Self-supervised fluorescence microscopy image denoising method, apparatus and its readable storage medium

Abstract

This invention proposes a self-supervised fluorescence microscopy image denoising method, apparatus, and readable storage medium, belonging to the field of digital image processing. To address the problems of detail loss due to fixed masks in existing self-supervised methods and signal collapse due to neglecting heteroscedasticity in traditional loss methods, this method extracts the local variance of the image to be processed to generate a structure-adaptive probability mask map, and performs pixel replacement to obtain the mask image. The mask image is input into a network embedded in a visual state space module, and global contextual features are extracted through multi-directional selective scanning to output a predicted image. A spatial weight map is generated based on local signal intensity, and an intensity-aware weighted loss is calculated at the mask location to optimize the network. This invention is mainly used for fluorescence microscopy image processing, effectively balancing background denoising and high-frequency structure preservation.

Description

Technical Field

[0001] This invention relates to the fields of digital image processing and deep learning technology, and in particular to a self-supervised fluorescence microscopy image denoising method, apparatus and readable storage medium thereof. Background Technology

[0002] Fluorescence microscopy has wide applications in the biomedical field, but due to photon scarcity during imaging, it is prone to severe shot noise. Since it is extremely difficult to obtain perfectly aligned "clean-noisy" image pairs in actual imaging, the application of traditional supervised learning-based denoising algorithms is severely limited. In recent years, self-supervised denoising algorithms, represented by Noise2Void (N2V), have solved the dependence on noise-free reference images by using a "blind spot network" mechanism (i.e., randomly masking some pixels and using local neighboring pixels for prediction).

[0003] Existing N2V-type algorithms typically sample the entire image using a uniform fixed probability mask (e.g., 1.5%) and employ traditional loss functions, such as mean squared error (MSE) or mean absolute error (MAE), which assume a globally constant noise variance, to calculate the reconstruction error between the network's predictions and the noisy image, thereby driving network training. Summary of the Invention

[0004] This invention provides a self-supervised fluorescence microscopy image denoising method, apparatus, and readable storage medium. It addresses the problems of existing self-supervised denoising algorithms that use fixed probability masks, resulting in incomplete denoising of smooth backgrounds and the erasure of complex structural details. At the same time, traditional loss functions ignore the heteroscedasticity of photon shot noise variance as the signal intensity increases, leading to severe signal collapse and blurring in high-brightness areas.

[0005] The core technology of this invention is to dynamically allocate mask probabilities based on local variance to adaptively protect biological structures, and to embed a visual state space module in the deepest layer to extract global contextual features. At the same time, it uses the local mean of the image to construct an intensity-aware weighted loss function to correct physical deviations caused by heteroscedastic noise.

[0006] In a first aspect, the present invention provides a self-supervised fluorescence microscopy image denoising method, the method comprising the following steps:

[0007] Acquire the fluorescence microscopic image to be denoised;

[0008] The local variance features of the fluorescence microscopy image to be denoised are obtained by local smooth pooling and then global normalized to obtain a normalized variance map. The values ​​of each pixel position in the normalized variance map are mapped to a preset probability allocation function to calculate the mask probability that is negatively correlated with the normalized variance value. The mask probabilities of all pixel positions constitute a structure-adaptive probability mask map. The target pixel to be masked is randomly determined according to the mask probability, and the pixel value of a non-nearby pixel in the local neighborhood of the target pixel is randomly selected as a replacement value to obtain the mask image.

[0009] The masked image is input into a pre-constructed global information perception network (GIFN). The GIFN adopts a symmetric encoder-decoder architecture with skip connections. The encoder contains multiple downsampling layers, and the decoder contains corresponding multiple upsampling layers. A visual state space module is embedded in the deep bottleneck structure of the GIFN, connecting the lowest-level downsampling features output by the encoder with the upsampling features input by the decoder. The visual state space module performs multi-directional selective scanning of the feature map of the masked image, including: flattening the input two-dimensional feature map along multiple different spatial diagonal directions to transform it into a one-dimensional feature sequence for multi-directional selective scanning, in order to extract global context features and output a predicted image. The multiple different spatial diagonal directions include at least the top-left to bottom-right direction, the bottom-right to top-left direction, the top-right to bottom-left direction, and the bottom-left to top-right direction.

