A method for constructing a fluorescence microscopy signal detection model and its application

Abstract

This invention relates to the field of signal recognition and localization technology, and discloses a method for constructing a fluorescence microscopy signal detection model and its application. This invention is a fitting method based on self-supervised deep learning. By representing fluorescence microscopy signals as Pattern of Interest (POI) or combinations thereof, and utilizing a self-supervised deep learning framework to perform parallel POI fitting and global denoising in a large receptive field, it achieves ultra-high-density fluorescence microscopy signal detection in a non-supervised learning manner under severe noise and strong fluorescence backgrounds, while maintaining high versatility and ease of use. This method can be easily adapted to various fluorescence microscopy signal detection tasks by simply setting POI examples provided to the fluorescence microscopy signal detection model and adjusting a sensitivity-related hyperparameter, achieving significant performance improvements in all cases.

Description

Technical Field

[0001] This invention relates to the field of signal recognition and localization technology, and in particular to a method for constructing a fluorescence microscopy signal detection model and its application. Background Technology

[0002] Fluorescence microscopy, with its unique ability to visualize and quantify vivid details of molecular and cellular processes, greatly contributes to the analysis and understanding of biological mechanisms. However, due to the influence of sample conditions, sensor performance, and imaging environment, fluorescence microscopy images often suffer from problems such as weak signals, overlap, multi-source compound noise contamination, and complex background interference, which hinder accurate and efficient signal detection.

[0003] Currently, various fluorescence microscopy image processing methods exist to address the aforementioned problems, but none fully meet the needs of different fields and applications. For example, methods based on spatiotemporal redundancy utilize the unpredictability of noise and the redundancy of signals to achieve denoising and signal enhancement. However, these methods are unsuitable for processing rapidly changing signals in time and space, and cannot determine the presence and location of signals. Furthermore, they can only remove noise, not distinguish between signals and background with similar spatiotemporal redundancy. Another example is benchmark-based methods, which train models on labeled datasets to learn the mapping between images and their corresponding benchmark signals for signal detection. However, this method requires a large number of high-quality, manually labeled tags, increasing the workload and cost of signal processing. Moreover, for some high-precision labeling tasks, such as Single Molecule Localization Microscopy (SMLM), manual labeling is impractical. Finally, methods based on conventional fitting use algorithms such as maximum likelihood estimation (MLE) to sequentially fit signals from images for signal detection. However, their inherent characteristics, such as sequence fitting and coarse modeling of interference factors, significantly degrade performance when processing weak or overlapping signals, severe noise, or complex backgrounds.

[0004] In summary, current methods face trade-offs between versatility, performance, and practicality, and a universal, high-performance, and easy-to-use fluorescence microscopy signal detection method is still lacking under various complex conditions. Summary of the Invention

[0005] This invention provides a method for constructing a fluorescence microscopy signal detection model and its application, in order to address the lack of a universal, high-performance and easy-to-use fluorescence microscopy signal detection method.

[0006] This invention provides a method for constructing a fluorescence microscopy signal detection model, comprising:

[0007] Acquiring raw fluorescence microscopy images Iori Among them, the original fluorescence microscopy image I ori The fluorescence microscopy signals in the samples have been characterized as Pattern of Interest (POI) or combinations thereof;

[0008] Initialize a deep learning model

[0009] Using deep learning models Original fluorescence microscope image I ori Convert to fitted graph I fit Among them, the fitted graph I fit Encoded with original fluorescence microscopy images I ori Location information and magnitude of POIs in the middle;

[0010] Fitted Figure I fit Group convolution with preset POI examples yields denoised image I. dns ;

[0011] Maximize the original fluorescence microscope image I ori and denoised image I dns The L2 similarity between them, and minimizing the fitted graph I fit The L1 regularization term is used to optimize the deep learning model.

[0012] Initialize the decoding module

[0013] Using the decoding module From the fitted graph I fit The original fluorescence microscope image I was obtained by decoding. ori Location information of POIs in the middle;

[0014] From the original fluorescence microscope image I ori Medium subtraction denoising image I dns The noisy image I is obtained. ns ;

[0015] Noise image I ns With Fitted Figure I fit The sum along the channel dimension By merging the samples, a noise fitting plot can be obtained. And based on the noise fitting graph The original fluorescence microscope image I was estimated. ori The magnitude of POI;

[0016] Combined with the estimated original fluorescence microscopy image I ori Amplitude of POI, example of preset POI, and the decoded raw fluorescence microscope image Iori The location information of the POI is used to obtain the synthetic image I. syn ;

[0017] Maximize the synthesized image I syn and denoised image I dns L1 similarity between them to optimize the decoding module.

[0018] Combined with the optimized deep learning model and decoding module A fluorescence microscopy signal detection model was obtained.

[0019] According to the method for constructing a fluorescence microscopy signal detection model provided by the present invention, the expression of the preset POI example is:

[0020]

[0021] In the expressions of the default POI examples, Let C represent a set of C pre-obtained POI images along the channel dimension, where (x,y) represents the pixel coordinates. Let represent the minimum value of the c-th POI image. This represents the maximum value of the c-th POI image.

[0022] The present invention provides a method for constructing a fluorescence microscopy signal detection model, a deep learning model. A fully convolutional neural network with a U-net architecture, a deep learning model. loss function for:

[0023]

[0024] Deep learning models loss function middle, I ori Represents the original fluorescence microscope image, I dns I represents a denoised image. fit Let λ represent the fitted graph, and λ represent the regularization strength.

[0025] According to the present invention, a method for constructing a fluorescence microscopy signal detection model is provided, wherein the fitted figure I is... fit Group convolution with preset POI examples yields denoised image I. dns ,include:

[0026] According to the first expression, the fitted graph I will be... fit Group convolution with preset POI examples yields denoised image I. dns The first expression is:

[0027]

[0028] In the first expression, I dns I represents a denoised image. fit This represents a fitted plot, where the colon ":" indicates all x or y coordinate values ​​in that dimension. Represents two-dimensional convolution. A preset POI example representing the sum of the mean normalized.

