A method and device for removing streak artifacts from light sheet fluorescence microscopy images

Abstract

This paper relates to the field of image processing, and more particularly to a method and apparatus for eliminating stripe artifacts in light-sheet fluorescence microscopy images. The method includes: acquiring multiple striped images obtained by scanning a biological tissue sample at the same location from multiple angles using a light-sheet fluorescence microscope; constructing a stripe-free image based on the multiple striped images; using the multiple striped images and their corresponding stripe-free images as training datasets; training a deep adversarial network (DAN) based on the training datasets; and using the trained DAN to eliminate stripe artifacts in the striped images of the target biological tissue sample. The embodiments described in this paper eliminate stripe artifacts in striped images of biological tissue samples obtained by light-sheet fluorescence microscopy by training a DAN model, improving the clarity of the stripe-free image obtained after stripe artifact removal. Furthermore, this method is applicable to various light-sheet fluorescence microscopy systems, improving the adaptability of the stripe artifact removal method.

Description

Technical Field

[0001] This article relates to the field of image processing, and in particular to a method and apparatus for eliminating stripe artifacts in light-sheet fluorescence microscopy images. Background Technology

[0002] Light-sheet fluorescence microscopy (LSFM) is a popular imaging method in developmental biology, neuroscience, and pathology research, widely used for rapid and high-resolution imaging of various biomedical sample sections. However, during optical microscopy imaging, streak artifacts can occur in the image due to inhomogeneities, high absorption, or scattering objects in the excitation light path or sample.

[0003] Currently, methods for dealing with stripe artifacts in light-sheet fluorescence microscopy images mainly include optical methods, such as modifying the microscope's physical equipment by adding an elliptical holographic diffuser lens, or taking multiple images from different angles and combining the images to eliminate the stripe artifacts. Frequency domain filtering methods also exist, which use operations such as Fourier transform to convert the image to the frequency domain, find corresponding information in the frequency domain based on the characteristics of the stripes, and filter it out to eliminate the stripe artifacts.

[0004] Because different microscopes vary greatly, methods for modifying the physical equipment of a microscope are not universally applicable, and changes in microscope imaging parameters may also render these methods ineffective. Imaging multiple times from different angles is time-consuming and requires modifications to the microscope equipment each time. Frequency domain filtering has the drawback that the positions and sizes of the fringes in different images vary, corresponding to different frequency domain information, thus requiring different filtering methods and making it unsuitable for large-scale image data.

[0005] There is an urgent need for a method to eliminate stripe artifacts in light-sheet fluorescence microscopy images, in order to solve the problem that existing methods for eliminating stripe artifacts in light-sheet fluorescence microscopy images have low universality. Summary of the Invention

[0006] To address the problems in the prior art, this embodiment provides a method and apparatus for eliminating stripe artifacts in light-sheet fluorescence microscope images. This method eliminates stripe artifacts in images with stripes, can be applied to most light-sheet fluorescence microscopes, does not require changes to the physical equipment, and is suitable for large-scale image data.

[0007] To solve the above-mentioned technical problems, the specific technical solution presented in this paper is as follows:

[0008] On the one hand, the embodiments of this paper provide a method for eliminating stripe artifacts in light-sheet fluorescence microscopy images, including,

[0009] Multiple striped images of a biological sample tissue are obtained by scanning the same location at multiple angles using a light-sheet fluorescence microscope, and a stripe-free image of the biological sample tissue at that location is constructed based on the multiple striped images.

[0010] Multiple striped images and corresponding non-striped images are used as training datasets to train a deep adversarial network based on the multiple training datasets of the biological sample tissue.

[0011] The trained deep adversarial network is used to eliminate stripe artifacts in the striped image of the target biological sample tissue, which is obtained by scanning the target biological sample tissue using the light-sheet fluorescence microscope.

[0012] Furthermore, the deep adversarial network includes a generator network and a discriminator network. The generator network is used to calculate the non-striped image corresponding to the striped image, and the discriminator network is used to determine whether the non-striped image calculated by the generator network is true.

[0013] Furthermore, the left half of the generator network is an encoder, and the right half is a decoder;

[0014] The steps of training the generative network include training the encoder using the training dataset, and then training the decoder using the computation results of the encoder to obtain the training result of the generative network.

[0015] Furthermore, the encoder includes an encoder initialization module and n downsampling encoding modules consisting of convolutional layers and convolutional attention modules, where n is a positive integer greater than 1;

[0016] The steps for training the encoder are as follows: the training dataset is initialized using the encoder initialization module, and multiple downsampling encoding modules are trained using the initialized training dataset, wherein the output of the previous downsampling encoding module is the input of the next downsampling encoding module, and the initialized training dataset is used as the input of the first downsampling encoding module.

[0017] During the training process of the encoder initialization module, the input of the downsampling encoding module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels and number of output channels. Then, the calculation result of the convolution attention module is activated using the LeakyRelu activation function to provide the network with non-linear modeling capability. The calculation result of the LeakyRelu activation function is used as the output of the encoder initialization module.

[0018] During the training process of the first downsampling encoding module to the (n-1)th downsampling encoding module, the input of the downsampling encoding module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the result of the convolution calculation is normalized using the InstanceNorm2d normalization function. Next, the Convolutional Attention Module (CBAM) is used to perform feature selection and enhancement on the normalized result. Finally, the calculation result of the Convolutional Attention Module is activated using the LeakyRelu activation function to provide the network with non-linear modeling capabilities. The calculation result of the LeakyRelu activation function is used as the output of the downsampling encoding module.

[0019] During the training of the last downsampling encoding module, the input of the downsampling encoding module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the calculation result of the convolutional attention module is activated using the ReLU activation function to provide the network with non-linear modeling capabilities. The calculation result of the ReLU activation function is used as the calculation result of the encoder.

[0020] Furthermore, the decoder includes a decoder initialization module and n upsampling decoding modules consisting of upsampling convolutional kernels, skip-layer connections, and convolutional attention modules. The number of upsampling decoding modules included in the decoder is the same as the number of downsampling layers in the encoder and corresponds to each other in reverse order, where n is a positive integer greater than 1.

[0021] The steps for training the decoder are as follows: the encoder's calculation result is initialized using the decoder initialization module, and multiple upsampling decoding modules are trained using the initial encoder's calculation result. The output of the previous upsampling decoding module is the input of the next upsampling decoding module, and the initial encoder's calculation result is used as the input of the first upsampling decoding module.

