characteristics of low-light scenes. The key is to first perform pre-filtering on the image sequence acquired in step 1, then crop the image edge regions, estimate the translation transformation between adjacent image frames using frequency domain cross-power spectrum phase estimation, perform initial inter-frame registration using phase correlation, and then achieve sub-pixel accuracy through a mutual information-based optimized registration method. The optimizer parameters include the maximum number of iterations, step size, and relaxation factor, which are set to fixed values ​​based on experience. The pre-filtering is Gaussian filtering with a standard deviation of 1, used to smooth each frame in the image sequence to reduce noise.

[0021] Furthermore, the core function of deconvolution in step 3 is to improve the spatial resolution of the image while suppressing noise amplification. The calibration point spread function is obtained by fitting the actual PSF data through a double exponential Gaussian model. An iterative deconvolution algorithm is adopted, and the regularization mechanism achieves a balance between noise suppression and resolution improvement by dynamically controlling the number of iterations. The PSF model uses double Gaussian components and energy normalization processing. The number of iterations can be adjusted automatically according to signal correlation denoising and online noise estimation to balance the deconvolution effect and the risk of overfitting.

[0022] Furthermore, the baseline correction in step 4 employs a dual-baseline region exponential fitting baseline correction, the core function of which is to eliminate temporal fluctuations and obtain a stable signal baseline. By defining initial and late dual-baseline regions (e.g., initial region 1:99 frames and late region 300: frame end), a pixel-level exponential fitting model is applied for nonlinear least squares fitting (e.g., using the partNLSQ algorithm) to generate a baseline adjustment factor. This dual mechanism ensures baseline stability, adapts to the dynamic changes of light source drift and signal fluctuations, and provides a reliable reference for subsequent quantization. Temporal filtering uses an adaptive window, with the window size dynamically adjusted according to the signal fluctuation frequency. Dark current subtraction is based on noise statistics in flat areas of the image, and the baseline is defined as the mean of a preset number of frames before stimulation. A dual-baseline region dynamic selection mechanism is adopted, using a pixel-level exponential fitting model to smooth temporal fluctuations. When fitting fails, default coefficients are automatically activated to ensure uninterrupted process.

[0023] Furthermore, the deep learning fusion module described in step 5 includes a neural network unit and a biometric constraint unit. The deep learning fusion module undergoes self-supervised training through a regression task aligning low-light fluorescence microscopy image sequences with denoised image sequences. During the inference phase, the spatiotemporal feature vector of each pixel output by the neural network unit, the event response intensity score calculated based on sparse representation, and the biological temporal similarity score output by the biometric constraint unit are adaptively weighted and fused according to the noise level of the current image to generate a signal credibility weight map (representing the probability distribution of each pixel signal as a real biological signal). This innovatively introduces a biometric constraint mechanism, filtering effective signals based on the inherent temporal patterns of neural signals (such as the steep rise and slow fall characteristics and release decay patterns of neurotransmitter signals) and spatial distribution characteristics. This effectively distinguishes weak biological signals from noise, preserving the true trajectory and decay characteristics of weak signals such as neurotransmitters, achieving high-fidelity signal extraction without relying on manually labeled data. Meanwhile, the baseline-corrected image sequence is optimized by passing it through a deep one-dimensional convolutional network and combining it with a signal confidence weight map, outputting the optimized signal trajectory, and finally outputting the denoised image sequence; the deep learning fusion module filters effective signals by combining the inherent temporal patterns and spatial distribution characteristics of neural signals, quantifies the signal attenuation features, and obtains the denoised image sequence.

[0024] The neural network unit is specifically implemented as follows: The neural network unit adopts a hybrid architecture of an 8-layer convolutional neural network and a 4-layer bidirectional gated recurrent network cascaded together. The 8-layer convolutional neural network includes an input layer, a feature expansion layer, a spatial feature extraction layer with dilated convolution, a feature compression layer, a feature re-expansion layer, and a regression output layer. Each of its intermediate layers includes batch normalization and LeakyReLU activation functions. The 4-layer bidirectional gated recurrent network implements forward and backward propagation paths through a custom gated recurrent unit. Each layer has residual connections, a gating mechanism, and batch normalization. The input to this neural network unit is a baseline-corrected image sequence. The 8-layer convolutional neural network first extracts and reconstructs features from a single frame image, outputting the preliminary denoising results and their sparse representations for each frame. The 4-layer bidirectional gated recurrent network receives a pixel-level temporal sequence composed of the sparse representations of each frame, performs spatiotemporal feature modeling, and outputs a spatiotemporal feature vector for each pixel.