[0029] According to the method for constructing a fluorescence microscopy signal detection model provided by the present invention, Figure I is fitted. fit The sum along the channel dimension The expression is:

[0030]

[0031] According to the present invention, a method for constructing a fluorescence microscopy signal detection model is provided, wherein the decoding module is utilized. From the fitted graph I fit The original fluorescence microscope image I was obtained by decoding. ori Location information of POIs, including:

[0032] Based on the fitted graph I fit The sum along the channel dimension The second expression is used to segment and obtain pixel regions that may contain POIs. A set;

[0033] Based on pixel regions that may contain POIs The set of values ​​is used to calculate the centroid estimate of the i-th Region of Interest (ROI) through the third expression. Subpixel coordinates of POI As the original fluorescence microscope image I ori Location information of POIs in the middle.

[0034] According to the method for constructing a fluorescence microscopy signal detection model provided by the present invention, the second expression is:

[0035]

[0036] In the second expression, θ represents the fitted graph I. fit The sum along the channel dimension The pre-initialized segmentation threshold is set within the range of pixel values.

[0037] According to the method for constructing a fluorescence microscopy signal detection model provided by the present invention, the third expression is:

[0038]

[0039] According to the present invention, a method for constructing a fluorescence microscopy signal detection model is provided, wherein the original fluorescence microscopy image I... ori Medium subtraction denoising image I dns The noisy image I is obtained. ns ,include:

[0040] From the original fluorescence microscope image I ori Medium subtraction denoising image I dns The noisy image I is obtained. ns I ns =I ori -I dns ;

[0041] Preset POI example using mean sum normalization The maximum value for noisy image I ns Normalize.

[0042] According to the present invention, a method for constructing a fluorescence microscopy signal detection model is provided, wherein the noisy image I... ns With Fitted Figure I fit The sum along the channel dimension By merging the samples, a noise fitting plot can be obtained. And based on the noise fitting graph The original fluorescence microscope image I was estimated. ori The magnitude of POIs includes:

[0043] The normalized noisy image I is obtained through the fourth expression. ns With Fitted Figure I fit The sum along the channel dimension By merging the samples, a noise fitting plot can be obtained.

[0044] The fifth expression is used to calculate the maximum value after convolving each ROI with the preset POI example corresponding to its z-axis position, thus estimating the i-th ROI. The amplitude A of the included POI i .

[0045] According to the method for constructing a fluorescence microscopy signal detection model provided by the present invention, the fourth expression is:

[0046]

[0047] In the fourth expression, Represents the normalized noisy image I ns The value at pixel coordinates (x, y). Indicates the fitted graph Ifit The sum along the channel dimension The value at pixel coordinates (x, y). Represents the noise fitting plot The value at pixel coordinates (x, y).

[0048] According to the method for constructing a fluorescence microscopy signal detection model provided by the present invention, the fifth expression is:

[0049]

[0050] In the fifth expression, "⊙" represents matrix element-wise multiplication, and M(·) represents a function that outputs a mask: pixels within the input set are 1, and all others are 0; at position... The preset POI example obtained is obtained through interpolation.

[0051] According to the present invention, a method for constructing a fluorescence microscopy signal detection model is provided, wherein the original fluorescence microscopy image I obtained by combining the estimation is... ori Amplitude of POI, example of preset POI, and the decoded raw fluorescence microscope image I ori The location information of the POI is used to obtain the synthetic image I. syn ,include:

[0052] The sixth expression, combined with the estimated original fluorescence microscope image I, is used to... ori Amplitude of POI, example of preset POI, and the decoded raw fluorescence microscope image I ori The location information of the POI is used to obtain the synthetic image I. syn The sixth expression is:

[0053]

[0054] In the sixth expression, H and W represent the height and width of the preset POI example, respectively, in position... and The pixel values ​​of the preset POI example are obtained through interpolation.

[0055] According to the present invention, a method for constructing a fluorescence microscopy signal detection model includes a decoding module. The loss function is:

[0056]

[0057] According to the present invention, a method for constructing a fluorescence microscopy signal detection model is provided, wherein the fluorescence microscopy signal detection model uses any one of the following or any combination thereof to optimize the model performance: center-aligned upsampling, non-uniform background learning, and POI diffusion.

[0058] The present invention also provides a method for detecting fluorescence microscopy signals, comprising:

[0059] Receive fluorescence microscope images to be tested;

[0060] Based on the fluorescence microscope image to be tested, the fluorescence microscopy signal detection model obtained by constructing the fluorescence microscopy signal detection model as described in any of the above-mentioned methods is used to obtain the fluorescence microscopy signal detection result of the fluorescence microscope image to be tested, wherein the fluorescence microscopy signal detection result includes the position information and / or amplitude information of the fluorescence microscopy signal.

[0061] The present invention also provides a fluorescence microscopy signal detection system, comprising:

[0062] The image receiving module is used to receive fluorescence microscope images under test.

[0063] The signal detection module is used to: obtain the fluorescence microscopy signal detection result of the fluorescence microscopy image under test based on the fluorescence microscopy image under test, using the fluorescence microscopy signal detection model obtained by any of the above-described fluorescence microscopy signal detection model construction methods, wherein the fluorescence microscopy signal detection result includes the position information and / or amplitude information of the fluorescence microscopy signal.

[0064] The present invention also provides an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the computer program to implement the method for constructing a fluorescence microscopy signal detection model and / or the fluorescence microscopy signal detection method described above.

[0065] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for constructing the fluorescence microscopy signal detection model and / or the fluorescence microscopy signal detection method described above.

[0066] The present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer is able to execute any of the above-described methods for constructing a fluorescence microscopy signal detection model and / or for detecting fluorescence microscopy signals.