[0022] During the training process of the decoder initialization module, the encoder's calculation results are first convolved using the corresponding convolution function ConvTranspose2d. Then, the convolution calculation results are normalized using the normalization function InstanceNorm2d. Next, the normalized results are feature-selected and enhanced using the convolutional attention module CBAM. Finally, the calculation results of the convolutional attention module are activated using the activation function ReLU to provide the network with non-linear modeling capabilities. The calculation result of the activation function ReLU is used as the output of the first upsampling decoding module.

[0023] During the training process of the first upsampling decoding module to the (n-1)th upsampling decoding module, skip-layer connections are first used to link the input of the upsampling decoding module with the output of the corresponding downsampling encoding module to obtain the linking result. Then, the linking result is convolved using the convolution function ConvTranspose2d with set kernel size, stride, padding, number of input channels, and number of output channels. The result of the convolution calculation is then normalized using the normalization function InstanceNorm2d. The normalized result is then enhanced using the convolutional attention module CBAM. Finally, the result of the convolutional attention module is activated using the activation function ReLU to provide the non-linear modeling capability of the network. The result of the activation function ReLU is used as the output of the upsampling decoding module.

[0024] During the training of the last upsampling decoding module, skip-layer connections are first used to link the input of the upsampling decoding module with the output of the corresponding downsampling encoding module to obtain the linking result. Then, the linking result is convolved using the ConvTranspose2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Finally, the convolution result is activated using the Tanh activation function to provide the network with non-linear modeling capabilities. The calculation result of the Tanh activation function is used as the stripe-free image obtained by the generator network from the striped images corresponding to the training dataset.

[0025] Furthermore, the discriminant network includes a discriminant network initialization module and m convolutional modules, where m is a positive integer greater than 1;

[0026] The calculation process of the discriminant network is as follows: the discriminant network initialization module initializes the stripeless image calculated by the generator network, and the initialized stripeless image is used to train multiple convolutional modules. The output of the previous convolutional module is the input of the next convolutional module, the initialized stripeless image is the input of the first convolutional module, and the output of the last convolutional module is the discrimination result of the discriminant network.

[0027] During the calculation process of the discriminant network initialization module, the stripeless image obtained by the generator network is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the result of the convolution calculation is activated using the LeakyRelu activation function to provide the network with nonlinear modeling capabilities. The calculation result of the LeakyRelu activation function is used as the output of the discriminant network initialization module, and the output of the discriminant network initialization module is input into the first convolution module.

[0028] During the computation of the first to the (m-1)th convolutional module, the input of the convolutional module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the result of the convolution calculation is normalized using the InstanceNorm2d normalization function. Finally, the normalized result is activated using the LeakyRelu activation function to provide the network with nonlinear modeling capabilities. The calculation result of the LeakyRelu activation function is used as the output of the convolutional module.

[0029] In the calculation process of the last convolution module, the input of the convolution module is first convolved according to the Conv2d convolution function with set kernel size, stride, padding, number of input channels and number of output channels. Finally, the sigmoid activation function is used to activate the result of the convolution calculation, providing the nonlinear modeling capability of the network. The calculation result of the sigmoid activation function is used as the discrimination result of the discriminant network.

[0030] Furthermore, the loss function during the training of the deep adversarial network is,

[0031] ,

[0032] Where L represents the loss value, This represents the loss function of the adversarial network. Describes the L1 loss function. This represents the VGG perceptual loss function. and As weight;

[0033] Adversarial network loss function The formula is,

[0034] ,

[0035] Where E[] represents the average value of the discrimination matrix, This refers to the discriminative network. This refers to the generator network. and The output is The discriminant matrix, This represents the output of the generating network.

[0036] On the other hand, this embodiment also provides a device for eliminating streak artifacts in light-sheet fluorescence microscopy images, including,

[0037] The training data acquisition unit is used to acquire multiple striped images obtained by scanning biological sample tissue at the same location from multiple angles using a light-sheet fluorescence microscope, and to construct a stripe-free image of the biological sample tissue at that location based on the multiple striped images.

[0038] A deep adversarial network training unit is used to train a deep adversarial network based on multiple training datasets of the biological sample tissue, using multiple striped images and corresponding non-striped images as training datasets.

[0039] The stripe artifact removal unit is used to remove stripe artifacts from the striped image of the target biological sample tissue using the trained deep adversarial network. The striped image of the target biological sample tissue is obtained by scanning the target biological sample tissue using the light-sheet fluorescence microscope.

[0040] On the other hand, embodiments of the present invention also provide a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the above-described method.

[0041] On the other hand, embodiments of the present invention also provide a computer storage medium storing a computer program thereon, which executes the above-described method when run by the processor of a computer device.

[0042] Using the embodiments described herein, no changes to the physical equipment of the light-sheet fluorescence microscope are required, nor is any physical lens added, which greatly reduces the cost of eliminating stripe artifacts. By training a deep adversarial network model to eliminate stripe artifacts in the striped images of biological samples obtained by the light-sheet fluorescence microscope, the clarity of the stripe-free images obtained after eliminating stripe artifacts is improved. In addition, the stripe artifact elimination method for light-sheet fluorescence microscope images described herein is applicable to various light-sheet fluorescence microscope systems, improving the adaptability of the stripe artifact elimination method. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments or prior art described herein, the accompanying drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this article. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of the implementation system for a method to eliminate stripe artifacts in light sheet fluorescence microscope images according to an embodiment of the present invention;

[0045] Figure 2The diagram shown is a flowchart illustrating a method for eliminating stripe artifacts in light-sheet fluorescence microscopy images according to an embodiment of this paper.

[0046] Figure 3 The diagram illustrates the steps involved in this paper to eliminate stripe artifacts in striped images of target biological tissue samples using the trained deep adversarial network.

[0047] Figure 4 The diagram shown is a schematic representation of a stripe artifact elimination device for light-sheet fluorescence microscopy images according to an embodiment of this paper.

[0048] Figure 5 The diagram shown is a structural schematic of the computer device in the embodiment of this article.