[0025] The biometric constraint unit is implemented as follows: Based on the neural signal dynamics model, the typical activity patterns of neurons are mathematically modeled; the biometric constraint unit first selects the top 10% of pixels with the highest response intensity (referring to the change amplitude of the fluorescence signal of the pixel in time) from the spatiotemporal feature vectors of all pixels, and averages their signal trajectories to generate a global signal template; then, a single / double exponential decay model is used to fit the global signal template, and the model parameters include the rise time constant and the double decay time constant; finally, the temporal correlation between the original signal trajectory of each pixel and the fitted global signal template is calculated, and the normalized correlation coefficient is output as the biometric temporal similarity score of the pixel (measuring the degree of temporal matching between the pixel signal and the typical activity pattern of the neuron);

[0026] Further, step 6, quantization and enhancement, serves the core function of achieving signal quantization and improving the visualization of weak signals. The key design uses a stable baseline as a reference to calculate the relative change value of the signal, and enables background subtraction to calculate the relative change value of the signal (ΔF / F0). Combined with pseudo-color coding and channel-specific enhancement, it highlights the spatiotemporal dynamic characteristics and spatial specific characteristics of the signal, facilitating subsequent synaptic signal analysis and verification. Compared to traditional global enhancement methods, the identification of signal temporal differences and spatial distribution characteristics is significantly improved. A dynamic range compression technique based on the 99.9th percentile is adopted (determining the upper limit of the signal dynamic range through the 99.9th percentile to avoid extreme value interference, while compressing the signal interval to highlight weak signals). Combined with a linear amplification mechanism and a dual-threshold linkage mechanism, the initial screening threshold and the fine screening threshold are separated to control signal sensitivity.

[0027] Furthermore, the present invention also provides an imaging system, which includes an image acquisition module, a processing module, and a result output module:

[0028] The image acquisition module is used to acquire low-light fluorescence microscopy image sequences; the processing module is a computing device equipped with a data processing environment, pre-installed with deconvolution tools, deep learning toolboxes, curve fitting tools, and dedicated analysis functions, supporting batch data parallel processing; the processing module integrates adaptive optimization algorithms (including dynamic noise level estimation, dynamic filter parameter adjustment, dynamic optimization parameter adjustment, initial parameter adaptive adjustment, background pixel percentage dynamic estimation, and threshold adaptive calculation), which can dynamically adjust processing parameters according to the characteristics of input data; the result output module is used to output the processed image sequences, quantized data, and visualization results, compatible with mainstream analysis software.

[0029] The beneficial effects of this invention are as follows:

[0030] 1. High motion correction accuracy: The optimized registration technique is adapted to the low-light and low signal-to-noise ratio characteristics of imaging freely moving animals. The inter-frame translation estimation bias is significantly reduced, and the motion artifact suppression effect is better than that of traditional grayscale matching algorithms. A two-stage registration strategy is adopted, with phase correlation method to quickly estimate the initial displacement, mutual information optimization to achieve sub-pixel accuracy, and optimizer parameters dynamically and adaptively adjusted according to the image noise level, significantly improving the registration robustness.

[0031] 2. Strong fidelity in weak signals: The deep learning module for biometric constraints adopts an autoencoder architecture for end-to-end self-supervised training. It learns spatiotemporal structural features by reconstructing input data, and combines it with a biometric constraint mechanism built from real data (extracting a global template from the top 10% of pixels with the highest response intensity and fitting an exponential decay model). This effectively distinguishes weak biological signals from noise, preserves weak signals such as neurotransmitters and real decay features, and avoids the signal loss or trajectory distortion problems of traditional methods. This architecture combines sparse representation with spatiotemporal modeling capabilities and achieves optimization through a self-supervised learning framework.

