Self-supervised three-dimensional fluorescence microscopic imaging isotropic resolution enhancement method based on diffusion model

Abstract

The invention discloses a self-supervised three-dimensional fluorescence microscopic imaging isotropic resolution enhancement method based on a diffusion model, and aims to solve the problems of insufficient biological structure analysis and complex hardware improvement scheme caused by anisotropic resolution in the existing three-dimensional fluorescence microscopic imaging technology and strong hypothesis dependence of the existing deep learning method. According to the method, model training can be completed only through axial slicing, the strong hypothesis limitation of a traditional method on axial-transverse structural similarity is thoroughly eliminated, and the universality of biological samples is remarkably improved. Compared with a scheme based on adversarial learning, the training process is more stable, and training is easier.

Description

Technical Field

[0001] The invention belongs to the technical field of three-dimensional fluorescence microscopy imaging, and in particular relates to a self-supervised three-dimensional fluorescence microscopy imaging isotropic resolution enhancement method based on a diffusion model. Background Art

[0002] As an important tool in the life sciences, three-dimensional fluorescence microscopy enables high-resolution three-dimensional observation of cellular and molecular structures, supporting long-term tracking of dynamic processes in living systems. It plays a key role in the study of cellular interactions, the analysis of disease mechanisms, and drug development. However, due to the anisotropic point spread function of the microscopic system and the large axial sampling interval, the axial resolution of traditional three-dimensional fluorescence microscopy is significantly lower than the lateral resolution. This anisotropy severely restricts the detailed analysis of complex biological three-dimensional structures. Therefore, achieving three-dimensional isotropic or near-isotropic resolution imaging is of great scientific value.

[0003] In existing technologies, the main approach to improving resolution isotropy is through improvements in the optical path. Typical methods include: multi-view deconvolution technology, such as multi-field line scanning confocal microscopy, which uses three objective lenses to sequentially scan diffraction-limited line illumination and combines multi-field image fusion to improve resolution balance; dual-field plane illumination microscopy, which achieves isotropic resolution through orthogonal volume data fusion; three-dimensional structured light illumination microscopy based on four-beam interference, which uses sample-opposed mirrors to achieve near-isotropic imaging; and axial scanning light sheet microscopy, which improves axial resolution through remote focusing synchronous scanning and rolling shutter acquisition. However, the above hardware improvement solutions have technical bottlenecks such as limited data acquisition rate and complex system transformation.

[0004] In recent years, deep learning-based microscopic image processing has made significant progress in areas such as denoising, deblurring, super-resolution, and aberration compensation. However, enhancing the axial resolution of diffraction-limited anisotropic volume data requires reconstructing the axial high-frequency features while preserving the lateral high-frequency information, a technical challenge significantly greater than conventional image restoration. Existing deep learning schemes mainly include two categories: content-aware image restoration network (CARE) based on one-dimensional fuzzy downsampling to construct semi-synthetic paired data to train the U-Net model [Weiger rt, M., Schmidt, U., Boothe, T. et al. Content-aware image restoration: pushing the limits of fluorescence microscopy. Nat Methods 15, 1090–1097 (2018). https: / / doi.org / 10.1038 / s41592-018-0216-7], and single-stage based on optimal transfer cycle adversarial network (OT-CycleGAN) [Park, H., Na, M., Kim, B. et al. Deep learning enables reference-free isotropic super-resolution for volumetric fluorescence microscopy. Nat Commun 13, 3297 (2022). https: / / doi.org / 10.1038 / s41467-022-30949-6 ] and dual-stage [Ning, K., Lu, B., Wang, https: / / doi.org / 10.1038 / s41377-023-01230-2] Self-supervised framework [Sim, B., Oh, G., Kim, J., Jung, C. & Ye, JC Optimal transport driven cycle GAN for unsupervised learning in inverse problems. SIAM J. Imaging Sci. 13, 2281–2306 (2020).]. However, such methods require a strong assumption that axial and transverse slices have structural similarity, which significantly limits the applicability of biological samples, and OT-CycleGAN has the inherent defect of insufficient training stability. It is worth noting that the denoising diffusion probabilistic model (DDPM) [Ho, Jonathan, Ajay Jain, and Pieter Abbeel."Denoising diffusion probabilistic models." Advances in neutral information processing systems 33 (2020): 6840-6851.] as a new generation of generative models, has demonstrated excellent performance beyond traditional GAN ​​and VAE with its stable training characteristics and unconstrained generation capabilities. Among them, the iterative latent variable optimization method (ILVR) that does not require learning [Choi, Jooyoung, et al. "Ilvr: Conditioning method fordenoising diffusion probabilistic models. In 2021IEEE." CVF international conference on computer vision (ICCV). Vol. 1. 2021.] can also use unconditional DDPM to achieve multimodal conditional generation.