[0049] [Explanation of Figure Markers]:

[0050] 101. Terminal;

[0051] 102. Server;

[0052] 401. Training Data Acquisition Unit;

[0053] 402. Deep Adversarial Network Training Unit;

[0054] 403. Stripe artifact elimination unit;

[0055] 502. Computer equipment;

[0056] 504, Processor;

[0057] 506. Memory;

[0058] 508. Drive mechanism;

[0059] 510. Input / output module;

[0060] 512. Input devices;

[0061] 514. Output devices;

[0062] 516. Presentation equipment;

[0063] 518. Graphical User Interface;

[0064] 520. Network interface;

[0065] 522. Communication link;

[0066] 524. Communication bus. Detailed Implementation

[0067] The technical solutions in the embodiments described below will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments described herein, and not all of the embodiments. Based on the embodiments described herein, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this document.

[0068] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings herein are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0069] It should be noted that the acquisition, storage, use, and processing of data in the technical solution of this application all comply with the relevant provisions of national laws and regulations.

[0070] like Figure 1 The diagram illustrates a system implementation of a method for eliminating stripe artifacts in light-sheet fluorescence microscopy images according to an embodiment of the present invention. The system includes a terminal 101 and a server 102, with a communication connection established between the terminal 101 and the server 102 to enable data exchange. Operators can input images acquired by light-sheet fluorescence microscopy into the server 102 via the terminal 101. The server 102 trains a deep adversarial network (DAN) based on the input images from the terminal 101, obtaining and storing the DAN model for eliminating stripe artifacts. When stripe artifacts in an image with stripes need to be eliminated, operators can input the striped image containing the artifacts to be eliminated into the server 102 via the terminal 101. The server 102 uses the trained DAN model to calculate the striped image input from the terminal 101, obtaining a corresponding stripe-free image, and then sends the obtained stripe-free image back to the terminal 101 for further processing by the user.

[0071] In the embodiments of this specification, the server 102 may be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0072] In one optional embodiment, terminal 101 may be an electronic device, including but not limited to desktop computers, tablets, laptops, etc. Optionally, the operating system running on the electronic device may include, but is not limited to, Android, iOS, Linux, Windows, etc.

[0073] In addition, it should be noted that, Figure 1 The example shown is merely one application environment provided by this disclosure. In practical applications, other application environments may also be included, and this specification does not impose any limitations.

[0074] To address the problems existing in the prior art, this embodiment provides a method for eliminating stripe artifacts in light-sheet fluorescence microscope images. This method eliminates stripe artifacts in images with stripes, can be applied to most light-sheet fluorescence microscopes, does not require changes to the physical equipment, and is suitable for large-scale image data. Figure 2 The diagram illustrates a flowchart of a method for eliminating stripe artifacts in light-sheet fluorescence microscopy images according to an embodiment of this paper. This diagram depicts the process of training a deep adversarial network model and eliminating stripe artifacts, but conventional or non-creative work may include more or fewer steps. The order of steps listed in the embodiment is merely one possible execution order among many and does not represent the only possible order. In actual system or device products, the methods shown in the embodiment or the accompanying drawings can be executed sequentially or in parallel. Specifically, as shown... Figure 2 As shown, the method, which can be executed by a processor, may include:

[0075] Step 201: Obtain multiple striped images of the biological sample tissue at the same location by scanning from multiple angles using a light-sheet fluorescence microscope, and construct a stripe-free image of the biological sample tissue at that location based on the multiple striped images.

[0076] Step 202: Using multiple striped images and corresponding non-striped images as training datasets, train a deep adversarial network based on the multiple training datasets of the biological sample tissue;

[0077] Step 203: Use the trained deep adversarial network to eliminate stripe artifacts in the striped image of the target biological sample tissue, which is obtained by scanning the target biological sample tissue with the light-sheet fluorescence microscope.

[0078] Using the embodiments described herein, no changes to the physical equipment of the light-sheet fluorescence microscope are required, nor is any physical lens added, which greatly reduces the cost of eliminating stripe artifacts. By training a deep adversarial network model to eliminate stripe artifacts in the striped images of biological samples obtained by the light-sheet fluorescence microscope, the clarity of the stripe-free images obtained after eliminating stripe artifacts is improved. In addition, the stripe artifact elimination method for light-sheet fluorescence microscope images described herein is applicable to various light-sheet fluorescence microscope systems, improving the adaptability of the stripe artifact elimination method.

[0079] In this embodiment, the image of a biological tissue sample at the same location from one angle obtained by light-sheet fluorescence microscopy does not show stripes in all parts; it also includes stripe-free portions. Therefore, the stripe-free portions can be fused from images obtained from multiple angles to construct a stripe-free image for that location. Then, the striped and stripe-free images are preprocessed, such as cropping, and combined to form a training dataset. ,in For images with stripes, These are stripe-free images; all images are... A 16-bit grayscale image of pixels.

[0080] In addition, stripeless images can be constructed, and stripe artifacts can be manually added to the constructed stripeless images to form a training dataset. This embodiment does not impose any limitations.

[0081] In this embodiment, the deep adversarial network includes a generator network and a discriminator network. The generator network is used to calculate the non-striped image corresponding to the striped image. The non-striped image is a calculated value. The discriminator network is used to determine whether the non-striped image calculated by the generator network is true.

[0082] In this embodiment of the invention, the left half of the generator network is an encoder, and the right half is a decoder. The step of training the generator network includes training the encoder using the training dataset, and then training the decoder using the computation results of the encoder to obtain the training result of the generator network.

[0083] According to one embodiment of this document, the encoder includes an encoder initialization module and n downsampling encoding modules consisting of convolutional layers and convolutional attention modules, where n is a positive integer greater than 1;

[0084] The steps for training the encoder are as follows: the training dataset is initialized using the encoder initialization module, and multiple downsampling encoding modules are trained using the initialized training dataset, wherein the output of the previous downsampling encoding module is the input of the next downsampling encoding module, and the initialized training dataset is used as the input of the first downsampling encoding module.

[0085] During the training process of the encoder initialization module, the input of the downsampling encoding module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels and number of output channels. Then, the calculation result of the convolution attention module is activated using the LeakyRelu activation function to provide the network with non-linear modeling capability. The calculation result of the LeakyRelu activation function is used as the output of the encoder initialization module.

[0086] During the training process of the first downsampling encoding module to the (n-1)th downsampling encoding module, the input of the downsampling encoding module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the result of the convolution calculation is normalized using the InstanceNorm2d normalization function. Next, the Convolutional Attention Module (CBAM) is used to perform feature selection and enhancement on the normalized result. Finally, the calculation result of the Convolutional Attention Module is activated using the LeakyRelu activation function to provide the network with non-linear modeling capabilities. The calculation result of the LeakyRelu activation function is used as the output of the downsampling encoding module.

