A fluorescence image background removal method based on a physical prior-driven training-free neural network

Abstract

This invention relates to a method for removing background from fluorescence images based on a physical prior-driven, training-free neural network, comprising: S1, establishing an untrained neural network and randomly determining the initial parameters of the neural network; S2, removing the background from the original observation image according to the current parameters of the neural network to obtain a background-removed observation image; S3, adding the estimated values ​​of the focus point spread function, the estimated values ​​of the defocus point spread function, and preset imaging noise to the background-removed observation image to obtain a reconstructed observation image with background; S4, establishing a loss function related to the preset imaging noise, calculating the loss, and obtaining a loss value; S5, cyclically executing steps S2 to S4, iteratively updating the parameters of the neural network according to the loss value until the convergence condition is met, determining the optimal parameters of the neural network, and removing the background from the original observation image using the neural network with the optimal parameters to obtain the optimal background-removed observation image.

Description

Technical Field

[0001] This invention relates to the field of image background removal, specifically to a method for removing fluorescent image background based on a physical prior-driven, training-free neural network. Background Technology

[0002] Fluorescence microscopy, with its excellent sensitivity and specificity, has become an indispensable research tool in the biomedical field, widely used in cell structure observation, molecular localization, and dynamic process analysis. High-quality fluorescence images are a prerequisite for accurate quantitative analysis and biological interpretation. However, in practical wide-field fluorescence imaging, fluorescence signals from defocused regions can create strong background interference, leading to a significant decrease in image contrast and signal-to-noise ratio, severely hindering clear identification and accurate measurement of target structures.

[0003] To suppress defocused backgrounds, frequency domain filtering is a common strategy in existing computational processing. This method relies on the assumption that background signals are concentrated in the low-frequency spatial spectrum and removes the background by suppressing low-frequency components. However, this method struggles to distinguish between the real, slowly changing background and the low-frequency structural information inherent in the sample itself, easily leading to the loss of useful signals and image distortion. Alternatively, deconvolution algorithms based on probabilistic models and optical transfer functions can effectively improve resolution and suppress some defocus blur under ideal conditions of accurate optical parameters and high image signal-to-noise ratio. However, these algorithms are extremely sensitive to noise. Under actual imaging conditions with low signal-to-noise ratios, the iterative process easily amplifies noise and produces artifacts, severely affecting the reliability and quantitative accuracy of the reconstruction results.

[0004] Deep learning-based methods have demonstrated powerful image processing capabilities. By utilizing large amounts of paired "background-image" data for supervised training, these methods can achieve end-to-end background suppression. However, their performance heavily relies on acquiring large-scale, high-quality paired datasets, which are typically time-consuming, costly, and even difficult to obtain in biological experiments. Furthermore, models trained on specific datasets often have limited generalization ability, and their performance may deteriorate sharply when applied to new data with different imaging conditions, sample types, or dye properties.

[0005] The purpose of this invention is to design a fluorescence image background removal method based on a physical prior-driven, training-free neural network to address the problems existing in the prior art. Summary of the Invention

[0006] To address the problems existing in the prior art, the present invention provides a method for removing background from fluorescent images based on a physical prior-driven, training-free neural network, which can effectively solve at least one of the problems existing in the prior art.

[0007] The technical solution of this invention is: A method for removing background from fluorescence images based on a physics-prior-driven, training-free neural network, comprising the following steps: S1, Establish an untrained neural network and randomly determine the initial parameters of the neural network; S2, Based on the current parameters of the neural network, remove the background from the original observed image to obtain a background-removed observed image; S3, add the estimated value of the focused point spread function, the estimated value of the defocused point spread function, and preset imaging noise to the background-reconstructed observation image to obtain the background-reconstructed observation image; S4, establish a loss function related to the preset imaging noise, calculate the loss between the reconstructed observation image and the original observation image using the loss function, and obtain the loss value; S5, repeat steps S2~S4, iteratively update the parameters of the neural network according to the loss value until the convergence condition is met, determine the optimal parameters of the neural network, and use the neural network with the optimal parameters to remove the background from the original observation image to obtain the optimal background-removed observation image.