[0010] After performing local mean filtering on the fluorescence microscopy image to be denoised without random pixel replacement processing, a spatial weight map is generated by combining a preset intensity adjustment factor and a basic weight constant, in which the weight values ​​are positively correlated with the local photon flux. For each mask position where a replacement occurs, the squared error between the predicted image and the fluorescence microscopy image to be denoised is calculated. The squared error is multiplied by the weight value at the corresponding position in the spatial weight map, and the average of the weighted product results of all mask positions is calculated to obtain the intensity-perceived weighted loss.

[0011] The parameters of the global information perception network are optimized based on the intensity-aware weighted loss to obtain the trained denoising model.

[0012] Furthermore, local variance features of the fluorescence micrograph to be denoised are obtained through local smoothing pooling operations, including:

[0013] The original fluorescence microscopy image to be denoised is globally min-max normalized and mapped to the [0,1] interval to obtain a normalized image patch X; a size of X is used. The sliding window is used to calculate the local variance at each pixel position (i,j). The calculation formula is:

[0014]

[0015] in, This indicates the mean pooling operation. This indicates the element-wise squaring operation.

[0016] Furthermore, the preset probability allocation function is specifically as follows:

[0017]

[0018] in, Let be the mask probability at pixel position (i,j). The normalized variance value. Scaling factor Based on the bias.

[0019] Furthermore, the target pixel to be masked is randomly determined based on the masking probability, including:

[0020] Generate a random number that follows a uniform distribution in the interval [0,1] for each pixel location in the image. If the corresponding position If the pixel position is determined to be the target pixel to be masked, it is added to the mask position set. Among them This represents the mask probability at the corresponding position.

[0021] Furthermore, during multi-directional selective scanning, the visual state space module performs dynamic perceptual adjustments on the one-dimensional feature sequence, including:

[0022] During the scanning process, the perceptual adoption weight of the state transition is dynamically adjusted based on the feature content of the current pixel.

[0023] In smooth background regions of the image, the adoption weight of current noisy pixels is reduced to extract background consistency features over a wide range;

[0024] In the structural regions of the image signal, the capture of current detailed features is enhanced to maintain the coherence of biological structures.

[0025] Furthermore, the visual state space module extracts global contextual features and outputs a predicted image, including:

[0026] An information channel spanning spatial distance is established using an internal memory mechanism to integrate enhanced features from multiple different spatial diagonal scanning outputs. The integrated features are then remapped back to two-dimensional space to generate global context features. After fusing the global context features with local features, the output predicted image is reconstructed via a decoder.

[0027] Furthermore, the specific calculation process for the intensity-perception weighted loss is as follows:

[0028] Using the formula: Generate a spatial weighted graph W;

[0029] in, is the intensity adjustment constant, and Y is the original fluorescence micrograph to be denoised without any mask replacement processing. This indicates the process of denoising the original fluorescence micrograph. Mean filtering processing;

[0030] For the set of mask locations where replacement occurs Intensity-perception weighted loss The calculation formula is:

[0031]

[0032] It uses relational expressions: The total number of mask pixels. These represent the weight values ​​at corresponding positions in the spatial weight map. To predict the pixel value at the corresponding location in the image, The pixel value at the corresponding position in the original fluorescence microscopy image to be denoised.

[0033] In a second aspect, the present invention provides a self-supervised fluorescence microscopy image denoising device, comprising:

[0034] Image acquisition module, used to acquire fluorescence microscopy images to be denoised;

[0035] The mask generation and replacement module is used to obtain the local variance features of the fluorescence microscopy image to be denoised through local smooth pooling operations and perform global normalization to obtain a normalized variance map. The values ​​of each pixel position in the normalized variance map are mapped to a preset probability allocation function to calculate the mask probability that is negatively correlated with the normalized variance value. The mask probabilities of all pixel positions constitute a structure-adaptive probability mask map. The target pixel to be masked is randomly determined according to the mask probability, and the pixel value of the non-near-neighbor pixel in the local neighborhood of the target pixel is randomly selected as the replacement value to obtain the mask image.