[0032] 3. Comprehensive Quantification Capabilities: By fitting output attenuation parameters through an exponential model, it supports synaptic-level neural signal analysis and transmission pattern research, providing quantitative evidence for the interpretation of physiological mechanisms, and simultaneously meeting the quantification needs of calcium signals and multiple neurotransmitter signals. Based on 99.9th percentile dynamic range compression and a dual-threshold linkage mechanism, it achieves intelligent adaptation of the signal dynamic range;

[0033] 4. High degree of integration: The entire process is automated, and the main parameters of each module are adaptively adjusted according to the dynamic noise estimation results, adapting to various in vivo imaging scenarios in neuroscience, including long-term imaging in both fixed and free-moving states, significantly improving processing efficiency. Dynamic noise estimation is based on multi-region joint statistics, and the filter parameters (Gaussian / median), network weights, and initial fitting values ​​are all adjusted in real time according to the noise level, achieving closed-loop optimization throughout the entire process;

[0034] 5. Wide Range of Applications: Adaptable to imaging of various neurotransmitters and calcium signals, providing technical support for research on related disease mechanisms, especially in neurodegenerative disease models, helping to reveal the underlying mechanisms of neural circuit dysfunction and filling the gap in the compatibility of multiple types of neural signals in existing technologies. This method, through modular design and adaptive learning mechanisms, possesses strong cross-platform generalization capabilities, maintaining stable performance under different experimental conditions and reducing deployment and parameter tuning costs. The method supports cross-species applications; through parameter fine-tuning mechanisms, it can achieve stable generalization in different model organisms (mice, rats, etc.) brain regions, demonstrating significantly better cross-scenario adaptability than traditional deep learning methods that rely on specific imaging conditions for training. Attached Figure Description

[0035] Figure 1: Overall flowchart of the method of the present invention, showing the connection between the steps of motion correction, deconvolution, baseline correction, image denoising, axial drift compensation, quantization and enhancement;

[0036] Figure 2: Imaging system configuration diagram, showing the connection relationship between the image acquisition module (CMOS), processing module (computing device), and result output module;

[0037] Figure 3: Signal denoising module structure diagram, showing the complete data stream of the denoised sequence from baseline adjustment, joint denoising, to sparse representation and feature expansion modules, and finally through adaptive weight fusion (GRU and 1D-CNN) and exponential fitting unit, and integrating noise adaptation and biological constraint mechanisms.

[0038] Figure 4: Comparison of quantization and enhancement effects. The top shows the neurotransmitter ΔF / F0 signal and the pseudo-color green channel signal (selected from a specific number of frames). The bottom shows the signal trajectory comparison before and after denoising (477 frames in total, solid line: before denoising, dashed line: after denoising).

[0039] Figure 5: Analysis of neurotransmitter fluorescence signal attenuation, including reference frame image, maximum ΔF / F projection map, ROI attenuation curve, full image pixel attenuation curve, attenuation length distribution histogram, and analysis summary parameters (such as effective ROI number, overall λ value, average λ value, pixel size, range, etc.).

[0040] Figure 6: PSF calibration results. The top figure shows the original PSF data and the double exponential fitting curve (including the 95% confidence interval), and the bottom figure shows the distribution of the fitting residuals (to verify the fitting accuracy). Detailed Implementation

[0041] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention. The invention will be specifically described below with reference to the accompanying drawings and embodiments.

[0042] Example: Imaging of gene-encoded neurotransmitters in mice under free-movement conditions

[0043] I. Experimental Preparation

[0044] Image acquisition: A CMOS imaging system (such as Miniscope) is used, with an excitation light power density ≥2 mW / mm². The system supports flexible configuration of key parameters to adapt to different experimental scenarios: the frame rate is adjustable (typical range 10-60 Hz), and the image size and pixel dimensions can be set and cropped according to the microscope model and imaging analysis requirements. Fluorescence image sequences of neurotransmitters (such as dopamine and acetylcholine) in the mouse brain are acquired and output in TIFF format. The total number of frames and the stimulus initiation frame are customized by the user according to the experimental design.

[0045] Own dataset: This method is adapted to the neurotransmitter fluorescence image sequences of mice in a free-moving state acquired in the above image acquisition experiment. The naming format is "experiment identifier_frame number.tif", and it is stored in the specified data directory.

[0046] Public Dataset Adaptation: This method is also compatible with various publicly available neuroimaging datasets. Specific acquisition and application methods are as follows:

[0047] NeuroData (https: / / neurodata.io / ): Provides in vivo low-light fluorescence imaging datasets for model organisms such as mice and rats, covering neurotransmitter signal data such as calcium and dopamine. It supports TIFF format download and can be directly used for cross-scenario validation of this method.