[0008] Further, in step S3, the background-removed observation image is processed using the following formula: ; in, This represents a reconstructed observation image with background. It is to observe the image without background. and These are the in-focus spread function and the out-of-focus spread function, respectively. It is the preset imaging noise. Represents pixel coordinates.

[0009] Further, before step S4, the following steps are performed: the preset imaging noise is analyzed using a pre-trained noise decomposition model, the preset imaging noise is decomposed into multiple imaging noises, and the contribution ratio of each imaging noise is output. The imaging noise is classified into photon shot noise sources and detector readout noise sources, and the corresponding contribution ratios are weighted to obtain the photon shot noise contribution rate and the detector readout noise contribution rate.

[0010] Furthermore, the pre-trained noise decomposition model is trained using the following method: Extract multi-dimensional statistical features of simulated imaging noise containing known noise composition ratios, and construct an input feature vector; Extract the noise sources and their contribution percentages from the simulated imaging noise, which contain known noise composition proportions, and construct the output label vector. Using the input feature vector as input and the output label vector as output, a random forest regression model is trained to obtain the noise decomposition model.

[0011] Furthermore, the imaging noise is classified into photon shot noise sources and detector readout noise sources, and the corresponding contribution ratios are weighted to obtain the contribution rates of photon shot noise and detector readout noise, including: The higher the degree of coupling between the noise source and the in-focus signal and the background signal, or the higher the spatial non-uniformity of the noise source, the higher the weight assigned to the noise source. The contribution percentages and weights of imaging noise from the same source are weighted and summed to obtain the contribution of photon shot noise and detector readout noise.

[0012] Furthermore, the photon shot noise sources include sample fluorescence shot noise, background fluorescence shot noise, dark current shot noise, and stray light and ambient light leakage shot noise, with corresponding weights of 1, 0.8~1, 0.2~0.4, and 0.1~0.3, respectively. The sources of detector readout noise include pixel response nonuniformity and fixed pattern noise, source follower thermal noise, column amplifier and programmable gain amplifier noise, and analog-to-digital conversion quantization noise, with corresponding weights of 1, 0.5~0.7, 0.4~0.6, and 0.1~0.2, respectively.

[0013] Further, step S4 includes: S4.1, the shot noise variance of each pixel is calculated based on the pixel intensity and system gain coefficient of the reconstructed observation image. The readout noise variance is obtained by converting the system gain coefficient with the detector readout noise electron count. The shot noise variance and the readout noise variance are weighted and fused by the photon shot noise contribution and the detector readout noise contribution to obtain the predicted noise variance of each pixel in the grayscale domain. S4.2, Construct a spatially adaptive weighted loss function based on the predicted noise variance. The loss function uses the reciprocal of the predicted noise variance of each pixel as the weight of the pixel reconstruction error. Calculate the loss value of the reconstructed observation image and the original observation image using the weighted loss function.

[0014] Further, in step S4.1, the shot noise variance and the readout noise variance are weighted and fused using the photon shot noise contribution and the detector readout noise contribution to obtain the predicted noise variance for each pixel in the grayscale domain, including: The contribution of photon shot noise and the contribution of detector readout noise are normalized so that the sum of the normalized contribution of photon shot noise and the contribution of detector readout noise is 1. The prediction noise variance for each pixel in the grayscale domain = shot noise variance × normalized photon shot noise contribution + readout noise variance × normalized detector readout noise contribution.