[0067] This invention provides a method for constructing a fluorescence microscopy signal detection model and its application. This method achieves high-throughput and high-sensitivity fluorescence microscopy signal detection under strong interference by simply setting POI examples provided to the fluorescence microscopy signal detection model (referred to herein as the DEPAF model) and adjusting a sensitivity-related hyperparameter, without requiring any benchmark ground truth or domain-specific imaging physics models. While maintaining performance, DEPAF's prediction speed for locating low-density POIs is comparable to traditional fitting-based methods, but it outperforms them when using high-density POIs. This is because the computational complexity of DEPAF's first branch depends only on the image size, not the number of POIs; although the second branch involves sequential processing, it does not include any iterative processes found in traditional fitting-based methods. This invention fully demonstrates the high versatility, performance, and practicality of DEPAF in various tasks. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0069] Figure 1 This is one of the flowcharts for a fluorescence microscopy signal detection method provided by the present invention.

[0070] Figure 2 This is a second schematic diagram of a fluorescence microscopy signal detection method provided by the present invention. In the figure, a shows the first branch of the method flow; b shows the second branch of the method flow; and c shows a fully convolutional neural network based on U-net. The structure is used to fit all POIs in the input image and denoise the input image. This neural network includes a downsampling stage followed by an upsampling stage. Each stage consists of multiple (de)convolutional layers with 3×3 filters, and non-trainable operations including max pooling (stride of 2), Dropout (probability of 50%), and Rectified Linear Unit (ReLU) activation. The first convolutional layer in the downsampling stage has N filters, where N is the maximum of 64 and the number of POIs. In each downsampling stage, the resolution is halved while the number of filters is doubled; the opposite is true in each upsampling stage. Black arrows indicate skip connections. Figure d shows the decoding module. The prior knowledge used. The diagram illustrates the decoding module. The processing steps are used to decode the location information of POIs from the fitted image. Conv. represents a convolutional layer; MaxPool. represents a max-pooling layer; Deconv. represents a deconvolutional layer.

[0071] Figure 3 This is a schematic diagram of the structure of a fluorescence microscopy signal detection system provided by the present invention.

[0072] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0074] Combination Figures 1 to 4 This invention describes a method for detecting fluorescence microscopic signals. The method can be divided into two aspects: the construction and application of a fluorescence microscopic signal detection model. In the method for constructing the fluorescence microscopic signal detection model, steps S110-150 constitute the first branch of the method flow, steps S210-260 constitute the second branch, and step S310 is the final step after the completion of the first and second branches. The execution subject of this fluorescence microscopic signal detection method can be any applicable terminal-side device or network-side device, such as a fluorescence microscopic signal detection apparatus.

[0075] See Figure 1 The present invention provides a method for detecting fluorescence microscopic signals, which may include:

[0076] S110, Obtain the original fluorescence microscope image I ori Among them, the original fluorescence microscopy image I ori Fluorescence microscopy signals in the samples have been characterized as Pattern of Interest (POI) or combinations thereof.

[0077] S120. Initialize the deep learning model

[0078] S130, Utilizing Deep Learning Models Original fluorescence microscope image I ori Convert to fitted graph I fit Among them, the fitted graph I fit Encoded with original fluorescence microscopy images i ori Location information and magnitude of POIs.

[0079] S140, Fitting graph i fit Group convolution with preset POI examples to obtain denoised image i dns .

[0080] S150, Maximize the original fluorescence microscope image i ori and denoised image I dns The L2 similarity between them, and minimizing the fitted graph I fit The L1 regularization term is used to optimize the deep learning model.

[0081] In this embodiment, a set of C pre-obtained POI images along the channel dimension are created in advance. It is then standardized to the 0-1 range to be modeled as a preset POI example. The expression for the preset POI example is:

[0082]

[0083] In the expressions of the default POI examples, Let c represent a set of c pre-obtained POI images along the channel dimension, where (x,y) represents the pixel coordinates. Let represent the minimum value of the c-th POI image. This represents the maximum value of the c-th POI image.

[0084] In this embodiment, based on the first expression, the fitted graph I is... fit Group convolution with preset POI examples to obtain denoised image i dns The first expression is:

[0085]

[0086] In the first expression, i dns Represents a denoised image, i fit This represents a fitted plot, where the colon ":" indicates all x or y coordinate values ​​in that dimension. Represents two-dimensional convolution. A preset POI example representing the sum of the mean normalized.

[0087] The deep learning model in this embodiment This is a fully convolutional neural network with a U-net architecture, used to simultaneously fit all points of interest (POIs) in the input image and denoise the input image; it is a deep learning model. loss function for:

[0088]

[0089] Deep learning models loss function middle, I ori Represents the original fluorescence microscope image, I dns I represents a denoised image. fit Let λ represent the fitted graph, and λ represent the regularization strength.

[0090] Among them, the loss function It includes an L2 term and an L1 regularization term. The L2 term promotes I... dns and I ori The similarity between them is used to achieve POI reconstruction in the image, and the L1 regularization term promotes I fit The sparsity of λ is used to achieve denoising. The regularization intensity λ is an adjustable hyperparameter, thus controlling the denoising intensity and the sensitivity of POI detection. This embodiment can effectively achieve denoising through sparse I... fit This method suppresses all other image information that differs from the spatial distribution of the Point of Interest (POI), thereby eliminating various composite noises. It does not require noise modeling and relies solely on POI information.

[0091] After completing the first branch in the method flow, start the second branch.

[0092] The second branch of this embodiment uses an optimizable decoding module. From I fit The POI location is decoded from the code. Because the POI example is related to I... ori The noise in the I (not just POI) also has similarities, fit This may contain fitted values ​​generated by noise rather than POIs. These noisy fitted values ​​typically do not exceed the noise level. Therefore, this embodiment uses a threshold θ to filter them out. However, since the noise level is difficult to estimate, this embodiment employs an iterative optimization method to robustly address this issue. Specifically, this embodiment first optimizes I along the channel dimension. fit Summing to obtain That is, the fitted graph I fit The sum along the channel dimension The expression is:

[0093]

[0094] S210, Initialize the Optimizable Decoding Module Used to decode the location of POI from the fitted graph.

[0095] S220, using the decoding module From the fitted graph I fit The original fluorescence microscope image I was obtained by decoding. ori Location information of POIs in the middle.