[0036] The denoising prediction module is used to input the mask image into a pre-constructed global information perception network. The global information perception network adopts a symmetric encoder-decoder architecture with skip connections. The encoder contains multiple downsampling layers, and the decoder contains corresponding multiple upsampling layers. The visual state space module is embedded in the deep bottleneck structure of the global information perception network to connect the lowest-level downsampling features output by the encoder with the upsampling features input by the decoder. The visual state space module performs multi-directional selective scanning on the feature map of the mask image, including: flattening the input two-dimensional feature map along multiple different spatial diagonal directions to transform it into a one-dimensional feature sequence for multi-directional selective scanning to extract global context features and output a predicted image. The multiple different spatial diagonal directions include at least the top-left to bottom-right direction, the bottom-right to top-left direction, the top-right to bottom-left direction, and the bottom-left to top-right direction.

[0037] The error weighting module is used to perform local mean filtering on the fluorescence microscopy image to be denoised without random pixel replacement processing, and then generate a spatial weight map with weight values ​​positively correlated with local photon flux by combining a preset intensity adjustment factor and a basic weight constant. For each mask position where a replacement occurs, the squared error between the predicted image and the fluorescence microscopy image to be denoised is calculated. The squared error is multiplied by the weight value of the corresponding position in the spatial weight map, and the average of the weighted product results of all mask positions is calculated to obtain the intensity-perceived weighted loss.

[0038] The model optimization module is used to optimize the parameters of the global information perception network based on the intensity-aware weighted loss to obtain the trained denoising model.

[0039] Thirdly, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the above-described self-supervised fluorescence microscopy image denoising method.

[0040] Fourthly, the present invention provides a readable storage medium storing a computer program, the computer program including program code for controlling a process to execute the process, the process including the self-supervised fluorescence microscopy image denoising method described above.

[0041] The main contributions and innovations of this invention are as follows:

[0042] 1. This invention proposes a structure-adaptive probabilistic masking mechanism. By extracting local variance and performing nonlinear probability mapping, it achieves variable-probability sampling based on local information density, breaking the deadlock of traditional algorithms that cannot simultaneously achieve "background denoising depth" and "structure preservation accuracy." High-probability sampling is automatically used in smooth background areas to effectively suppress residual background noise, while extremely low-probability sampling is automatically used in complex texture areas to protect high-frequency biological structures from damage.

[0043] 2. This invention constructs a global information perception network with a visual state space module. By utilizing a multi-directional two-dimensional selective scanning mechanism to establish spatial information channels across long distances, it overcomes the limitations of the local receptive field in traditional convolutional neural networks. Even under severe noise occlusion, it can fill in missing pixels by referencing globally similar structural information, ensuring the continuity and unbroken topological features of large-scale biological structures (such as neurons and mitochondria).

[0044] 3. This invention introduces an intensity-sensing weighted loss function based on signal strength. By utilizing a dynamic weight map generated from the local mean of the image to be processed, the network is forced to increase the error penalty weight for high-brightness regions, correcting the network learning direction shift caused by photon shot noise. This effectively solves the inherent signal collapse problem in high-brightness regions in self-supervised algorithms, and significantly improves the fidelity and accuracy of the denoised image in linear brightness measurement.

[0045] Details of one or more embodiments of the present invention are set forth in the following drawings and description, so that other features, objects and advantages of the invention will be more readily understood. Attached Figure Description

[0046] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0047] Figure 1 This is a flowchart of the model training process for a self-supervised fluorescence microscopy image denoising method according to an embodiment of the present invention;

[0048] Figure 2 This is a flowchart of a self-supervised fluorescence microscopy image denoising method according to an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0050] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0051] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0052] Example 1

[0053] This embodiment provides a self-supervised fluorescence microscopy image denoising method based on structure and strength priors and a state-space model. (Refer to the appendix of the specification.) Figure 1 The model training process architecture diagram shown indicates that the denoising method of this invention is mainly divided into a training phase and a test inference phase.

[0054] During the training phase, such as Figure 2 As shown, the specific steps include:

[0055] Step S1: Obtain the original fluorescence microscopic image to be denoised (i.e., Figure 1 The original image in the image is used to extract local variance features.