[0015] Further, in step S4.2, a spatially adaptive weighted loss function is constructed based on the predicted noise variance. This loss function uses the reciprocal of the predicted noise variance for each pixel as the weight of the pixel reconstruction error, and includes: S4.2.1, construct the following loss function: , in, This represents the network parameters of a convolutional neural network. These represent the number of pixels horizontally and vertically in the image, respectively. Represents the coordinates of a pixel. This represents the variance of the prediction noise for the corresponding pixel in the grayscale domain. This represents the grayscale value of the corresponding pixel in the original observed image. This represents the grayscale value of the corresponding pixel in the reconstructed observed image. This represents a small regularization constant used to prevent the denominator from being zero.

[0016] Further, in step S5, the parameters of the neural network are iteratively updated based on the loss value until the convergence condition is met. The optimal parameters of the neural network are then determined, including: Calculate the gradient of the weighted loss function with respect to the parameters of the neural network, and update the parameters using the gradient descent optimization algorithm; The optimal parameters are obtained when the relative change of the weighted loss function value after a preset number of iterations is lower than a preset threshold, or when the number of iterations reaches a preset maximum number of iterations.

[0017] Therefore, the present invention provides the following effects and / or advantages: This application addresses the ill-conditioned inverse problem of recovering a sharp image from a blurred image. This research achieves excellent solution results by combining imaging physics mechanisms with neural networks. The method is based on the principle of optical path reversibility and the idea of ​​computational reconstruction. It uses a neural network to inversely estimate the object distribution before imaging, recovering the sharp image in focus; then, it uses a forward physical model of fluorescence imaging to calculate and generate the predicted blurred image. This closed-loop process constrains the training convergence of the neural network through physical mechanisms, thereby achieving adaptive optimization of the network parameters.

[0018] This application further decomposes a preset noisy image into multiple imaging noises, determines the contribution ratio of each imaging noise, and then weights them together using the weights corresponding to each imaging noise to obtain the contribution of photon shot noise and detector readout noise. The predicted noise variance of each pixel in the grayscale domain is constructed using the contribution of photon shot noise and detector readout noise. Finally, a spatially adaptive weighted loss function is constructed using the predicted noise variance, which automatically relaxes the fitting constraints in high-noise areas and retains fine weights in low-noise areas. Therefore, it is not necessary to pre-train for specific imaging conditions or sample types. The optimal network parameters of the convolutional neural network can be obtained after only a preset number of iterations, achieving an automatic balance between noise suppression and detail fidelity.

[0019] This application decomposes imaging noise into noise from multiple sources using a large model and assigns a corresponding contribution percentage to each noise source. This better reflects the true degree of harm of various noises to the background removal task, thus providing basic data for the subsequent integration of shot noise variance and readout noise variance. Furthermore, the predicted noise variance obtained by weighted fusion of contribution percentages has both spatial variability and noise source adaptability.

[0020] This application optimizes the loss function from the traditional method of uniform noise space, and changes the loss function from equal-weighted reconstruction error to noise variance-weighted reconstruction error. It also performs a refined physical modeling of the error term, enabling it to distinguish the noise confidence of different pixels.

[0021] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0022] It should be understood that the above summary and the following detailed description of the invention are exemplary and explanatory, and are intended to provide further explanation of the invention as claimed. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating one embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of a comparative experiment on fluorescence imaging of mouse brain slices using various background removal methods. Detailed Implementation

[0025] To facilitate understanding by those skilled in the art, the present invention will now be described in further detail with reference to the embodiments: Example 1 refer to Figure 1A method for removing background from fluorescence images based on a physics-prior-driven, training-free neural network includes the following steps: S1, Establish an untrained neural network and randomly determine the initial parameters of the neural network; S2, Based on the current parameters of the neural network, remove the background from the original observed image to obtain a background-removed observed image; S3, add the estimated value of the focused point spread function, the estimated value of the defocused point spread function, and preset imaging noise to the background-reconstructed observation image to obtain the background-reconstructed observation image; S4, establish a loss function related to the preset imaging noise, calculate the loss between the reconstructed observation image and the original observation image using the loss function, and obtain the loss value; S5, repeat steps S2~S4, iteratively update the parameters of the neural network according to the loss value until the convergence condition is met, determine the optimal parameters of the neural network, and use the neural network with the optimal parameters to remove the background from the original observation image to obtain the optimal background-removed observation image.