[0005] In summary, existing technologies have the following key drawbacks: (1) Hardware improvement solutions lead to increased system complexity and decreased acquisition efficiency; (2) Supervised learning methods rely on artificially synthesized paired data and are limited by the assumption of structural similarity; (3) Self-supervised solutions suffer from model training instability; (4) Traditional generative models struggle to balance data fidelity and artifact suppression. Therefore, there is an urgent need to develop a high-fidelity 3D isotropic resolution enhancement method that does not rely on hardware modification, does not require structural prior assumptions, and has high fidelity. Summary of the Invention

[0006] In view of the above, the present invention provides a method for isotropic resolution enhancement of self-supervised three-dimensional fluorescence microscopy based on a diffusion model to solve the problems of insufficient biological structure analysis caused by anisotropic resolution in existing three-dimensional fluorescence microscopy technology, complex hardware improvement solutions, and strong assumption dependence of existing deep learning methods. This method only requires axial slicing to complete model training, completely getting rid of the strong assumption limitation of traditional methods on axial-transverse structural similarity, and significantly improving the universality of biological samples. Compared with the solution based on adversarial learning, the training process of this method is more stable and easier to train. The present invention is achieved through the following technical solutions:

[0007] The present invention discloses a method for isotropic resolution enhancement of self-supervised three-dimensional fluorescence microscopy imaging based on a diffusion model, comprising the following steps:

[0008] 1) In a 3D fluorescence microscopy system, excitation light is used to excite fluorescently labeled biological samples, and data is collected through a detection objective. The process of acquiring a 3D image stack involves performing Z-axis layered scanning of the sample, combined with optical sectioning technology and high-precision mechanical control, to ultimately obtain 3D biological sample data;

[0009] 2) obtaining the point spread function of the corresponding three-dimensional fluorescence microscopy imaging system through simulation or actual measurement, then normalizing the point spread function and calculating its back-projection operator;

[0010] 3) performing axial interpolation processing on the three-dimensional biological sample data so that the axial pixel size is consistent with the lateral pixel size, and performing non-blind deconvolution preprocessing on the collected biological sample image data using the Richardson-Lucy deconvolution algorithm or the Wiener-Butterworth filter deconvolution algorithm to obtain a two-dimensional axial slice of the preprocessed data;

[0011] 4) pre-training the unconditional generative diffusion model using only the two-dimensional axial slices of the pre-processed data to learn the distribution characteristics of the two-dimensional axial slices of the data and obtain a pre-trained unconditional diffusion model;

[0012] 5) The two-dimensional axial slices of the preprocessed data in step 3) have orthogonal high-resolution and low-resolution directions. A theoretically estimated one-dimensional Gaussian blur kernel is used to perform one-dimensional blur processing on the high-resolution direction of the two-dimensional axial slices of the preprocessed data. Then, downsampling and interpolation are performed along this direction to a pixel size consistent with the original axial pixel size. Then, upsampling and interpolation are performed to a pixel size consistent with the lateral pixel size to obtain a two-dimensional axial slice of the processed data.