[0048] Allen Brain Atlas (https: / / portal.brain-map.org / ): Contains a standardized neural circuit fluorescence imaging dataset, providing detailed experimental parameter annotations (such as excitation power, frame rate, and animal model information) to suit the parameter fine-tuning needs of this method.

[0049] Application of public datasets: The downloaded public datasets must ensure that the frame sequence is continuous and the format is TIFF. Only the parameters such as "number of stimulus frames" and "baseline window range" of each module in this method need to be adjusted according to the imaging parameters of the dataset (such as frame rate and pixel size). The core processing flow does not need to be modified to achieve accurate signal capture and quantization.

[0050] Data storage: Image naming format is "experiment identifier_frame number.tif";

[0051] II. Method Implementation

[0052] 1) Motion interference elimination: This is achieved through motion correction, specifically implemented using the technical solution in step 2. First, the image sequence acquired in step 1 undergoes pre-filtering (Gaussian filter parameter with a standard deviation of 1). Then, the image edge regions are cropped (cropping ratio of 0.1 to remove invalid interference areas). The translation transformation between adjacent image frames is estimated using the cross-power spectrum phase estimation in the frequency domain. The Fourier phase correlation registration method is used to convert the image to the frequency domain and calculate the cross-power spectrum to estimate the translation transformation between adjacent image frames in the live low-light fluorescence microscopy image sequence, thus eliminating motion artifacts during imaging. The optimizer parameters (maximum number of iterations, step size, and relaxation factor) are dynamically and adaptively adjusted based on inter-frame differences. The input image sequence outputs registered data. This step effectively eliminates inter-frame translation artifacts caused by free animal movement, providing a spatially aligned image basis for subsequent processing.

[0053] 2) Deconvolution processing is performed based on the calibrated point spread function (PSF). The PSF is obtained by fitting the actual captured data using a double exponential Gaussian model (see fitting results). Figure 6 The top figure shows the original PSF data and the double exponential fitted curve, including the 95% confidence interval; the bottom figure shows the fitted residual distribution, verifying that the fitting accuracy meets the requirements. The Richardson-Lucy iterative algorithm is used, combined with a regularization mechanism to control noise. The number of iterations is preset to a fixed value (e.g., 50 times) to balance the deconvolution effect with the risk of overfitting. This step can enhance the spatial detail resolution of synaptic signals, making the edge details of neurotransmitter release sites clearer.

[0054] 3) Baseline Correction: Define dual baseline regions (e.g., initial region 1:99 frames and late region 300: frame end). For the time series of each pixel, extract the baseline region signal value, perform nonlinear least squares fitting using an exponential function (e.g., using the partNLSQ algorithm), calculate the fitting coefficients and adjustment basis, and then generate a pixel-level baseline adjustment factor. If fitting fails, the default coefficients [0,0] and adjustment basis [1,1] are automatically enabled. The adjustment factor is applied to the entire image sequence, outputting a stable baseline-corrected image sequence. This step improves efficiency through vectorization and ensures process robustness, providing a reliable reference baseline for subsequent quantization.

[0055] 4) Adaptive image denoising: Run the deep learning fusion module (the module structure is shown in Figure 3. The input is the baseline adjustment sequence, which is jointly denoised, sparsely represented and feature expanded. Combined with noise adaptation mechanism and biological constraint enhancement, the output denoised sequence is obtained by adaptive weight fusion of the bidirectional GRU module and the 1D-CNN module and fitting of the single / double exponential model based on nonlinear optimization).

[0056] This module employs an 8-layer convolutional neural network structure (including dilated convolution, batch normalization, and the LeakyReLU activation function; LeakyReLU is a ReLU activation function with a negative slope, which can alleviate the gradient vanishing problem). It also integrates a bidirectional gated recurrent network (GRU, which controls information flow through a gating mechanism and is adept at capturing long-term temporal dependencies) composed of multiple layers for temporal feature extraction. Based on the input imaging sequence data, it performs end-to-end supervised training and fine-tuning to generate a confidence weight map of the signal (representing the probability distribution of each pixel signal as a real biological signal) and preliminary denoising results. Combining biological feature constraints (based on a global temporal template and spatial distribution prior generated by fitting a single / double exponential function, simulating the "steep rise and slow fall" characteristics of neural response) to filter effective signals, it outputs a spatial attention map and fitting parameters.