[0026] This embodiment proposes a physics-driven, training-free neural network method for suppressing defocused backgrounds in fluorescence images. This method embeds the forward physical process of fluorescence imaging—the PSF convolution operation in the image formation model—into the U-Net network architecture as a physical constraint to guide network optimization. No pre-training or externally labeled data is required; only a single image to be processed is needed to adaptively update the network parameters, effectively suppressing defocused backgrounds. This embodiment incorporates estimates of the in-focus and out-of-focus point spread functions, along with preset imaging noise, into the reconstructed observation image. If the background-removed observation image output by the neural network is clear, the resulting reconstructed observation image with background should be very close to the original observation image. In this case, the loss value calculated by the loss function should be sufficiently small. This principle is used to iteratively update the parameters of the neural network, thereby obtaining the optimal background-removed observation image.

[0027] Next, we will elaborate on how each step is specifically implemented.

[0028] The process of this application is as follows: Figure 1 As shown.

[0029] Further, in step S3, the background-removed observation image is processed using the following formula: (1) in, This represents a reconstructed observation image with background. It is to observe the image without background. and These are the in-focus spread function and the out-of-focus spread function, respectively. It is the preset imaging noise. Represents pixel coordinates.

[0030] In this step, the spatial distribution of the real sample is continuous. Within the depth of focus of the detection, it can be assumed that the three-dimensional structure of the sample does not change abruptly. Therefore, the two-dimensional sample distribution within the focal plane can be used to replace the three-dimensional sample within the detection area. First, the image captured under wide-field imaging conditions is represented as the convolution of the real sample and the system point spread function, plus the measured noise. Ignoring the differences in excitation and divergence wavelengths of fluorescence, the original observation image... It can be represented as: (2) in, and These represent the spatial coordinates of the image plane; and These represent the spatial coordinates of the object plane; It is the three-dimensional point spread function of the objective lens; It refers to the spatial distribution of objects; It is noise. In fluorescence microscopy imaging, the detector's detection of photon discreteness leads to shot noise, as well as circuit readout noise and dark current noise, resulting in Poisson-Gaussian mixed noise in fluorescence imaging. In the actual imaging process, fluorescence signals inside and outside the focal plane are excited, leading to aliasing detection of signals inside and outside the focal plane, thereby reducing the clarity of the image. It can be assumed that the article on fluorescence microscopy imaging considers noise to be the cause of the final image quality reduction. This assumption leads to the previous methods being able to recover the image well in the noise-dominated imaging process. However, as the imaging sample becomes thicker, when the proportion of the defocus background in the image is greater than that of the noise, the previous denoising schemes often have limited background removal effect. Within the depth of focus range of the detection, it can be assumed that the three-dimensional structure of the sample does not change abruptly, so the two-dimensional distributed sample in the focal plane can be used to replace the three-dimensional sample in the detection area. Then formula (2) can be written as formula (1).

[0031] In step S3, based on the assumption that the neural network can correctly solve this inverse problem, its clear output signal should be able to reconstruct the original input image through the forward fluorescence imaging process (Equation 1). Based on this assumption, we embed the forward imaging process into the network training as a physical constraint, enabling the network to achieve adaptive parameter optimization without labeled data. Therefore, using the convolutional imaging formula established in step S3, the estimated in-focus spread and out-of-focus spread estimates are superimposed with the preset imaging noise to obtain the reconstructed observation image.