[0087] During the training of the last downsampling encoding module, the input of the downsampling encoding module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the calculation result of the convolutional attention module is activated using the ReLU activation function to provide the network with non-linear modeling capabilities. The calculation result of the ReLU activation function is used as the calculation result of the encoder.

[0088] In this embodiment, the encoder may include an encoder initialization module, a first downsampling encoding module, a second downsampling encoding module, a third downsampling encoding module, a fourth downsampling encoding module, a fifth downsampling encoding module, a sixth downsampling encoding module, and a seventh downsampling encoding module;

[0089] The encoder initialization module takes images from the training dataset as input. The kernel size is ( The step size is ( Fill with The number of input channels is 1, the number of output channels is 64, and the LeakyRelu activation function is used to provide the nonlinear modeling capability of the network. The formula corresponding to the initialization module is (1).

[0090] (1)

[0091] in, This indicates the output of the initialization module. This represents the LeakyRelu activation function. Represents the convolution function;

[0092] The kernel size of the first downsampling encoding module ( The step size is ( Fill with The number of input channels is 64 and the number of output channels is 128. The InstanceNorm2d function is used for normalization, and the convolutional attention module is used for feature selection and enhancement to highlight the main features and suppress irrelevant features. Finally, the LeakyRelu activation function is used to provide the non-linear modeling capability of the network. The formula corresponding to the first downsampling encoding module is (2).

[0093] (2)

[0094] in, This represents the output of the first downsampling encoding module. This represents the convolutional attention module. Represents the normalization function;

[0095] The kernel size of the second downsampling encoding module ( The step size is ( Fill with The number of input channels is 128, the number of output channels is 256, and the InstanceNorm2d function is used for normalization. The convolutional attention module is used for feature selection and enhancement to highlight the main features and suppress irrelevant features. Finally, the LeakyRelu activation function is used to provide the non-linear modeling capability of the network. The formula corresponding to the second downsampling encoding module is (3).

[0096] (3)

[0097] in, This indicates the output of the second downsampling encoding module;

[0098] The kernel size of the third downsampling encoding module ( The step size is ( Fill with The number of input channels is 256 and the number of output channels is 512. The InstanceNorm2d function is used for normalization, and the convolutional attention module is used for feature selection and enhancement to highlight the main features and suppress irrelevant features. Finally, the LeakyRelu activation function is used to provide the non-linear modeling capability of the network. The formula corresponding to the third downsampling encoding module is (4).

[0099] (4)

[0100] in, This represents the output of the third downsampling encoding module;

[0101] The kernel size of the fourth downsampling encoding module ( The step size is ( Fill with The number of input channels is 512, the number of output channels is 512, and the InstanceNorm2d function is used for normalization. The convolutional attention module is used for feature selection and enhancement to highlight the main features and suppress irrelevant features. Finally, the LeakyRelu activation function is used to provide the non-linear modeling capability of the network. The formula corresponding to the fourth downsampling encoding module is (5).

[0102] (5)

[0103] in, This represents the output of the fourth downsampling encoding module;

[0104] The kernel size of the fifth downsampling encoding module ( The step size is ( Fill with The number of input channels is 512, the number of output channels is 512, and the InstanceNorm2d function is used for normalization. The convolutional attention module is used for feature selection and enhancement to highlight the main features and suppress irrelevant features. Finally, the LeakyRelu activation function is used to provide the non-linear modeling capability of the network. The formula corresponding to the fifth downsampling encoding module is (6).

[0105] (6)

[0106] in, This represents the output of the fifth downsampling encoding module;

[0107] The kernel size of the sixth downsampling encoding module ( The step size is ( Fill with The number of input channels is 512, the number of output channels is 512, and the InstanceNorm2d function is used for normalization. The convolutional attention module is used for feature selection and enhancement to highlight the main features and suppress irrelevant features. Finally, the LeakyRelu activation function is used to provide the non-linear modeling capability of the network. The formula corresponding to the sixth downsampling encoding module is (7).

[0108] (7)

[0109] in, This represents the output of the sixth downsampling encoding module;

[0110] The kernel size of the seventh downsampling encoding module ( The step size is ( Fill with The number of input channels is 512 and the number of output channels is 512. The ReLU activation function is used to provide the nonlinear modeling capability of the network. The formula corresponding to the seventh downsampling coding module is (8).

[0111] (8)

[0112] in, This represents the output of the sixth downsampling encoding module.

[0113] According to one embodiment of this article, the decoder includes a decoder initialization module and n upsampling decoding modules consisting of upsampling convolutional kernels, skip-layer connections, and convolutional attention modules. The number of upsampling decoding modules included in the decoder is the same as the number of downsampling layers in the encoder and corresponds to them in reverse order, where n is a positive integer greater than 1.

[0114] The steps for training the decoder are as follows: the encoder's calculation result is initialized using the decoder initialization module, and multiple upsampling decoding modules are trained using the initial encoder's calculation result. The output of the previous upsampling decoding module is the input of the next upsampling decoding module, and the initial encoder's calculation result is used as the input of the first upsampling decoding module.

[0115] During the training process of the decoder initialization module, the encoder's calculation results are first convolved using the corresponding convolution function ConvTranspose2d. Then, the convolution calculation results are normalized using the normalization function InstanceNorm2d. Next, the normalized results are feature-selected and enhanced using the convolutional attention module CBAM. Finally, the calculation results of the convolutional attention module are activated using the activation function ReLU to provide the network with non-linear modeling capabilities. The calculation result of the activation function ReLU is used as the output of the first upsampling decoding module.

[0116] During the training process of the first upsampling decoding module to the (n-1)th upsampling decoding module, skip-layer connections are first used to link the input of the upsampling decoding module with the output of the corresponding downsampling encoding module to obtain the linking result. Then, the linking result is convolved using the convolution function ConvTranspose2d with set kernel size, stride, padding, number of input channels, and number of output channels. The result of the convolution calculation is then normalized using the normalization function InstanceNorm2d. The normalized result is then enhanced using the convolutional attention module CBAM. Finally, the result of the convolutional attention module is activated using the activation function ReLU to provide the non-linear modeling capability of the network. The result of the activation function ReLU is used as the output of the upsampling decoding module.