[0057] The noise reduction effect is shown in Figure 4, with neurotransmitters at the top. The signal and the pseudo-color green channel signal (frames 10, 160, 180, 200, and 300 are selected to show the signal distribution at different time sequences). Below is a comparison of the signal trajectory of 477 frames (solid line is before denoising, dashed line is after denoising). It can be seen that the fluctuation of the signal trajectory is significantly reduced after denoising, while the transient peak and attenuation characteristics of the neurotransmitter signal are completely preserved.

[0058] 5) Axial drift compensation: Identify the reference frame with the highest intensity in the denoised image sequence, calculate the pixel-level difference map between the reference frame and the average frame of the sequence, and generate an effective pixel mask based on a preset threshold; through vectorized three-dimensional matrix operations, apply an effective pixel-weighted average to the baseline data to generate a pixel-level baseline correction curve, achieving intensity normalization and drift compensation. This step achieves pixel-level intensity correction by constructing an axial correction curve. The correction curve parameters are dynamically optimized based on the imaging time. Vectorized three-dimensional matrix operations are used instead of traditional frame-by-frame iteration, improving processing efficiency and effectively overcoming the problem of blurred deep tissue imaging caused by microscope focal length axial drift during long-term imaging.

[0059] 6) Signal Quantization and Enhancement: Calculating the relative change value of the signal. ,in Indicates the fluorescence intensity of the current image frame compared to the baseline intensity. The difference, To generate pseudo-color images and videos, the average value of the resting frames before stimulation is used. The formula is defined as follows:

[0060] When background subtraction is not enabled:

[0061] When background subtraction is enabled: ;

[0062] The parameters are defined as follows: Indicates the fluorescence intensity of the current image frame; The mean fluorescence intensity of the resting frame before stimulation (i.e., the baseline calculated in step 3) represents the fluorescence intensity before stimulation. ); The average fluorescence intensity represents the background region of the image, calculated by selecting flat areas with no signal activity at the four corners of the image.

[0063] After quantization, pseudo-color images and videos were generated (see the pseudo-color image at the top of Figure 4). Dynamic range compression based on the 99.9th percentile was employed, combined with a linear amplification mechanism and a dual-threshold linkage mechanism (separately setting the initial screening threshold and the fine screening threshold) to highlight the spatiotemporal dynamic characteristics of weak signals. The signal attenuation analysis results are shown in Figure 5, including the reference frame image, the maximum ΔF / F projection map, the ROI attenuation curve, the full image pixel attenuation curve, the attenuation length distribution histogram, and analysis summary parameters (effective ROI number 20, overall λ value 0.457 μm, average λ value 0.559±0.232 μm, pixel size 0.448 μm, etc.). The goodness of fit R² ≥ 0.308, conforming to the single / double exponential attenuation law of neurotransmitter signals.

[0064] Furthermore, the imaging system described in this invention:

[0065] Image acquisition module: It adopts a CMOS imaging system with a frame rate adapted to the imaging scene (such as a frame rate of 10-60Hz for imaging of freely moving animals), outputs TIFF format low-light fluorescence image sequences, adapts to the high sensitivity requirements of low-light signal acquisition, and is compatible with miniscope and other microscopic systems.

[0066] Processing module: A computing device equipped with a data processing environment, pre-installed with deconvolution tools, deep learning toolboxes, curve fitting tools and dedicated analysis functions, supporting batch data parallel processing; the processing module integrates an adaptive optimization algorithm, which can dynamically adjust processing parameters according to the characteristics of input data, improving the robustness and generalization performance of the system under different imaging conditions;

[0067] Results output module: Supports outputting processed image sequences (TIFF / BMP), ΔF / F0 quantized data and fitting parameters (Excel / PNG), and visualized videos (MP4 / AVI, H.264 encoded), compatible with mainstream analysis software such as ImageJ and FIJI, facilitating subsequent research.

[0068] III. Result Verification

[0069] The results of this method in processing in vivo neurotransmitter imaging data are as follows:

[0070] Motion correction effect: Improved spatial alignment consistency between image frames; the registered data can meet the basic requirements for subsequent spatial resolution improvement and signal quantization.

[0071] Deconvolution effect: The spatial detail resolution of neurotransmitter release sites is improved, the edge contour of the signal region is clear, and the residuals of the PSF fitting used are all less than 0.1 (see Figure 6 Residual Plot), with no noise amplification or overfitting.