[0032] refer to Figure 2 , Figure 2This is a comparative experiment using fluorescence imaging of mouse brain slices. (a) shows the network input (with background). (b) shows the effect of the RL deconvolution algorithm. (c) shows the background removal in the dark channel. (d) shows the N2N denoising effect. (e) shows the output of the neural network obtained in this embodiment (with background removed). (f) shows... Figure 2 (a)-(e) Normalized intensity maps of the underlined areas in the lower right corner. Figure 2 (a)-(e) Scale bar: 15 micrometers.

[0033] Example 2 This embodiment further performs the following steps based on Embodiment 1.

[0034] Furthermore, Before step S4, the following steps are performed: the preset imaging noise is analyzed using a pre-trained noise decomposition model, the preset imaging noise is decomposed into multiple imaging noises, and the contribution ratio of each imaging noise is output. The imaging noise is classified into photon shot noise sources and detector readout noise sources, and the corresponding contribution ratios are weighted to obtain the photon shot noise contribution rate and the detector readout noise contribution rate.

[0035] Specifically, imaging noise is classified into photon shot noise sources and detector readout noise sources. Photon shot noise sources can include sample fluorescence shot noise, background fluorescence shot noise, dark current shot noise, and stray light and ambient light leakage shot noise. Detector readout noise sources can include pixel response non-uniformity and fixed-mode noise, source follower thermal noise, column amplifier and programmable gain amplifier noise, and analog-to-digital conversion quantization noise. Since different noise sources have varying impacts on the preset imaging noise performance, they need to be decomposed and their corresponding contributions obtained. Simultaneously, the contributions of photon shot noise and detector readout noise are comprehensively considered based on different contribution levels and sources.

[0036] Furthermore, the pre-trained noise decomposition model is trained using the following method: Extract multi-dimensional statistical features of simulated imaging noise containing known noise composition ratios, and construct an input feature vector; Extract the noise sources and their contribution percentages from the simulated imaging noise, which contain known noise composition proportions, and construct the output label vector. Using the input feature vector as input and the output label vector as output, a random forest regression model is trained to obtain the noise decomposition model.

[0037] In this step, the simulated imaging noise with known noise composition ratios can first be obtained in a simulation environment or a known environment. Preferably, the simulated imaging noise with known noise composition ratios can be obtained by using a single variable. For example, the proportion of local fluorescence shot noise in the imaging noise can be set to 10%, and the proportion of other noises can be set to 0%. Then, the corresponding imaging noise can be obtained, allowing the model to learn to decouple the contribution ratio of photon shot noise and detector readout noise from statistical features.

[0038] Furthermore, the imaging noise is classified into photon shot noise sources and detector readout noise sources, and the corresponding contribution ratios are weighted to obtain the contribution rates of photon shot noise and detector readout noise, including: The higher the degree of coupling between the noise source and the in-focus signal and the background signal, or the higher the spatial non-uniformity of the noise source, the higher the weight assigned to the noise source. The contribution percentages and weights of imaging noise from the same source are weighted and summed to obtain the contribution of photon shot noise and detector readout noise.

[0039] In this step, the weights assigned to noise sources fully consider the signal dependence and spatial structure interference of imaging noise. If a noise source is directly coupled to the in-focus signal or background signal, or if its spatial distribution has a non-uniform structure, it will cause more severe interference to background separation. Therefore, it needs to be assigned a higher weight to highlight the influence of these noise components when calculating their contribution. See Tables 1 and 2 for details.

[0040] Table 1. Weighting of Photon Shot Noise Sources

[0041] Table 2, Weighting of Detector Readout Noise Sources

[0042] The contribution percentages and weights of imaging noise from the same source are weighted and summed to obtain the contribution of photon shot noise and detector readout noise.