[0117] During the training of the last upsampling decoding module, skip-layer connections are first used to link the input of the upsampling decoding module with the output of the corresponding downsampling encoding module to obtain the linking result. Then, the linking result is convolved using the ConvTranspose2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Finally, the convolution result is activated using the Tanh activation function to provide the network with non-linear modeling capabilities. The calculation result of the Tanh activation function is used as the stripe-free image obtained by the generator network from the striped images corresponding to the training dataset.

[0118] In this embodiment, the decoder may include a decoder initialization module, a first upsampling decoding module, a second upsampling decoding module, a third upsampling decoding module, a fourth upsampling decoding module, a fifth upsampling decoding module, a sixth upsampling decoding module, and a seventh upsampling decoding module;

[0119] The kernel size of the decoder initialization module is ( The step size is ( Fill with The number of input channels is 512, and the InstanceNorm2d function is used for normalization. The convolutional attention module is used for feature selection and enhancement to highlight the main features and suppress irrelevant features. Finally, the ReLU activation function is used to provide the non-linear modeling capability of the network. The formula corresponding to the decoder initialization module is (9).

[0120] (9)

[0121] in, This represents the output of the decoder initialization module;

[0122] The kernel size of the first upsampling decoding module is ( The step size is ( Fill with The input channel has 1024 channels and the output channel has 512 channels. Normalization is performed using the InstanceNorm2d function. A convolutional attention module is used for feature selection and enhancement, highlighting key features and suppressing irrelevant features. Finally, the ReLU activation function is used to provide the network with non-linear modeling capabilities. The formula corresponding to the first upsampling decoding module is (10), which indicates a skip link.

[0123] (10)

[0124] in, This represents the output of the first upsampling decoding module;

[0125] The kernel size of the second upsampling decoding module is ( The step size is ( Fill with The input channel has 1024 channels and the output channel has 512 channels. Normalization is performed using the InstanceNorm2d function. A convolutional attention module is used for feature selection and enhancement, highlighting key features and suppressing irrelevant features. Finally, the ReLU activation function is used to provide the network with non-linear modeling capabilities. The formula corresponding to the second upsampling decoding module is (11), which indicates a skip link.

[0126] (11)

[0127] in, This represents the output of the second upsampling decoding module;

[0128] The kernel size of the third upsampling decoding module is ( The step size is ( Fill with The input channel has 1024 channels and the output channel has 512 channels. Normalization is performed using the InstanceNorm2d function. A convolutional attention module is used for feature selection and enhancement, highlighting key features and suppressing irrelevant features. Finally, the ReLU activation function is used to provide the network with non-linear modeling capabilities. The formula corresponding to the third upsampling decoding module is (12), which indicates a skip-layer link.

[0129] (12)

[0130] in, This represents the output of the third upsampling decoding module;

[0131] The kernel size of the fourth upsampling decoding module is ( The step size is ( Fill with The input channel has 1024 channels and the output channel has 256 channels. Normalization is performed using the InstanceNorm2d function. A convolutional attention module is used for feature selection and enhancement, highlighting key features and suppressing irrelevant features. Finally, the ReLU activation function is used to provide the network with non-linear modeling capabilities. The formula corresponding to the fourth upsampling decoding module is (13), which indicates a skip link.

[0132] (13)

[0133] in, This represents the output of the fourth upsampling decoding module;

[0134] The kernel size of the fifth upsampling decoding module is ( The step size is ( Fill with The input has 512 channels and the output has 128 channels. Normalization is performed using the InstanceNorm2d function. A convolutional attention module is used for feature selection and enhancement, highlighting key features and suppressing irrelevant features. Finally, the ReLU activation function is used to provide the network with non-linear modeling capabilities. The formula corresponding to the fifth upsampling decoding module is (14), which indicates a skip link.

[0135] (14)

[0136] in, This represents the output of the fifth upsampling decoding module;

[0137] The kernel size of the sixth upsampling decoding module is ( The step size is ( Fill with The input has 256 channels and the output has 64 channels. Normalization is performed using the InstanceNorm2d function. A convolutional attention module is used for feature selection and enhancement, highlighting key features and suppressing irrelevant features. Finally, the ReLU activation function is used to provide the network with non-linear modeling capabilities. The formula corresponding to the sixth upsampling decoding module is (15), which indicates a skip link.

[0138] (15)

[0139] in, This represents the output of the sixth upsampling decoding module;

[0140] The kernel size of the seventh upsampling decoding module is ( The step size is ( Fill with The network has 128 input channels and 1 output channel, and uses the Tanh activation function to provide nonlinear modeling capabilities. The formula corresponding to the seventh upsampling decoding module is (16), which indicates a skip link.

[0141] (16)

[0142] in, This represents the output of the seventh upsampling decoding module, and the output of the seventh upsampling encoding module. The stripe-free image is obtained by the generator network from the striped images in the training dataset. .

[0143] This can be understood as follows: unlike the traditional Unet generative network which uses the BatchNorm2d normalization function, this paper uses the InstanceNorm2d normalization function for normalization, which is more conducive to generating high-resolution grayscale images and improves output and training effects.

[0144] According to one embodiment of this article, the discriminant network includes a discriminant network initialization module and m convolutional modules, where m is a positive integer greater than 1;

[0145] The calculation process of the discriminant network is as follows: the discriminant network initialization module initializes the stripeless image calculated by the generator network, and the initialized stripeless image is used to train multiple convolutional modules. The output of the previous convolutional module is the input of the next convolutional module, the initialized stripeless image is the input of the first convolutional module, and the output of the last convolutional module is the discrimination result of the discriminant network.

[0146] During the calculation process of the discriminant network initialization module, the stripeless image obtained by the generator network is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the result of the convolution calculation is activated using the LeakyRelu activation function to provide the network with nonlinear modeling capabilities. The calculation result of the LeakyRelu activation function is used as the output of the discriminant network initialization module, and the output of the discriminant network initialization module is input into the first convolution module.

[0147] During the computation of the first to the (m-1)th convolutional module, the input of the convolutional module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the result of the convolution calculation is normalized using the InstanceNorm2d normalization function. Finally, the normalized result is activated using the LeakyRelu activation function to provide the network with nonlinear modeling capabilities. The calculation result of the LeakyRelu activation function is used as the output of the convolutional module.