[0072] Baseline stability: The baseline fluctuation range is controlled within ±5% of the mean of the resting frames before stimulation, stabilizing the signal baseline and providing a reliable basis for signal quantization;

[0073] Fitting and weak signal performance: The goodness of fit of single / double exponential signals R² ≥ 0.308, which conforms to the attenuation law of biological signals; the number of effective ROIs reached 20, the overall attenuation length λ = 0.457 micrometers, and the average attenuation length λ = 0.559 ± 0.232 micrometers (see Figure 5 for analysis summary). The transient peak and attenuation characteristics of weak neurotransmitter signals were completely preserved.

[0074] Axial correction effect: The defocus blur caused by axial drift is eliminated, the clarity of deep tissue imaging is improved, the intensity consistency of each frame in a long imaging sequence is enhanced, and there is no obvious local blur or intensity change phenomenon.

[0075] Signal enhancement effect: The spatiotemporal dynamic characteristics of weak signals are improved, and the pseudo-color image can clearly distinguish signal triggering events of different time sequences (see Figure 4, pseudo-color image). The fitting parameters and quantitative data provide quantitative basis for subsequent physiological mechanism analysis.

[0076] This method can also be extended to imaging of neurotransmitters such as dopamine and oxytocin. Only parameters such as the number of stimulation frames need to be adjusted, without changing the core process. It provides a breakthrough tool for neural circuit research and is particularly suitable for long-term monitoring of the miniscope system in freely moving animals.

[0077] Any equivalent substitutions or structural adjustments based on the principles of this invention should be included within the scope of protection of this patent. Those skilled in the art should understand that the specific embodiments described in the specification and drawings are merely illustrative of the technical solutions of this invention and should not be considered as limitations on the scope of patent protection. Any reasonable modifications, equivalent substitutions, or improvements made to the method steps, parameter adjustments, or system modules within the spirit and principles of this patent are within the scope of protection of this patent.

Claims

1. A method for precise capture and quantification of neural signals for low-light fluorescence microscopy, characterized in that, Includes the following steps: Step 1: Obtain a sequence of live low-light fluorescence micrographs; Step 2, Motion Correction: The Fourier phase correlation registration method is used to convert the image to the frequency domain and calculate the cross power spectrum to estimate the translation transformation between adjacent image frames in the live low-light fluorescence microscopy image sequence, which is used to eliminate motion artifacts in the imaging process. Step 3, Deconvolution: Based on the calibrated point spread function, the motion-corrected image sequence is processed through iterative deconvolution and regularization mechanisms to obtain a high-resolution deconvolution image sequence; Step 4, Baseline Correction: A dual baseline region dynamic selection and pixel-level exponential fitting model is adopted. The baseline adjustment factor is calculated through nonlinear least squares fitting to smooth the temporal fluctuations of the deconvolution image sequence and generate a stable baseline-corrected image sequence. Step 5, Image Denoising: Denoising the baseline-corrected image sequence using the constructed deep learning fusion module; Step 6, Quantization and Enhancement: Using the baseline of the denoised image sequence as a reference, calculate the relative change value of the signal, remove abnormal artifacts, and use a spatiotemporal coding and channel-specific enhancement collaborative strategy to amplify the weak fluorescence signal, and finally output the quantized data.

2. The method for precise capture and quantization of neural signals for low-light fluorescence microscopy as described in claim 1, characterized in that, The deep learning fusion module described in step (5) includes a neural network unit and a biometric constraint unit; The deep learning fusion module performs self-supervised regression training on the baseline-corrected image sequence; during the inference phase, the neural network unit outputs the spatiotemporal feature vector of each pixel and calculates the event response intensity score based on sparse representation; The biometric constraining unit outputs a biometric temporal similarity score, which is adaptively weighted and fused based on the noise level of the current image to generate a signal credibility weight map. Simultaneously, the baseline-corrected image sequence is passed through a deep one-dimensional convolutional network and optimized in conjunction with the signal credibility weight map to output an optimized signal trajectory, ultimately outputting a denoised image sequence. The deep learning fusion module selects effective signals by combining the inherent temporal patterns and spatial distribution characteristics of neural signals, quantifies signal attenuation features, and obtains the denoised image sequence.