[0043] In this step, imaging noise from the same source is grouped together, and then the contribution percentages and weights of the imaging noise in the same group are weighted and summed. This yields the combined contribution of photon shot noise and detector readout noise, representing both. Even though different noise components have the same energy, their interference with background separation varies significantly. For example, sample fluorescence shot noise is directly coupled to the signal, forming strong signal-dependent random fluctuations in bright areas. If its variance is not fully estimated, the network will treat the noise as a signal for fitting. In contrast, uniform dark current shot noise or quantization noise, while potentially having considerable energy, is spatially uniform and interferes with background estimation and structure restoration far less than the former two types. By calculating the contribution percentage and corresponding weights, high-hazard and low-hazard noise are treated differently, ensuring that the subsequent predicted noise variance accurately reflects the impact of noise on the reconstruction task at that pixel. Through differentiated weighting, the contribution of high-hazard noise is amplified, while low-hazard noise is suppressed.

[0044] Further, step S4 includes: S4.1, the shot noise variance of each pixel is calculated based on the pixel intensity and system gain coefficient of the reconstructed observation image. The readout noise variance is obtained by converting the system gain coefficient with the detector readout noise electron count. The shot noise variance and the readout noise variance are weighted and fused by the photon shot noise contribution and the detector readout noise contribution to obtain the predicted noise variance of each pixel in the grayscale domain. The noise variance of each pixel is calculated by squared the ratio of the system gain coefficient to the noise standard deviation of the number of noise electrons read out by the detector. This step is a direct adoption of existing technology.

[0045] In step S4.1, the shot noise variance and the readout noise variance are weighted and fused using the photon shot noise contribution and the detector readout noise contribution to obtain the predicted noise variance for each pixel in the grayscale domain, including: The contribution of photon shot noise and the contribution of detector readout noise are normalized so that the sum of the normalized contribution of photon shot noise and the contribution of detector readout noise is 1. The prediction noise variance for each pixel in the grayscale domain = shot noise variance × normalized photon shot noise contribution + readout noise variance × normalized detector readout noise contribution.

[0046] In this step, the contribution of photon shot noise and detector readout noise after normalization reflects the relative dominance of shot noise and readout noise in the formation of preset imaging noise. It dynamically captures the statistical characteristics of the overall noise composition in the image, so that the weighted fusion of noise variance is always synchronized with the actual noise situation of the current image.

[0047] S4.2, Construct a spatially adaptive weighted loss function based on the predicted noise variance. The loss function uses the reciprocal of the predicted noise variance of each pixel as the weight of the pixel reconstruction error. Calculate the loss value of the reconstructed observation image and the original observation image using the weighted loss function.

[0048] Further, in step S4.2, a spatially adaptive weighted loss function is constructed based on the predicted noise variance. This loss function uses the reciprocal of the predicted noise variance for each pixel as the weight of the pixel reconstruction error, and includes: S4.2.1, construct the following loss function: (3) in, This represents the network parameters of a convolutional neural network. These represent the number of pixels horizontally and vertically in the image, respectively. Represents the coordinates of a pixel. This represents the variance of the prediction noise for the corresponding pixel in the grayscale domain. This represents the grayscale value of the corresponding pixel in the original observed image. This represents the grayscale value of the corresponding pixel in the reconstructed observed image. This represents a small regularization constant used to prevent the denominator from being zero.

[0049] In this step, existing technologies generally use the following loss function: , in, It is the image output by the neural network. It contains both Gaussian and Poisson noise. As input to the neural network, it is an image containing an out-of-focus background. These are weighting coefficients. This is the regularization term. The loss function uses a completely uniform noise space, treating the reconstruction errors of all pixels equally, without distinguishing between high-noise and low-noise pixels. Represents norm operations.