[0148] In the calculation process of the last convolution module, the input of the convolution module is first convolved according to the Conv2d convolution function with set kernel size, stride, padding, number of input channels and number of output channels. Finally, the sigmoid activation function is used to activate the result of the convolution calculation, providing the nonlinear modeling capability of the network. The calculation result of the sigmoid activation function is used as the discrimination result of the discriminant network.

[0149] In this embodiment, the discriminant network may include a discriminant network initialization module, a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module, and a fifth convolutional module;

[0150] The stripeless image obtained by the generative network is The kernel size of the convolutional module in the network initialization module is determined as ( The step size is ( Fill with The number of input channels is 1 and the number of output channels is 32. Then, the LeakyRelu activation function is used to activate the result of the convolution calculation, providing the nonlinear modeling capability of the network. The formula for the discrimination network initialization module is (17).

[0151] (17)

[0152] in, This represents the output of the discrimination network initialization module;

[0153] The kernel size of the first convolutional module is ( The step size is ( Fill with The number of input channels is 32 and the number of output channels is 64. Then, the normalization function InstanceNorm2d is used to normalize the result of the convolution calculation. Finally, the activation function LeakyRelu is used to activate the normalized result, providing the nonlinear modeling capability of the network. The formula for the first convolution module is (18).

[0154] (18)

[0155] in, This represents the output of the first convolutional module;

[0156] The kernel size of the second convolutional module is ( The step size is ( Fill with The number of input channels is 64 and the number of output channels is 128. Then, the normalization function InstanceNorm2d is used to normalize the result of the convolution calculation. Finally, the activation function LeakyRelu is used to activate the normalized result, providing the nonlinear modeling capability of the network. The formula for the second convolution module is (19).

[0157] (19)

[0158] in, This represents the output of the second convolutional module;

[0159] The kernel size of the third convolutional module is ( The step size is ( Fill with The number of input channels is 128 and the number of output channels is 256. Then, the normalization function InstanceNorm2d is used to normalize the result of the convolution calculation. Finally, the activation function LeakyRelu is used to activate the normalized result, providing the nonlinear modeling capability of the network. The formula for the third convolution module is (20).

[0160] (20)

[0161] in, This represents the output of the third convolutional module;

[0162] The kernel size of the fourth convolutional module is ( The step size is ( Fill with The number of input channels is 256 and the number of output channels is 512. Then, the normalization function InstanceNorm2d is used to normalize the result of the convolution calculation. Finally, the activation function LeakyRelu is used to activate the normalized result, providing the nonlinear modeling capability of the network. The formula for the fourth convolution module is (21).

[0163] (twenty one)

[0164] in, This represents the output of the fourth convolutional module;

[0165] The kernel size of the fifth convolutional module is ( The step size is ( Fill with The number of input channels is 512 and the number of output channels is 1. Then, the sigmoid activation function is used to activate the convolution result to provide the nonlinear modeling capability of the network. The formula for the fifth convolution module is (22).

[0166] (twenty two)

[0167] in, This represents the output of the fifth convolutional module. As the discrimination result of the discriminant network Where P is The discrimination matrix P is used, where each element represents the probability of correctly classifying that part. Compared to traditional discriminators, this paper uses the discrimination matrix P to judge the results of the generator network, which can improve the discrimination results for high-definition images.

[0168] According to one embodiment of this paper, the loss function in the process of training a deep adversarial network is Equation (23).

[0169] (twenty three)

[0170] Where L represents the loss value, This represents the loss function of the adversarial network. Describes the L1 loss function. This represents the VGG perceptual loss function. and As weight;

[0171] The L1 loss function is given by formula (24):

[0172] (twenty four)

[0173] Where N is the number of pixels in the image, Represents real images The 1 pixel, Representative generated image The 1 pixel;

[0174] The VGG perceptual loss function firstly... and The image is expanded into a three-channel image, and a loss function is constructed using a pre-trained VGG16 network. This network can be downloaded directly using the PyTorch deep learning framework, and will not be described in detail here.

[0175] Adversarial network loss function The formula is (25).

[0176] (25)

[0177] Where E[] represents the average value of the discrimination matrix, This refers to the discriminative network. This refers to the generator network. and The output is The discriminant matrix, This represents the output of the generating network.

[0178] Preferably, Set to 200. Set to 1, the learning rate is The Adam optimization algorithm is used to continuously reduce the loss function. When both the L1 loss function and the VGG loss function are less than 0.05, and When the loss function converges, training is considered complete. The trained generative network can then be used to eliminate image stripes. Used to calculate the average of an array.

[0179] In the embodiments described herein, such as Figure 3 As shown, eliminating stripe artifacts in striped images of target biological tissue samples using the trained deep adversarial network may include the following steps:

[0180] Step 301: Use the beam scanning method of light-sheet fluorescence microscopy to acquire biological sample tissue images as raw images. The image type is 16-bit grayscale.

[0181] Step 302: Perform content detection on the original image and segment the original image into multiple images. The small images are recorded, and the location information corresponding to each small image is recorded;

[0182] Step 303: Normalize each small image, use the deep adversarial network trained in this embodiment to eliminate image stripes and enhance image quality, and then convert the image back to a 16-bit grayscale image.

[0183] Step 304: Based on the position information corresponding to each small image recorded, the small images after the stripes have been removed are re-stitched together to restore an image of the same size as the original image, thus completing the removal of stripe artifacts from the original image.

[0184] In the embodiments described herein, the images after stripe artifact removal can be used for research such as neural tracking and cell segmentation, and the embodiments described herein do not impose any limitations.

[0185] Based on the same inventive concept, this embodiment also provides a device for eliminating streak artifacts in light-sheet fluorescence microscope images, such as... Figure 4 As shown, it includes:

[0186] The training data acquisition unit 401 is used to acquire multiple striped images obtained by scanning biological sample tissue at the same location from multiple angles using a light-sheet fluorescence microscope, and to construct a stripe-free image of the biological sample tissue at that location based on the multiple striped images.

[0187] The deep adversarial network training unit 402 is used to train the deep adversarial network based on multiple training datasets of the biological sample tissue, using multiple striped images and corresponding non-striped images as training datasets.