[0096] Specifically, this embodiment is based on the fitted graph I. fit The sum along the channel dimension The second expression is used to segment and obtain pixel regions that may contain POIs. The set; where the second expression is:

[0097]

[0098] In the second expression, θ represents the fitted graph I. fit The sum along the channel dimension The initial segmentation threshold is set within the range of pixel values.

[0099] To further delineate the Region of Interest (ROI) containing a single POI, this embodiment summarizes the following two priors:

[0100] (a) Shape prior: Due to the discrete nature and inherent size of pixels, a point in continuous space (i.e. a point in subpixel coordinates in the image) is represented in discrete pixel space as 1 to 4 pixels in a 2×2 pixel block, which corresponds to the shape of each ROI.

[0101] (b) Priority Prior: For ROIs containing 1 to 4 non-zero pixels, the range of sub-pixel coordinates they can represent increases sequentially. According to the classical probability model, their probabilities of occurrence also increase sequentially. Therefore, their partitioning priority also increases sequentially.

[0102] Therefore, based on the above priors, the ROI containing the fitting position and magnitude of a single POI is divided using Algorithm 1.

[0103]

[0104] It is worth noting that computational efficiency can be significantly improved by processing each 8-separated pixel region separately (i.e., all pixels in each region have no 8-neighborhood relationship with each other), rather than applying the algorithm directly to the large original image.

[0105] Then, this embodiment is based on pixel regions that may contain POIs. The set of values ​​is used to calculate the centroid estimate of the i-th ROI through the third expression. Subpixel coordinates of POI As the original fluorescence microscope image I ori The location information of the POI, where the third expression is:

[0106]

[0107] S230, from the original fluorescence microscope image I ori Medium subtraction denoising image I dns The noisy image I is obtained. ns I ns =I ori -I dns And using the preset POI example with mean sum normalization. The maximum value for noisy image I ns Normalize.

[0108] S240, Transfer the noisy image I ns With Fitted Figure I fit The sum along the channel dimension By merging the samples, a noise fitting plot can be obtained. And based on the noise fitting graph The original fluorescence microscope image I was estimated. ori The magnitude of POIs in the middle.

[0109] Specifically, in this embodiment, the normalized noisy image I is expressed using the fourth expression. ns With Fitted Figure I fit The sum along the channel dimension By merging the samples, a noise fitting plot can be obtained. The fifth expression is used to calculate the maximum value after convolving each ROI with the preset POI example corresponding to its z-axis position, thus estimating the i-th ROI. The amplitude A of the included POI i .

[0110] The fourth expression is:

[0111]

[0112] In the fourth expression, Represents the normalized noisy image I ns The value at pixel coordinates (x, y). Indicates the fitted graph I fit The sum along the channel dimension The value at pixel coordinates (x, y). Represents the noise fitting plot The value at pixel coordinates (x, y).

[0113] The fifth expression is:

[0114]

[0115] In the fifth expression, "⊙" represents matrix element-wise multiplication, and M(·) represents a function that outputs a mask: pixels within the input set are 1, and all others are 0; at position... The preset POI example obtained is obtained through interpolation.

[0116] S250. Using the sixth expression, combined with the estimated original fluorescence microscope image I ori Amplitude of POI, example of preset POI, and the decoded raw fluorescence microscope image I ori The location information of the POI is used to obtain the synthetic image I. syn The sixth expression is:

[0117]

[0118] In the sixth expression, H and W represent the height and width of the preset POI example, respectively, in position... and The pixel values ​​of the preset POI example are obtained through interpolation.

[0119] S260, Maximize the synthesized image I syn and denoised image I dns L1 similarity between them to optimize the decoding module.

[0120] Based on the above construction, a key trade-off emerges: choosing a θ value lower than the noise level leads to... The noise-fitted values ​​in I are preserved. This will result in... syn Multiple generated from Noise POIs derived from noise in the context of [the noise source]. Due to the randomness and instability of noise, these noise POIs cannot be correlated with [the noise source]. dns The virtual POI matching in the model is achieved by convolving the noise-fitted values ​​with POI examples. Therefore, I... syn Will deviate from I dns Conversely, choosing an excessively large θ value above the noise level can lead to incorrect filtering. The non-noise fitted value in I. This results in syn The non-noise POI generated in the ratio I dns The amount in the middle is less, which also leads to I syn Deviation from I dns Therefore, in this embodiment, θ can be obtained by maximizing I. dns and I synThe similarity between them is used for optimization, that is, minimizing the loss function, decoding module The loss function is:

[0121]

[0122] because Since θ is non-differentiable and the optimization is univariate, this embodiment uses a gradient-free optimization method. Specifically, this embodiment first uses Bayesian optimization to quickly estimate... The curve is shaped and a suboptimal θ value is found. Then, a grid search is performed around this suboptimal θ value to find the optimal θ value. This method aims to balance the speed and accuracy of optimization.

[0123] After completing the first and second branches in the method flow, S130 can be executed.

[0124] S310, combined with the optimized deep learning model and decoding module A fluorescence microscopy signal detection model was obtained.

[0125] In one embodiment, the fluorescence microscopy signal detection model may employ any of the following or any combination thereof to optimize model performance: center-aligned upsampling technique to improve the accuracy of POI localization; non-uniform background learning to enhance anti-interference capability; and POI diffusion to promote model convergence in multi-POI detection scenarios.

[0126] (a) Center-aligned upsampling: In the second branch of the method, the inherent pixel size may limit... The ability to distinguish overlapping POIs and accurately estimate their locations is crucial. To address this limitation, the pixel size can be reduced by applying the same upsampling factor to the POI examples and the input image at the start of the pipeline. However, an excessively large upsampling factor may cause the method to falsely detect a single POI as multiple POIs, increasing the false alarm rate. Therefore, in scenarios requiring high localization accuracy and the ability to distinguish overlapping POIs (such as SMLM), this embodiment uses an upsampling factor of 1.5. In general, a factor of 1 (i.e., no upsampling) is sufficient.