[0050] This embodiment is optimized to the form of formula (3). The influence of the noise term no longer appears in an additive form, but is implicitly included in the predicted noise variance in the denominator as its reciprocal as a multiplicative weight. Furthermore, the loss function is changed from equal-weighted reconstruction error to reconstruction error weighted by noise variance. A refined physical model is performed on the error term to enable it to distinguish the noise reliability of different pixels. In addition, for each pixel in the fluorescence image, the error source between its true signal and the observed value is mainly imaging noise. In the previous steps, the total noise variance of the pixel position has been estimated by fusing physical modeling and noise contribution. Therefore, the reliability of the pixel observation is inversely proportional to the predicted noise variance. The larger the predicted noise variance, the less reliable the observation, and the lower the weight of its reconstruction error in the loss function should be. Through this weighted loss function, the untrained network can autonomously distinguish between reliable and unreliable pixels without any labeled data, achieving the optimal balance between background suppression and detail preservation.

[0051] In step S5, after each network parameter update, the background-removed observation image is regenerated, the observation image is reconstructed, and the noise variance map is updated. As the network parameters are gradually optimized, the original observation image becomes increasingly similar to the reconstructed observation image. The reconstruction error component in the reconstructed observation image gradually decreases, while the proportion of the noise component gradually increases. The contribution of the noise decomposition model output is also dynamically adjusted accordingly, resulting in more accurate prediction of the noise variance, which in turn provides positive feedback to the optimization of the loss weights in the next round. This closed-loop mechanism ensures the self-consistency and convergence stability of the entire method.

[0052] Further, in step S5, the parameters of the neural network are iteratively updated based on the loss value until the convergence condition is met. The optimal parameters of the neural network are then determined, including: Calculate the gradient of the weighted loss function with respect to the parameters of the neural network, and update the parameters using the gradient descent optimization algorithm; The optimal parameters are obtained when the relative change of the weighted loss function value after a preset number of iterations is lower than a preset threshold, or when the number of iterations reaches a preset maximum number of iterations.

[0053] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0054] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0055] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0056] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0057] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms should not be construed as necessarily referring to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

Claims

1. A method for removing background of fluorescence image based on physical prior-driven training-free neural network, characterized in that: Includes the following steps: S1, Establish an untrained neural network and randomly determine the initial parameters of the neural network; S2, Based on the current parameters of the neural network, remove the background from the original observed image to obtain a background-removed observed image; S3, add the estimated value of the focused point spread function, the estimated value of the defocused point spread function, and preset imaging noise to the background-reconstructed observation image to obtain the background-reconstructed observation image; The preset imaging noise is analyzed by a pre-trained noise decomposition model, which decomposes the preset imaging noise into multiple imaging noises and outputs the contribution ratio of each imaging noise. The imaging noise is classified into photon shot noise sources and detector readout noise sources, and the corresponding contribution ratios are weighted to obtain the contribution degree of photon shot noise and the contribution degree of detector readout noise. S4, establish a loss function related to the preset imaging noise, calculate the loss between the reconstructed observation image and the original observation image using the loss function, and obtain the loss value; S5, repeat steps S2~S4, iteratively update the parameters of the neural network according to the loss value until the convergence condition is met, determine the optimal parameters of the neural network, and use the neural network with the optimal parameters to remove the background from the original observation image to obtain the optimal background-removed observation image.

2. The method of claim 1, wherein the method is characterized by: In step S3, the background-removed observation image is processed using the following formula: ； wherein, denotes a reconstructed observation image with background, is a de-background observation image, and are in-focus point spread function and out-of-focus point spread function, respectively, is a preset imaging noise, denotes a pixel coordinate, denotes a convolution operation.

3. The method for removing background from fluorescence images based on a physical prior-driven, training-free neural network according to claim 1, characterized in that: The pre-trained noise decomposition model was trained using the following method: Extract multi-dimensional statistical features of simulated imaging noise containing known noise composition ratios, and construct an input feature vector; Extract the noise sources and their contribution percentages from the simulated imaging noise, which contain known noise composition proportions, and construct the output label vector. Using the input feature vector as input and the output label vector as output, a random forest regression model is trained to obtain the noise decomposition model.