[0188] The stripe artifact removal unit 403 is used to remove stripe artifacts from the striped image of the target biological sample tissue using the trained deep adversarial network. The striped image of the target biological sample tissue is obtained by scanning the target biological sample tissue using the light-sheet fluorescence microscope.

[0189] Since the principle of the above-mentioned device in solving the problem is similar to that of the above-mentioned method, the implementation of the above-mentioned device can refer to the implementation of the above-mentioned method, and the repeated parts will not be described again.

[0190] like Figure 5As shown, a computer device provided in this embodiment is described. The apparatus described herein can be the computer device described in this embodiment, performing the methods described above. The computer device 502 may include one or more processors 504, such as one or more central processing units (CPUs), each of which can implement one or more hardware threads. The computer device 502 may also include any memory 506 for storing information of any kind, such as code, settings, data, etc. Without limitation, for example, memory 506 may include any type of RAM, any type of ROM, flash memory, hard disk, optical disk, etc. More generally, any memory can use any technology to store information. Further, any memory can provide volatile or non-volatile retention of information. Further, any memory may represent a fixed or removable component of the computer device 502. In one case, when processor 504 executes associated instructions stored in any memory or combination of memories, the computer device 502 can perform any operation of the associated instructions. The computer device 502 also includes one or more drive mechanisms 508 for interacting with any memory, such as hard disk drive mechanisms, optical disk drive mechanisms, etc.

[0191] Computer device 502 may also include an input / output module 510 (I / O) for receiving various inputs (via input device 512) and providing various outputs (via output device 514). A specific output mechanism may include a presentation device 516 and an associated graphical user interface (GUI) 518. In other embodiments, the input / output module 510 (I / O), input device 512, and output device 514 may be omitted, and the device may function solely as a computer device within a network. Computer device 502 may also include one or more network interfaces 520 for exchanging data with other devices via one or more communication links 522. One or more communication buses 524 couple the components described above together.

[0192] Communication link 522 can be implemented in any way, such as via a local area network, a wide area network (e.g., the Internet), a point-to-point connection, or any combination thereof. Communication link 522 may include any combination of hardwired links, wireless links, routers, gateway functions, name servers, etc., governed by any protocol or combination of protocols.

[0193] Corresponding to Figures 2-3 In addition to the method described above, this embodiment also provides a computer-readable storage medium storing a computer program that is executed by a processor to perform the above steps.

[0194] This embodiment also provides a computer-readable instruction, wherein when a processor executes the instruction, the program therein causes the processor to perform the following: Figures 2-3 The method of demonstration.

[0195] It should be understood that in the various embodiments of this document, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this document.

[0196] It should also be understood that, in the embodiments herein, the term "and / or" is merely a description of the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following associated objects have an "or" relationship.

[0197] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this document.

[0198] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0199] In the embodiments provided herein, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections through some interfaces, devices, or units, or they may be electrical, mechanical, or other forms of connection.

[0200] 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 units can be selected to achieve the purpose of the embodiments described herein, depending on actual needs.

[0201] Furthermore, the functional units in the various embodiments of this document can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0202] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this paper, in essence, or the part that contributes to the prior art, or all or 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 this paper. 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.

[0203] This document uses specific embodiments to illustrate the principles and implementation methods of this document. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and core ideas of this document. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this document. Therefore, the content of this specification should not be construed as a limitation of this document.

Claims

1. A method of stripe artifact removal for light sheet fluorescence microscopy images, the method comprising: include: ​ Multiple striped images of a biological sample tissue are obtained by scanning the same location at multiple angles using a light-sheet fluorescence microscope, and a stripe-free image of the biological sample tissue at that location is constructed based on the multiple striped images. Multiple striped images and corresponding non-striped images are used as training datasets to train a deep adversarial network based on the multiple training datasets of the biological sample tissue. The trained deep adversarial network is used to eliminate stripe artifacts in striped images of target biological sample tissues, which are obtained by scanning the target biological sample tissues using the light-sheet fluorescence microscope. The deep adversarial network includes a generator network and a discriminator network. The generator network is used to calculate the non-striped image corresponding to the striped image, and the discriminator network is used to determine whether the non-striped image calculated by the generator network is true. The left half of the generator network is an encoder, and the right half is a decoder; The step of training the generative network includes training the encoder using the training dataset, and then training the decoder using the computation results of the encoder to obtain the training result of the generative network; The encoder includes an encoder initialization module and n downsampling encoding modules consisting of convolutional layers and convolutional attention modules, where n is a positive integer greater than 1; The steps for training the encoder are as follows: the training dataset is initialized using the encoder initialization module, and multiple downsampling encoding modules are trained using the initialized training dataset, wherein the output of the previous downsampling encoding module is the input of the next downsampling encoding module, and the initialized training dataset is used as the input of the first downsampling encoding module. During the training process of the encoder initialization module, the input of the downsampling encoding module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels and number of output channels. Then, the result of the convolution calculation is activated using the LeakyRelu activation function to provide the network with nonlinear modeling capabilities. The calculation result of the LeakyRelu activation function is used as the output of the encoder initialization module. During the training process of the first downsampling encoding module to the (n-1)th downsampling encoding module, the input of the downsampling encoding module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the result of the convolution calculation is normalized using the InstanceNorm2d normalization function. Next, the Convolutional Attention Module (CBAM) is used to perform feature selection and enhancement on the normalized result. Finally, the calculation result of the Convolutional Attention Module is activated using the LeakyRelu activation function to provide the network with non-linear modeling capabilities. The calculation result of the LeakyRelu activation function is used as the output of the downsampling encoding module. During the training of the last downsampling encoding module, the input of the downsampling encoding module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the ReLU activation function is used to activate the result of the convolution calculation, providing the network with nonlinear modeling capabilities. The calculation result of the ReLU activation function is used as the calculation result of the encoder.