[0127] This upsampling should also preserve the shape of the POI example and the POI in the input image, as well as the center position of the POI example. The shape of the POI is affected by the interpolation algorithm used during the upsampling process. Different tasks may require different interpolation algorithms. For example, cubic spline interpolation has been shown to accurately capture the shape of any point spread function (PSF), so it was attempted to be used in the PSF-related task in this embodiment. However, this embodiment found that the choice of interpolation algorithm is not critical, and cubic spline interpolation is effective even for non-PSF-related tasks. Furthermore, the same interpolation algorithm used to obtain subpixel values ​​in the pipeline is the same as that used here.

[0128] To illustrate how to maintain the center position of the POI example after upsampling, this embodiment considers the following single-channel two-dimensional POI example with height H and width W. Let the pixel coordinate vector along the x-axis in the POI example be:

[0129] x ori = (0, 1, ..., W-1).

[0130] Let the initial upsampling factor be k > 1. To ensure that the side length of the POI example remains odd after upsampling, this embodiment first corrects the upsampling factor in the following way:

[0131]

[0132] Then, using the interpolation step size and upsampling width x ori Expand to x exp :

[0133]

[0134] Then for x ori and x exp Perform center alignment to obtain the x-axis coordinates of the interpolation query point. qry :

[0135]

[0136] in and They are x ori and x exp The mean. Similarly, the y-axis coordinate of the interpolation query point. qry The same steps can be used to calculate it. Then, the coordinates of the interpolation query point are obtained through x. qry and y qry The complete mesh construction is as follows:

[0137]

[0138] To simplify the calculation, the upsampling process for the input image is the same as that for the POI examples, except that the side lengths after upsampling must remain odd. Therefore, after POI localization of the upsampled image, the POI position can be mapped back to the original coordinate system in the following way:

[0139]

[0140] Where x up This is the x-axis coordinate of the POI in the upsampled image. Similarly, It can also be calculated accordingly.

[0141] (b) Non-uniform background learning. Non-uniform backgrounds in the image can affect the performance of POI detection. To address this issue, this embodiment adds an additional constant image patch (i.e., a small image with uniform values ​​everywhere) three times the size of the central region of the 95% pixel value distribution of the signal POI example as a background POIP. bg and use Output channel generates background fitting image I fitBG This method enables the model to learn arbitrary non-uniform backgrounds when detecting POIs. When applying this technique, it is used for optimization. The loss function becomes:

[0142]

[0143] Note that because the background does not exhibit natural sparsity, therefore I fitBG It is not subject to L1 regularization.

[0144] (c) POI diffusion. When this method is applied to a multi-POI detection task, the number of POI examples determines the deep learning model. Output I fit The number of channels. Therefore, as the number of POI examples increases, I fit The search space will also expand, which may affect... To address this issue, this embodiment introduces a POI diffusion technique. This technique first merges all POI examples into a central channel using mean pooling. In the last convolutional layer, only the kernel corresponding to the central channel is activated for training. During training, if the validation loss does not decrease after a certain number of iterations, POI diffusion is triggered. At this point, new activation channels are determined using a binary search method, and POI examples are re-merged into their nearest activation channels using mean pooling. The kernels corresponding to these new activation channels inherit the initial weights of the previously activated kernels and are activated for training. This process is repeated until all channels are activated. This technique aims to make... The optimal solution is found within a smaller search space, and then progressively expanded to a larger search space, thus achieving incremental learning. In this process, the frequency of POI examples is minimized by initially merging them, and then gradually increased through POI diffusion during training. This aligns with the natural tendency of the model to gradually fit from low-frequency patterns to high-frequency patterns. These factors contribute to the significant effect of this technique in improving the convergence of this method in multi-POI detection tasks.

[0145] The fluorescence microscopy signal detection model obtained by the method for constructing a fluorescence microscopy signal detection model provided by this invention can be applied to the identification of various fluorescence signals (e.g., dynamic fluorescence signals), forming a fluorescence microscopy signal detection method, which may include the following steps: obtaining a fluorescence microscope image to be tested; based on the fluorescence microscope image to be tested, obtaining the fluorescence microscopy signal detection result of the fluorescence microscope image to be tested by the fluorescence microscopy signal detection model obtained by the method for constructing the fluorescence microscopy signal detection model described above, wherein the fluorescence microscopy signal detection result includes the position information and / or amplitude information of the fluorescence microscopy signal.

[0146] Dynamic fluorescence signals include, but are not limited to, FISH signals, CRISPR signals, and nanosignals.

[0147] The FISH signals include, but are not limited to, the following types: (a) Conventional DNA / RNA FISH: used to detect and locate specific DNA or RNA sequences. (b) Multiplex DNA / RNA FISH: used to simultaneously detect multiple different DNA or RNA sequences, including but not limited to MERFISH, seqFISH, and seqFISH+. (c) Quantitative DNA / RNA FISH: used to quantitatively determine the quantity of target DNA or RNA. (d) Alternating DNA / RNA FISH: used to detect complex genomic rearrangements. (e) Color-coded DNA / RNA FISH: used to distinguish multiple target sequences by different colors. (f) High-resolution DNA / RNA FISH: used to achieve fine localization at the subcellular level. (g) Super-resolution DNA / RNA FISH: combined with super-resolution microscopy to improve detection accuracy. (h) Reversible DNA / RNA FISH: Allows the same sample to be reused for the detection of different targets. (i) Rapid DNA / RNA FISH: Used to accelerate the detection process. (j) Automated DNA / RNA FISH: Used for high-throughput automated analysis. (k) All other fluorescence FISH methods published to date, including but not limited to: Genomic in situ hybridization (GISH), multicolor fluorescence in situ hybridization (M-FISH), primer in situ DNA synthesis (PRINS), in situ PCR hybridization (IS-PCR), DNA fiber fluorescence in situ hybridization (Fiber-FISH), bacterial artificial chromosome fluorescence in situ hybridization (BAC-FISH), tyramine signal amplification fluorescence in situ hybridization (TSA-FISH), molecular beacon fluorescence in situ hybridization (MB-FISH), peptide nucleic acid fluorescence in situ hybridization (PNA-FISH), flow cytometry fluorescence in situ hybridization (FlowFISH), Oligopaint and its related technologies, HD-FISH, Cas-FISH, CRISPRLiveFISH, GOLD FISH, etc.