3. The method for precise capture and quantization of neural signals for low-light fluorescence microscopy as described in claim 2, characterized in that, The neural network unit is specifically implemented as follows: The neural network unit employs a hybrid architecture consisting of an 8-layer convolutional neural network and a 4-layer bidirectional gated recurrent network. The 8-layer convolutional neural network includes an input layer, a feature dilation layer, a spatial feature extraction layer with dilated convolution, a feature compression layer, a feature re-dilation layer, and a regression output layer. Each of its intermediate layers includes batch normalization and the LeakyReLU activation function. The 4-layer bidirectional gated recurrent network implements forward and backward propagation paths through a custom-defined gated recurrent unit. Each layer incorporates residual connections, a gating mechanism, and batch normalization. The input to this neural network unit is a baseline-corrected image sequence. The 8-layer convolutional neural network first extracts and reconstructs features from a single frame image, outputting the preliminary denoising results and their sparse representations for each frame. The 4-layer bidirectional gated recurrent network receives a pixel-level temporal sequence composed of the sparse representations of each frame, performs spatiotemporal feature modeling, and outputs a spatiotemporal feature vector for each pixel.

4. The method for precise capture and quantification of neural signals for low-light fluorescence microscopy as described in claim 3, characterized in that, The biometric constraint unit is specifically implemented as follows: Based on the neural signal dynamics model, the typical activity patterns of neurons are mathematically modeled. The biofeature constraint unit first selects the top 10% of pixels with the highest response intensity from the spatiotemporal feature vectors of all pixels, and averages their signal trajectories to generate a global signal template. Then, a single / double exponential decay model is used to fit the global signal template, and the model parameters include the rise time constant and the double decay time constant. Finally, the temporal correlation between the original signal trajectory of each pixel and the fitted global signal template is calculated, and the normalized correlation coefficient is used as the biotemporal similarity score of the pixel.

5. The method for precise capture and quantification of neural signals for low-light fluorescence microscopy as described in claim 4, characterized in that, The motion correction in step 2 involves first performing pre-filtering on the low-light fluorescence microscopy image sequence acquired in step 1, then cropping the image edge regions, estimating the translation transformation between adjacent image frames using frequency domain cross-power spectrum phase estimation, performing initial inter-frame registration using the phase correlation method, and then achieving sub-pixel accuracy using an optimized registration method based on mutual information. The optimizer parameters include the maximum number of iterations, step size, and relaxation factor, which are set to fixed values ​​based on experience. The pre-filtering is a Gaussian filter with a standard deviation of 1, used to smooth each frame in the image sequence to reduce noise.

6. The method for precise capture and quantization of neural signals for low-light fluorescence microscopy as described in claim 4, characterized in that, In the deconvolution step 3, the calibration point spread function is obtained by fitting the actual PSF data using a double exponential Gaussian model. An iterative deconvolution algorithm combined with a regularization mechanism is used to achieve a balance between noise suppression and resolution improvement by dynamically controlling the number of iterations.

7. The method for precise capture and quantization of neural signals for low-light fluorescence microscopy as described in claim 4, characterized in that, When the total imaging time exceeds 1 minute or the observed object has physiological fluctuations, axial drift compensation is enabled. Specifically, it is implemented as follows: load the denoised image sequence and the baseline-corrected image sequence, identify the reference frame with the highest intensity; calculate the pixel-level difference map between the reference frame and the average frame, and generate an effective pixel mask through dynamic thresholding. Vectorized three-dimensional matrix operations are used to apply effective pixel weighted averaging to the baseline data to generate pixel-level baseline correction curves. Intensity normalization and offset compensation are performed using the baseline correction curves, and the resulting image sequence is output after axial drift compensation.

8. An imaging system, characterized in that, This system is used to implement the method described in any one of claims 1 to 7. The system includes an image acquisition module, a processing module, and a result output module. The image acquisition module is used to acquire a sequence of low-light fluorescence microscopy images. The processing module is a computing device equipped with a data processing environment, pre-installed with deconvolution tools, a deep learning toolbox, curve fitting tools, and dedicated analysis functions, supporting batch data parallel processing. The processing module integrates an adaptive optimization algorithm that can dynamically adjust processing parameters according to the characteristics of the input data. The result output module is used to output the processed image sequence, quantified data, and visualization results, and is compatible with mainstream analysis software.

Images