[0056] Specifically, in order to quantify the "information density" of each region of the image, for each normalized image block X in the input, a sliding window of size k×k is used (in this embodiment, k=5, i.e., a window size of 5×5), and the local variance σ of each pixel position (i,j) is calculated. 2 (i,j), the calculation formula is:

[0057]

[0058] in, This indicates a mean pooling (local smooth pooling) operation. The larger the variance value, the closer the region is to the edge of biological tissue or complex texture; the smaller the variance, the closer the region is to a smooth background.

[0059] It should be noted that the normalized image patch X is obtained by performing global Min-Max normalization on the original fluorescence microscopy image to be denoised, that is, linearly mapping its pixel values ​​to the [0,1] interval. Furthermore, in the above formula... This indicates that each pixel value in the normalized image patch X is squared element-wise, rather than matrix multiplication.

[0060] Step S2: Generate a structural adaptive probability mask (i.e., Figure 1The variance map in the image is used to perform blind spot replacement processing on the fluorescence micrograph to be denoised, thus obtaining the mask image (i.e., the variance map in the image). Figure 1 (Input image in the image).

[0061] Specifically, firstly, the local variance σ of the entire image... 2 Perform a global Min-Max normalization operation on (i,j) to force it to be mapped to the [0,1] interval, resulting in a normalized variance plot. .

[0062] Subsequently, the probability assignment function is applied to calculate the mask probability for each point. The formula is:

[0063]

[0064] in, The scaling factor is (preferably configured as 0.15 in this embodiment). The base bias is set to 0.005 in this embodiment. As shown in the formula, the mask probability is negatively correlated with the normalized variance. In an extremely smooth background region ( The mask probability approaches 15.5%, and high-intensity sampling is used to suppress residual background noise; while in structurally complex texture areas ( The mask probability approaches 0.5%, and the high-frequency structure is protected from damage by extremely low-intensity sampling.

[0065] After constructing an adaptive probabilistic mask based on the mask probabilities of all pixel positions, target pixels to be masked are randomly selected according to these probabilities. For each selected target pixel, instead of directly setting it to zero, a non-nearby pixel value is randomly selected within its 5×5 local neighborhood as a replacement value. This operation disrupts local autocorrelation between pixels while maintaining the consistency of image data distribution.

[0066] In this embodiment, at the specific algorithm code execution level, the step of randomly determining the target pixel to be masked based on the mask probability is as follows:

[0067] Generate a random number that follows a uniform distribution in the interval [0,1] for each pixel location in the image. If the corresponding position If the pixel position is determined to be the target pixel to be masked, it is added to the mask position set. Otherwise, retain the original pixel value.

[0068] Step S3: Input the mask image (input image) into a pre-built global information perception network (Mamba-Unet) containing a visual state space (SS2D) module for denoising prediction.

[0069] This network adopts a symmetric encoder-decoder (U-Net) architecture, with 4 downsampling layers and 4 upsampling layers, and the channel number gradients are set to [32, 64, 128, 256] respectively. Unlike traditional fully convolutional networks, this invention embeds a visual state space module (SS2D) in the deepest layer (bottleneck layer) of the U-Net.

[0070] This module performs two-dimensional selective scanning: First, the input two-dimensional feature map is flattened into a one-dimensional sequence along four different spatial diagonal directions (top left to bottom right, bottom right to top left, top right to bottom left, bottom left to top right). While scanning these one-dimensional sequences, the module dynamically adjusts the adoption weight of the perception state based on the feature content of the current pixel. In smooth background areas, the module automatically reduces the adoption weight of the current noisy pixels, focusing on extracting large-scale background consistency features; in signal structure areas, the module strengthens the capture of current details. With the help of an internal memory mechanism, this module can establish spatial information channels across long distances (e.g., using similar cellular structures in the distance of the image to complete pixels currently obscured by severe noise). Finally, the global enhancement features obtained from the four-directional scan are integrated and remapped back to two-dimensional space, fused with the original local features through skip connections, and reconstructed by the decoder to output the predicted image (i.e., Figure 1 (output image in the image).