[0148] CRISPR signals include, but are not limited to, the following types: (a) CRISPR-Cas9: a conventional CRISPR system for gene editing. (b) CRISPR-Cas12: a CRISPR system with an extended target recognition range. (c) CRISPR-Cas13: a CRISPR system for RNA editing. (d) CRISPRa (CRISPR activation): used for gene activation. (e) CRISPRi (CRISPR interference): used for gene repression. (f) Base Editing CRISPR: used for precise single-base modifications. (g) Pulsed CRISPR: used for time-controlled gene editing.

[0149] Nanofluorescent signals include, but are not limited to, the following types: (a) Nanoparticle signals: used for labeling and tracking specific molecules. (b) Quantum dot signals: used for high-brightness and high-stability fluorescent labeling. (c) Nanotube signals: used to enhance signal transmission and detection. (d) Gold nanoparticle signals: used to enhance optical contrast. (e) Nanoshell signals: used for multifunctional diagnostics and therapy. (f) Nanoprobe signals: used for highly sensitive molecular detection.

[0150] The fluorescence microscopy signal detection model provided by this invention can also realize a fluorescence microscopy signal detection method based on fluorescence lifetime. The fluorescence microscopy signals include, but are not limited to, the following types: (a) Fluorescence lifetime imaging microscopy signals (FLIM): used to measure the molecular environment and interactions. (b) Lifetime fluorescence sensor signals: used for real-time monitoring of intracellular processes. (c) Time-resolved fluorescence signals: used to improve detection sensitivity. (d) Two-photon fluorescence lifetime signals: used for imaging deep tissues. (e) Multiphoton fluorescence lifetime signals: used for imaging complex biological samples.

[0151] The fluorescence microscopy signal detection method provided by this invention can be used for the following operations, including but not limited to: (a) super-resolution imaging; (b) FISH fluorescence signal point detection; (c) denoising, spike signal estimation, and neuron segmentation based on neuronal calcium imaging images and videos, two-photon neuronal calcium imaging images and videos, and three-photon or higher neuronal calcium imaging images and videos; and (d) accurate estimation of arbitrary non-uniform fluorescence background.

[0152] The following are specific application examples:

[0153] (1) Using calibrated PSF images as examples of points of interest (POIs), DEPAF enables millisecond-level dynamic super-resolution imaging. By optimizing the direct stochastic optical reconstruction microscopy (dSTORM) technique, SMLM imaging was performed on dynamic live-cell samples, achieving an ultra-short on-time of fluorescent molecules of approximately 10 milliseconds and a capture frame rate of approximately 781 frames per second (FPS). Using DEPAF, super-resolution video was rendered with a temporal resolution of approximately 76.8 milliseconds (approximately 13 FPS) and a spatial resolution of approximately 30 nanometers. These results demonstrate that DEPAF enables millisecond-level dynamic 2D SMLM, making it possible to observe rapid and minute changes in dynamic subcellular structures.

[0154] (2) Using fitted Gaussian PSF images as examples of POIs, DEPAF enables high-throughput and high-sensitivity detection of FISH point signals. This demonstrates the application of DEPAF in MERFISH data collected from samples with strong backgrounds and high-density markers. Results show that DEPAF significantly improves the throughput and sensitivity of MERFISH technology because it is highly robust to highly overlapping point signals and has a high detection rate for excessively bright or dark point signals. Furthermore, DEPAF reduces the error rate per bit, indicating its potential to enable MERFISH to use longer error-resistant codes, thus allowing for the simultaneous detection of more RNA species. Moreover, since DEPAF does not use explicit thresholding to determine signal presence, it can globally detect point signals in images with consistent sensitivity after simple bleaching correction of hybridization images from different rounds. Overall, DEPAF effectively addresses the inherent challenges of MERFISH technology in data analysis, greatly promoting the development of MERFISH towards higher throughput and sensitivity, and enabling more detailed and comprehensive RNA abundance measurements and spatial organization analysis.

[0155] (3) Using fitted spike signals and standard Gaussian PSF images as POI examples, DEPAF can integrate accurate denoising, robust spike signal estimation, and cost-effective active neuron segmentation for two-photon neuronal calcium imaging videos. Considering that the data of each pixel along the time axis in the two-photon calcium imaging video is treated as a one-dimensional image, DEPAF can denoise the video and locate overlapping pulse signals in significant noise. Simultaneously, a noise-free signal separated from the background is generated through DEPAF's image synthesis module. After reshaping these noise-free signals into a three-dimensional video, an accurately denoised two-photon calcium imaging video retaining only the signals of active neurons can be obtained. Furthermore, DEPAF can robustly handle pulse amplitude differences between different neurons and cortical regions through simple normalization based on the median and first quartile of each one-dimensional image. This advantage is also attributed to DEPAF's avoidance of using explicit thresholding to determine signal presence. Based on the denoised video, DEPAF can detect and locate high-density PSF constituting regions of active neurons in the spatial dimension. Subsequently, active neurons were extracted and segmented by performing simple DBSCAN clustering and boundary generation on the located PSFs. Comparative experimental results show that, in this task, DEPAF outperforms other methods that rely on large amounts of real data, as a method that significantly reduces the need for real data. DEPAF demonstrates excellent performance while maintaining high practicality by unifying denoising, impulse estimation, and active neuron segmentation tasks in two-photon calcium imaging through a single framework. This further demonstrates the general advantages of DEPAF across a wider range of data modalities and tasks, and also shows its potential for using lower photon budgets for two-photon calcium imaging and analysis of finer structures in more challenging imaging environments.