2. The method of claim 1, wherein, The decoder includes a decoder initialization module and n upsampling decoding modules consisting of upsampling convolutional kernels, skip-layer connections, and convolutional attention modules. The number of upsampling decoding modules included in the decoder is the same as the number of downsampling layers in the encoder and corresponds to them in reverse order. n is a positive integer greater than 1. The steps for training the decoder are as follows: the encoder's calculation result is initialized using the decoder initialization module, and multiple upsampling decoding modules are trained using the initial encoder's calculation result. The output of the previous upsampling decoding module is the input of the next upsampling decoding module, and the initial encoder's calculation result is used as the input of the first upsampling decoding module. During the training process of the decoder initialization module, the encoder's calculation results are first convolved using the corresponding convolution function ConvTranspose2d. Then, the convolution calculation results are normalized using the normalization function InstanceNorm2d. Next, the normalized results are feature-selected and enhanced using the convolutional attention module CBAM. Finally, the calculation results of the convolutional attention module are activated using the activation function ReLU to provide the network with non-linear modeling capabilities. The calculation result of the activation function ReLU is used as the output of the first upsampling decoding module. During the training process of the first upsampling decoding module to the (n-1)th upsampling decoding module, skip-layer connections are first used to link the input of the upsampling decoding module with the output of the corresponding downsampling encoding module to obtain the linking result. Then, the linking result is convolved using the convolution function ConvTranspose2d with set kernel size, stride, padding, number of input channels, and number of output channels. The result of the convolution calculation is then normalized using the normalization function InstanceNorm2d. The normalized result is then enhanced using the convolutional attention module CBAM. Finally, the result of the convolutional attention module is activated using the activation function ReLU to provide the non-linear modeling capability of the network. The result of the activation function ReLU is used as the output of the upsampling decoding module. During the training of the last upsampling decoding module, skip-layer connections are first used to link the input of the upsampling decoding module with the output of the corresponding downsampling encoding module to obtain the linking result. Then, the linking result is convolved using the ConvTranspose2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Finally, the convolution result is activated using the Tanh activation function to provide the network with non-linear modeling capabilities. The calculation result of the Tanh activation function is used as the stripe-free image obtained by the generator network from the striped images corresponding to the training dataset.

3. The method of claim 1, wherein, The discriminant network includes a discriminant network initialization module and m convolutional modules, where m is a positive integer greater than 1; The calculation process of the discriminant network is as follows: the discriminant network initialization module initializes the stripeless image calculated by the generator network, and the initialized stripeless image is used to train multiple convolutional modules. The output of the previous convolutional module is the input of the next convolutional module, the initialized stripeless image is the input of the first convolutional module, and the output of the last convolutional module is the discrimination result of the discriminant network. During the calculation process of the discriminant network initialization module, the stripeless image obtained by the generator network is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the result of the convolution calculation is activated using the LeakyRelu activation function to provide the network with nonlinear modeling capabilities. The calculation result of the LeakyRelu activation function is used as the output of the discriminant network initialization module, and the output of the discriminant network initialization module is input into the first convolution module. During the computation of the first to the (m-1)th convolutional module, the input of the convolutional module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the result of the convolution calculation is normalized using the InstanceNorm2d normalization function. Finally, the normalized result is activated using the LeakyRelu activation function to provide the network with nonlinear modeling capabilities. The calculation result of the LeakyRelu activation function is used as the output of the convolutional module. In the calculation process of the last convolution module, the input of the convolution module is first convolved according to the Conv2d convolution function with set kernel size, stride, padding, number of input channels and number of output channels. Finally, the sigmoid activation function is used to activate the result of the convolution calculation, providing the nonlinear modeling capability of the network. The calculation result of the sigmoid activation function is used as the discrimination result of the discriminant network.

4. The method of claim 1, wherein, The loss function during the training of a deep adversarial network is: ， Where L represents the loss value, This represents the loss function of the adversarial network. Describes the L1 loss function. This represents the VGG perceptual loss function. and As weight; Adversarial network loss function The formula is ， Where E[] represents the average value of the discrimination matrix, This refers to the discriminative network. This refers to the generator network. and The output is The discriminant matrix, Represents the output of the generating network, wherein For images with stripes, This is a stripe-free image.

5. An apparatus for stripe artifact removal of light sheet fluorescence microscopy images, characterized in that, include: The training data acquisition unit is used to acquire multiple striped images obtained by scanning biological sample tissue at the same location from multiple angles using a light-sheet fluorescence microscope, and to construct a stripe-free image of the biological sample tissue at that location based on the multiple striped images. A deep adversarial network training unit is used to train a deep adversarial network based on multiple training datasets of the biological sample tissue, using multiple striped images and corresponding non-striped images as training datasets. The stripe artifact removal unit is used to remove stripe artifacts from the striped image of the target biological sample tissue using the trained deep adversarial network. The striped image of the target biological sample tissue is obtained by scanning the target biological sample tissue using the light-sheet fluorescence microscope. The deep adversarial network includes a generator network and a discriminator network. The generator network is used to calculate the non-striped image corresponding to the striped image, and the discriminator network is used to determine whether the non-striped image calculated by the generator network is true. The left half of the generator network is an encoder, and the right half is a decoder; The step of training the generative network includes training the encoder using the training dataset, and then training the decoder using the computation results of the encoder to obtain the training result of the generative network; The encoder includes an encoder initialization module and n downsampling encoding modules consisting of convolutional layers and convolutional attention modules, where n is a positive integer greater than 1; The steps for training the encoder are as follows: the training dataset is initialized using the encoder initialization module, and multiple downsampling encoding modules are trained using the initialized training dataset, wherein the output of the previous downsampling encoding module is the input of the next downsampling encoding module, and the initialized training dataset is used as the input of the first downsampling encoding module. During the training process of the encoder initialization module, the input of the downsampling encoding module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels and number of output channels. Then, the result of the convolution calculation is activated using the LeakyRelu activation function to provide the network with nonlinear modeling capabilities. The calculation result of the LeakyRelu activation function is used as the output of the encoder initialization module. During the training process of the first downsampling encoding module to the (n-1)th downsampling encoding module, the input of the downsampling encoding module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the result of the convolution calculation is normalized using the InstanceNorm2d normalization function. Next, the Convolutional Attention Module (CBAM) is used to perform feature selection and enhancement on the normalized result. Finally, the calculation result of the Convolutional Attention Module is activated using the LeakyRelu activation function to provide the network with non-linear modeling capabilities. The calculation result of the LeakyRelu activation function is used as the output of the downsampling encoding module. During the training of the last downsampling encoding module, the input of the downsampling encoding module is first convolved using the Conv2d convolution function with set kernel size, stride, padding, number of input channels, and number of output channels. Then, the ReLU activation function is used to activate the result of the convolution calculation, providing the network with nonlinear modeling capabilities. The calculation result of the ReLU activation function is used as the calculation result of the encoder.

6. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 4.

7. A computer readable storage medium characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 4.

Images