[0071] The Visual State Space Module (SS2D) is based on the time-scale discretization theory of continuous systems. When scanning the aforementioned one-dimensional sequence, it internally introduces a zero-order hold (ZOH) or a linear assumption to discretize the continuous state transition matrix and observation matrix into dynamic weight parameters that depend on the input. This allows for dynamic adjustment and long-range memory of the weights adopted for the current pixel's perception state without increasing computational complexity (maintaining linear time complexity).

[0072] Step S4: Calculate the intensity-perceived weighted loss (IA-Loss).

[0073] Considering that fluorescence imaging is dominated by photon shot noise, and the noise variance increases with signal intensity (heteroscedasticity), this embodiment uses the original fluorescence micrograph Y to be denoised (i.e., without any mask replacement processing) as an example. Figure 1 The original image (in the image) is first subjected to a 3×3 mean filter to prevent interference from isolated noise points, and then a spatial weight map W (i.e., Figure 1 (brightness diagram in the image)

[0074]

[0075] in, is the intensity adjustment constant (preferably 1.0). The values ​​in the weighting map W are positively correlated with the local photon flux (signal intensity).

[0076] Based on the generated spatial weight map, only at the set of locations where mask replacement occurs... At this point, calculate the predicted image. Weighted reconstruction error between the original image Y and the original image Y:

[0077]

[0078] in, The total number of mask pixels. The input is the mask image. This formula forcibly increases the error penalty weight for high-brightness regions, compelling the network to maintain the true signal strength in these regions, effectively solving the signal collapse problem in high-brightness areas in the traditional N2V algorithm.

[0079] Step S5: Optimize network parameters.

[0080] The training batch size is set to 32, the Adam optimizer is used, and the learning rate is set to 1e-4. Based on the intensity-aware weighted loss calculated above, backpropagation and parameter optimization are performed on the global information perception network until the network converges, and the trained denoising model is obtained.

[0081] During the testing and inference phase, a full-resolution noisy fluorescence micrograph of the image to be processed is directly acquired, skipping all the aforementioned structural adaptive probability masking and blind spot replacement steps, and directly input into the trained denoising model. The network performs global denoising directly through the learned contextual collaboration mechanism, outputting the final high-fidelity denoised image.

[0082] As a preferred hardware and software implementation environment, the network model training and testing of the present invention can be performed on a computer or server device equipped with an NVIDIA graphics processing unit (GPU) and its corresponding CUDA accelerated computing platform, based on deep learning frameworks such as PyTorch or TensorFlow.

[0083] Example 2

[0084] Based on the same concept, this invention also proposes a self-supervised fluorescence microscopy image denoising device, comprising:

[0085] Image acquisition module: used to acquire fluorescence microscopic images to be denoised;

[0086] Mask generation and replacement module: Based on the local variance features of the fluorescence micrograph to be denoised, generate a structure-adaptive probability mask map, and perform random replacement processing on local non-neighbor pixels of the original image according to the map, and output the mask image.

[0087] Denoising prediction module: Includes a pre-built Mamba-Unet global information perception network, which uses a visual state space module embedded in a deep bottleneck structure to perform four-way spatial diagonal flattening and dynamic selective scanning of feature maps, extract global context features and output prediction images;

[0088] Error weighting module: It is used to generate a spatial weight map based on the mean smoothing of the original image to be denoised. It calculates the squared error between the predicted image and the original image only at the mask position where the replacement occurs, and uses the weight map to perform product weighting to calculate the mean, so as to obtain the intensity-aware weighted loss.

[0089] Model optimization module: Used to optimize the global information perception network using intensity-aware weighted loss to obtain the trained denoising model.

[0090] Example 3

[0091] This embodiment also provides an electronic device, see reference. Figure 3 It includes a memory 404 and a processor 402, wherein the memory 404 stores a computer program and the processor 402 is configured to run the computer program to perform the steps in any of the above method embodiments.

[0092] Specifically, the processor 402 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement embodiments of the present invention.

[0093] Memory 404 may include a mass storage device for data or instructions. For example, and not limitingly, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to a data processing device. In a particular embodiment, memory 404 is non-volatile memory. In a particular embodiment, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random-Access Memory (FPMDRAM), Extended Data Out Dynamic Random-Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0094] The memory 404 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 402.