[0156] (4) By using constant image patches as POI examples, DEPAF can achieve accurate estimation of arbitrary heterogeneous backgrounds. For DEPAF, the key to achieving accurate estimation of arbitrary backgrounds lies in the design of the background POI examples. First, previous methods generally revealed that the background distribution is concentrated in low frequencies, indicating that the background POI examples should represent some form of low-frequency information. Second, since the background does not contain the image contribution of signals or noise, it should be as unrelated to these as possible to prevent DEPAF from fitting the background POI to signals or noise. Therefore, the background POI examples should be orthogonal to various signals and noises. Based on these considerations, the background POI examples are designed as uniform patches of a certain size, which are essentially small images with constant values. The reasons are as follows: (1) The inherent size of the uniform patch makes it unable to fit high-frequency information in the image, so it tends to be used to fit low-frequency information; (2) Due to the extensive periodic fluctuations of signals and the randomness of noise, this uniform patch can achieve extensive orthogonality to various signals and noises. These two factors meet the requirements well. Furthermore, benefiting from DEPAF's ability to simultaneously fit multiple POI examples, by using this background POI example as an additional example besides the signal POI example, DEPAF can simultaneously optimize background estimation and signal detection. This adaptive discrimination further improves the accuracy of background estimation. Experimental results show that DEPAF achieves high-accuracy background estimation in simulated data with different background frequencies and exhibits excellent performance on experimental data. This method provides a new perspective and approach for a wider range of fluorescence background estimation.

[0157] This invention proposes a general fluorescence microscopy signal detection method by treating signals as points of interest (POIs), exhibiting high versatility, performance, and practicality. In millisecond-level dynamic 2D SMLM imaging, the method demonstrates its ability to locate single-shape signals under high signal density and noise conditions. In 3D SMLM imaging, the method demonstrates its ability to simultaneously locate multi-shape signals under the same stringent imaging conditions. Its application to FISH point signal detection enables high-throughput and sensitive MERFISH analysis. To demonstrate its ability to handle cross-dimensional data and other potential applications, such as 1D signal detection, prior-free denoising, and geometric segmentation based on high-density point localization, this embodiment applies it to two-photon calcium imaging video processing, achieving accurate denoising, highly robust spike signal estimation, and cost-effective activated neuron segmentation. To demonstrate its ability to embed more generalized priors independent of signal shape, this embodiment applies it to fluorescence background estimation, achieving accurate separation of arbitrary non-uniform backgrounds.

[0158] The signal detection steps described above are achieved by simply modifying the POI examples provided to the DEPAF model and adjusting the sensitivity-related hyperparameters, without requiring any benchmark true values ​​or domain-specific imaging physics models. This is possible because the structural and regular features of fluorescence microscopy signals, such as PSF in SMLM, periodic fluctuations in spike estimation, and specific frequency distributions in fluorescence background estimation, can all be broadly represented using POI examples. By further integrating the powerful representations and utility of a self-supervised neural network architecture, the model can be easily tuned to extract signals based on variable signal features without compromising performance. These case studies repeatedly validate a novel perspective on signals as patterns and fully demonstrate the high versatility, performance, and practicality of DEPAF across a variety of tasks.

[0159] While maintaining performance, DEPAF's prediction speed for locating low-density POIs is comparable to traditional fitting-based methods, but it outperforms them when using high-density POIs. This is because the computational complexity of DEPAF's first branch depends only on the image size, not the number of POIs within it. Although the second branch involves sequential processing, it does not include any iterative processes used in traditional fitting-based methods during the prediction phase.

[0160] The fluorescence microscopy signal detection system provided by the present invention is described below. The fluorescence microscopy signal detection system described below can be referred to in correspondence with the fluorescence microscopy signal detection method described above.

[0161] See Figure 3 The present invention also provides a fluorescence microscopy signal detection system, comprising:

[0162] The image receiving module is used to receive fluorescence microscope images under test.

[0163] The signal detection module is used to: obtain the fluorescence microscopy signal detection result of the fluorescence microscopy image under test based on the fluorescence microscopy image under test, using the fluorescence microscopy signal detection model obtained by any of the above-described fluorescence microscopy signal detection model construction methods, wherein the fluorescence microscopy signal detection result includes the position information and / or amplitude information of the fluorescence microscopy signal.

[0164] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4As shown, the electronic device may include a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute any of the above-described methods for constructing a fluorescence microscopy signal detection model and / or for detecting fluorescence microscopy signals.

[0165] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0166] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program, the computer program being stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer is able to execute any of the above-described methods for constructing a fluorescence microscopy signal detection model and / or for detecting fluorescence microscopy signals.

[0167] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for constructing the fluorescence microscopy signal detection model and / or the fluorescence microscopy signal detection method described above.

[0168] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0169] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a fluorescence microscopy signal detection model, characterized in that, include: Acquiring raw fluorescence microscopy images I ori Among them, the original fluorescence microscopy image I ori The fluorescence microscopy signals in the samples have been characterized as Pattern of Interest (POI) or combinations thereof; Initialize a deep learning model Using deep learning models Original fluorescence microscope image I ori Convert to fitted graph I fit Among them, the fitted graph I fit Encoded with original fluorescence microscopy images I ori Location information and magnitude of POIs in the middle; Fitted Figure I fit Group convolution with preset POI examples yields denoised image I. dns ; Maximize the original fluorescence microscope image I ori and denoised image I dns The L2 similarity between them, and minimizing the fitted graph I fit The L1 regularization term is used to optimize the deep learning model. Initialize the decoding module Using the decoding module From the fitted graph I fit The original fluorescence microscope image I was obtained by decoding. ori Location information of POIs in the middle; From the original fluorescence microscope image I ori Medium subtraction denoising image I dns The noisy image I is obtained. ns ; Noise image I ns With Fitted Figure I fit The sum along the channel dimension By merging the samples, a noise fitting plot can be obtained. And based on the noise fitting graph The original fluorescence microscope image I was estimated. ori The magnitude of POI; Combined with the estimated original fluorescence microscopy image I ori Amplitude of POI, example of preset POI, and the decoded raw fluorescence microscope image I ori The location information of the POI is used to obtain the synthetic image I. syn ; Maximize the synthesized image I syn and denoised image I dns L1 similarity between them to optimize the decoding module. Combined with the optimized deep learning model and decoding module A fluorescence microscopy signal detection model was obtained.