4. The method for removing background from fluorescence images based on a physical prior-driven, training-free neural network according to claim 1, characterized in that: The imaging noise is classified into photon shot noise sources and detector readout noise sources, and the corresponding contribution ratios are weighted to obtain the contribution rates of photon shot noise and detector readout noise, including: The higher the degree of coupling between the noise source and the in-focus signal and the background signal, or the higher the spatial non-uniformity of the noise source, the higher the weight assigned to the noise source. The contribution percentages and weights of imaging noise from the same source are weighted and summed to obtain the contribution of photon shot noise and detector readout noise.

5. The method for removing background from fluorescence images based on a physical prior-driven, training-free neural network according to claim 4, characterized in that: The photon shot noise sources include sample fluorescence shot noise, background fluorescence shot noise, dark current shot noise, and stray light and ambient light leakage shot noise, with corresponding weights of 1, 0.8~1, 0.2~0.4, and 0.1~0.3, respectively. The sources of detector readout noise include pixel response nonuniformity and fixed pattern noise, source follower thermal noise, column amplifier and programmable gain amplifier noise, and analog-to-digital conversion quantization noise, with corresponding weights of 1, 0.5~0.7, 0.4~0.6, and 0.1~0.2, respectively.

6. The method for removing background from fluorescence images based on a physical prior-driven, training-free neural network according to claim 1, characterized in that: Step S4 includes: S4.1, the shot noise variance of each pixel is calculated based on the pixel intensity and system gain coefficient of the reconstructed observation image. The readout noise variance is obtained by converting the system gain coefficient with the detector readout noise electron count. The shot noise variance and the readout noise variance are weighted and fused by the photon shot noise contribution and the detector readout noise contribution to obtain the predicted noise variance of each pixel in the grayscale domain. S4.2, Construct a spatially adaptive weighted loss function based on the predicted noise variance. The loss function uses the reciprocal of the predicted noise variance of each pixel as the weight of the pixel reconstruction error. Calculate the loss value of the reconstructed observation image and the original observation image using the weighted loss function.

7. The method for removing background from fluorescence images based on a physical prior-driven, training-free neural network according to claim 6, characterized in that: In step S4.1, the shot noise variance and the readout noise variance are weighted and fused using the photon shot noise contribution and the detector readout noise contribution to obtain the predicted noise variance for each pixel in the grayscale domain, including: The contribution of photon shot noise and the contribution of detector readout noise are normalized so that the sum of the normalized contribution of photon shot noise and the contribution of detector readout noise is 1. The prediction noise variance for each pixel in the grayscale domain = shot noise variance × normalized photon shot noise contribution + readout noise variance × normalized detector readout noise contribution.

8. The method for removing background from fluorescence images based on a physical prior-driven, training-free neural network according to claim 6, characterized in that: In step S4.2, a spatially adaptive weighted loss function is constructed based on the predicted noise variance. This loss function uses the reciprocal of the predicted noise variance for each pixel as the weight of the pixel reconstruction error, and includes: S4.2.1, construct the following loss function: ， in, This represents the network parameters of a convolutional neural network. These represent the number of pixels horizontally and vertically in the image, respectively. Represents the coordinates of a pixel. This represents the variance of the prediction noise for the corresponding pixel in the grayscale domain. This represents the grayscale value of the corresponding pixel in the original observed image. This represents the grayscale value of the corresponding pixel in the reconstructed observed image. This represents a small regularization constant used to prevent the denominator from being zero.

9. The method for removing background from fluorescence images based on a physical prior-driven, training-free neural network according to claim 8, characterized in that: In step S5, the parameters of the neural network are iteratively updated based on the loss value until the convergence condition is met. The optimal parameters of the neural network are then determined, including: Calculate the gradient of the weighted loss function with respect to the parameters of the neural network, and update the parameters using the gradient descent optimization algorithm; The optimal parameters are obtained when the relative change of the weighted loss function value after a preset number of iterations is lower than a preset threshold, or when the number of iterations reaches a preset maximum number of iterations.

Images