[0095] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any of the self-supervised fluorescence microscopy image denoising methods in the above embodiments.

[0096] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408, wherein the transmission device 406 is connected to the processor 402, and the input / output device 408 is connected to the processor 402.

[0097] The transmission device 406 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 406 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0098] Input / output device 408 is used to input or output information.

[0099] Example 4

[0100] This embodiment also provides a readable storage medium storing a computer program, the computer program including program code for controlling a process to execute the process, the process including the self-supervised fluorescence microscopy image denoising method according to Embodiment 1.

[0101] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0102] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0103] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets, and / or macros can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. The computer program product may include one or more computer-executable components configured to perform the embodiments when the program is run. The one or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted in this respect that, as Figure 2 Any box in the logical flow can represent a program step, or interconnected logic circuits, boxes and functions, or a combination of program steps and logic circuits, boxes and functions. Software can be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0104] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] The above embodiments are merely illustrative of several implementations of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A self-supervised fluorescence microscopy image denoising method, characterized in that, Includes the following steps: Acquire the fluorescence microscopic image to be denoised; The local variance features of the fluorescence micrograph to be denoised are obtained by local smoothing pooling and then global normalized to obtain a normalized variance map. The values ​​of each pixel position in the normalized variance map are mapped to a preset probability allocation function to calculate the mask probability that is negatively correlated with the normalized variance value. An adaptive probability mask map is constructed from the mask probabilities of all pixel positions. The target pixel to be masked is randomly determined according to the mask probability, and the pixel value of a non-nearby pixel in the local neighborhood of the target pixel is randomly selected as a replacement value to obtain the mask image. The masked image is input into a pre-constructed global information perception network. This network employs a symmetric encoder-decoder architecture with skip connections. The encoder contains multiple downsampling layers, and the decoder contains corresponding multiple upsampling layers. A visual state space module is embedded in the deep bottleneck structure of the global information perception network, connecting the lowest-level downsampling features output by the encoder with the upsampling features input by the decoder. The visual state space module performs multi-directional selective scanning of the feature map of the masked image, including: flattening the input two-dimensional feature map along multiple different spatial diagonal directions to transform it into a one-dimensional feature sequence for multi-directional selective scanning, in order to extract global context features and output a predicted image. The multiple different spatial diagonal directions include at least the top-left to bottom-right direction, the bottom-right to top-left direction, the top-right to bottom-left direction, and the bottom-left to top-right direction. After performing local mean filtering on the fluorescence micrograph to be denoised without random pixel replacement processing, a spatial weight map is generated by combining a preset intensity adjustment factor and a basic weight constant, in which the weight values ​​are positively correlated with the local photon flux. For each mask position where a replacement occurs, the squared error between the predicted image and the fluorescence micrograph to be denoised is calculated. The squared error is multiplied by the weight value at the corresponding position in the spatial weight map, and the average of the weighted product results of all mask positions is calculated to obtain the intensity-perceived weighted loss. The parameters of the global information perception network are optimized based on the intensity-aware weighted loss to obtain the trained denoising model.

2. The self-supervised fluorescence microscopy image denoising method as described in claim 1, characterized in that, The local variance features of the fluorescence micrograph to be denoised are obtained through local smoothing pooling operations, including: The original fluorescence microscopy image to be denoised is globally min-max normalized and mapped to the [0,1] interval to obtain a normalized image patch X; a size of X is used. The sliding window is used to calculate the local variance at each pixel position (i,j). The calculation formula is: in, This indicates the mean pooling operation. This indicates the element-wise squaring operation.

3. The self-supervised fluorescence microscopy image denoising method as described in claim 2, characterized in that, The preset probability allocation function is as follows: in, Let be the mask probability at pixel position (i,j). The normalized variance value. Scaling factor Based on the bias.

4. The self-supervised fluorescence microscopy image denoising method as described in claim 1, characterized in that, The target pixel to be masked is randomly determined based on the mask probability, including: Generate a random number that follows a uniform distribution in the interval [0,1] for each pixel location in the image. If the corresponding position If the pixel position is determined to be the target pixel to be masked, it is added to the mask position set. Among them This represents the mask probability at the corresponding position.