2. The method for constructing the fluorescence microscopy signal detection model according to claim 1, characterized in that, The default expression for the POI example is: In the expressions of the default POI examples, Let C represent a set of C pre-obtained POI images along the channel dimension, where (x,y) represents the pixel coordinates. Let represent the minimum value of the c-th POI image. Represents the maximum value of the c-th POI image; and / or, Deep Learning Models A fully convolutional neural network with a U-net architecture, a deep learning model. loss function for: Deep Learning Models loss function middle, I ori Represents the original fluorescence microscope image, I dns I represents a denoised image. fit λ represents the fitted plot, and λ represents the regularization strength. And / or, The fitting graph I fit Group convolution with preset POI examples yields denoised image I. dns ,include: According to the first expression, the fitted graph I will be... fit Group convolution with preset POI examples yields denoised image I. dns The first expression is: In the first expression, I dns I represents a denoised image. fit This represents a fitted plot; the colon ":" indicates all x or y coordinate values ​​in that dimension. Represents two-dimensional convolution. This represents a preset POI example with mean sum normalization; and / or, Fitted Figure I fit The sum along the channel dimension The expression is:

3. The method for constructing the fluorescence microscopy signal detection model according to claim 2, characterized in that, The use of decoding module From the fitted graph I fit The original fluorescence microscope image I was obtained by decoding. ori Location information of POIs, including: Based on the fitted graph I fit The sum along the channel dimension The second expression is used to segment and obtain pixel regions that may contain POIs. A set; Based on pixel regions that may contain POIs The set of values ​​is used to calculate the centroid estimate of the i-th ROI through the third expression. Subpixel coordinates of POI As the original fluorescence microscope image I ori Location information of POIs in the middle; The second expression is: In the second expression, θ represents the fitted graph I. fit The sum along the channel dimension The pre-initialized segmentation threshold within the pixel value range; The third expression is:

4. The method for constructing the fluorescence microscopy signal detection model according to claim 3, characterized in that, The original fluorescence microscope image I ori Medium subtraction denoising image I dns The noisy image I is obtained. ns ,include: From the original fluorescence microscope image I ori Medium subtraction denoising image I dns The noisy image I is obtained. ns I ns =I ori -I dns ; Preset POI example using mean sum normalization The maximum value for noisy image I ns Perform normalization; and / or, The noise image I ns With Fitted Figure I fit The sum along the channel dimension By merging the samples, a noise fitting plot can be obtained. And based on the noise fitting graph The original fluorescence microscope image I was estimated. ori The magnitude of POIs includes: The normalized noisy image I is obtained through the fourth expression. ns With Fitted Figure I fit The sum along the channel dimension By merging the samples, a noise fitting plot can be obtained. The fifth expression is used to calculate the maximum value after convolving each Region of Interest (ROI) with the preset ROI example corresponding to its z-axis position, thus estimating the i-th ROI. The amplitude A of the included POI i ; The fourth expression is: In the fourth expression, Represents the normalized noisy image I ns The value at pixel coordinates (x, y). Indicates the fitted graph I fit The sum along the channel dimension The value at pixel coordinates (x, y). Represents the noise fitting plot The value at pixel coordinates (x, y); The fifth expression is: In the fifth expression, "⊙" represents matrix element-wise multiplication, and M(·) represents a function that outputs a mask: pixels within the input set are 1, and all others are 0; at position... The preset POI example obtained is obtained through interpolation.

5. The method for constructing the fluorescence microscopy signal detection model according to claim 4, characterized in that, The original fluorescence microscope image I obtained by combining estimation ori Amplitude of POI, example of preset POI, and the decoded raw fluorescence microscope image I ori The location information of the POI is used to obtain the synthetic image I. syn ,include: The sixth expression, combined with the estimated original fluorescence microscope image I, is used to... ori Amplitude of POI, example of preset POI, and the decoded raw fluorescence microscope image I ori The location information of the POI is used to obtain the synthetic image I. syn The sixth expression is: In the sixth expression, H and W represent the height and width of the preset POI example, respectively, in position... and The pixel values ​​of the preset POI example at that location are obtained through interpolation; and / or, Decoding module The loss function is:

6. The method for constructing a fluorescence microscopy signal detection model according to any one of claims 1 to 5, characterized in that, The fluorescence microscopy signal detection model uses any one of the following or any combination thereof to optimize model performance: center-aligned upsampling, non-uniform background learning, and POI diffusion.

7. A method for detecting fluorescence microscopic signals, characterized in that, include: Receive fluorescence microscope images to be tested; Based on the fluorescence microscope image to be tested, the fluorescence microscopy signal detection model obtained by the fluorescence microscopy signal detection model construction method according to any one of claims 1 to 6 is used to obtain the fluorescence microscopy signal detection result of the fluorescence microscope image to be tested, wherein the fluorescence microscopy signal detection result includes the position information and / or amplitude information of the fluorescence microscopy signal.

8. A fluorescence microscopy signal detection system, characterized in that, include: The image receiving module is used to receive fluorescence microscope images under test. The signal detection module is used to: obtain the fluorescence microscopy signal detection result of the fluorescence microscopy image under test based on the fluorescence microscopy image under test, by using the fluorescence microscopy signal detection model obtained by the construction method of the fluorescence microscopy signal detection model according to any one of claims 1 to 6, wherein the fluorescence microscopy signal detection result includes the position information and / or amplitude information of the fluorescence microscopy signal.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method for constructing a fluorescence microscopy signal detection model as described in any one of claims 1 to 6 and / or the fluorescence microscopy signal detection method as described in claim 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for constructing the fluorescence microscopy signal detection model as described in any one of claims 1 to 6 and / or the fluorescence microscopy signal detection method as described in claim 7.

Images