5. The self-supervised fluorescence microscopy image denoising method as described in claim 4, characterized in that, When performing multi-directional selective scanning, the visual state space module performs dynamic perceptual adjustments on the one-dimensional feature sequence, including: During the scanning process, the perceptual adoption weight of the state transition is dynamically adjusted based on the feature content of the current pixel. In smooth background regions of the image, the adoption weight of current noisy pixels is reduced to extract background consistency features over a wide range; In the structural regions of the image signal, the capture of current detailed features is enhanced to maintain the coherence of biological structures.

6. The self-supervised fluorescence microscopy image denoising method as described in claim 5, characterized in that, The visual state space module extracts global contextual features and outputs a predicted image, including: An information channel spanning spatial distance is established using an internal memory mechanism to integrate enhanced features from multiple different spatial diagonal scanning outputs; the integrated features are remapped back to two-dimensional space to generate global context features, and the global context features are fused with local features before being reconstructed and output as the predicted image via the decoder.

7. The self-supervised fluorescence microscopy image denoising method as described in claim 1, characterized in that, The specific calculation process for the intensity-sensing weighted loss is as follows: Using the formula: Generate a spatial weighted graph W; in, is the intensity adjustment constant, and Y is the original fluorescence micrograph to be denoised without any mask replacement processing. This indicates the process of denoising the original fluorescence micrograph. Mean filtering processing; For the set of mask locations where replacement occurs Intensity-perceived weighted loss The calculation formula is: It uses relational expressions: The total number of mask pixels. These represent the weight values ​​at corresponding positions in the spatial weight map. To predict the pixel value at the corresponding location in the image, The pixel value at the corresponding position in the original fluorescence microscopy image to be denoised.

8. A self-supervised fluorescence microscopy image denoising device, characterized in that, include: Image acquisition module, used to acquire fluorescence microscopy images to be denoised; The mask generation and replacement module is used to obtain the local variance features of the fluorescence micrograph to be denoised through local smooth pooling operations and perform global normalization to obtain a normalized variance map; map the values ​​of each pixel position in the normalized variance map to a preset probability allocation function, calculate the mask probability that is negatively correlated with the normalized variance value, and construct a structure-adaptive probability mask map from the mask probabilities of all pixel positions; randomly determine the target pixel to be masked according to the mask probability, and randomly select the pixel value of a non-nearby pixel in the local neighborhood of the target pixel as a replacement value to obtain the mask image. A denoising prediction module is used to input the mask image into a pre-constructed global information perception network. The global information perception network adopts a symmetric encoder-decoder architecture with skip connections. The encoder contains multiple downsampling layers, and the decoder contains corresponding multiple upsampling layers. A visual state space module is embedded in the deep bottleneck structure of the global information perception network, used to connect the lowest-level downsampling features output by the encoder with the upsampling features input by the decoder. The visual state space module performs multi-directional selective scanning of the feature map of the mask image, including: flattening the input two-dimensional feature map along multiple different spatial diagonal directions to transform it into a one-dimensional feature sequence for multi-directional selective scanning, in order to extract global context features and output a predicted image; the multiple different spatial diagonal directions include at least the top-left to bottom-right direction, the bottom-right to top-left direction, the top-right to bottom-left direction, and the bottom-left to top-right direction. The error weighting module is used to perform local mean filtering on the fluorescence microscopy image to be denoised without random pixel replacement processing, and then generate a spatial weight map with weight values ​​positively correlated with local photon flux by combining a preset intensity adjustment factor and a basic weight constant; for each mask position where a replacement occurs, the squared error between the predicted image and the fluorescence microscopy image to be denoised is calculated, the squared error is multiplied by the weight value of the corresponding position in the spatial weight map, and the average of the weighted product results of all mask positions is calculated to obtain the intensity-perceived weighted loss; The model optimization module is used to optimize the parameters of the global information perception network based on the intensity-aware weighted loss to obtain the trained denoising model.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the self-supervised fluorescence microscopy image denoising method according to any one of claims 1 to 7.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, the computer program including program code for controlling a process to execute the process, the process including the self-supervised fluorescence microscopy image denoising method according to any one of claims 1 to 7.

Images