[0013] 6) Using the two-dimensional axial slice processed in step 5) as a conditional image, conditionally generate the image using the pre-trained unconditional diffusion model in step 4), so that the resolution of the two orthogonal dimensions of the generated two-dimensional slice image is consistent with the original axial low resolution;

[0014] 7) Pairing the two-dimensional axial slices of the preprocessed data in step 3) with the two-dimensional low-resolution slices generated in step 6), and then using the paired data to train a neural network, with the latter as input and the former as output;

[0015] 8) rotating the two-dimensional axial slices of the preprocessed data in step 3) at equal intervals by several different angles, and then inputting the rotated images into the neural network trained in step 7) to obtain several corresponding output images, and then reversely rotating the output images by corresponding angles to obtain a final output image;

[0016] 9) Fourier transforming the output images obtained in step 8) respectively, taking the maximum value of their Fourier spectra for fusion, and then performing inverse Fourier transform to obtain isotropic resolution axial slices;

[0017] 10) Repeat step 9) for all axial slices in the XZ and YZ directions, and then average the results in the two directions to finally obtain isotropic high-resolution three-dimensional fluorescence microscopy stack data.

[0018] As a further improvement, the three-dimensional biological sample data collected in step 1) of the present invention is a volume image, which is obtained by stacking two-dimensional slice images, and the axial resolution is lower than the lateral resolution.

[0019] As a further improvement, the point spread function in step 2) of the present invention has three spatial dimensions and is a three-dimensional matrix structure.

[0020] As a further improvement, the pre-training process of the unconditional generative diffusion model in step 4) of the present invention is as follows:

[0021]

[0022] right Perform gradient descent optimization, and repeat the above process until convergence, where x0 is the two-dimensional axial slice of the preprocessed data; q(x0) is the distribution characteristics of the two-dimensional axial slice of the data; t is the number of time steps; Uniform() is a uniform distribution; T is the total number of time steps; noise ∈ is a standard Gaussian distribution with a variance of 1; ∈ θ A neural network for predicting noise distribution; is a coefficient and decreases as t increases.

[0023] As a further improvement, in step 5) of the present invention, a theoretically estimated one-dimensional Gaussian blur kernel is used to perform one-dimensional blur processing on the high-resolution direction of the two-dimensional axial slice of the preprocessed data, and then down-sampled and interpolated along this direction to a pixel size consistent with the original axial pixel size, and then up-sampled and interpolated to a pixel size consistent with the lateral pixel size, specifically:

[0024] y=U(D(Blur(x0)))

[0025] Where x0 is the two-dimensional axial slice of the preprocessed data; Blur() is the theoretically estimated one-dimensional Gaussian blur kernel; D() is a one-dimensional downsampling operator, which downsamples and interpolates along a single dimension to a pixel size consistent with the original axial pixel size; U() is a one-dimensional upsampling operator, which upsamples and interpolates along a single dimension to a pixel size consistent with the horizontal pixel size; y is the processed two-dimensional axial slice.

[0026] As a further improvement, step 6) of the present invention can be specifically expressed as follows:

[0027] Given a conditional image y, sampling The time step number t is traversed from T,…,1 from large to small, and the following process is repeated to finally obtain the new image x0, which is denoted as y0 for differentiation:

[0028] z~N(0,I);

[0029] x′ t-1 ~p θ (x′ t-1 ∣x t );

[0030] y′ t-1 ~q(y′ t-1 |y);

[0031] x t-1 ←(y′ t-1 +x′ t-1 ) / 2;

[0032] Among them, the noise z obeys the standard Gaussian distribution with a variance of 1 to ensure the randomness of the generation process; x t is a two-dimensional image at time step t; p θ (∣) is the process of generating using the pre-trained unconditional generative diffusion model; x' t-1 is a two-dimensional image at time step t-1, generated by the pre-trained unconditional generative diffusion model; q(|) is a two-dimensional image after adding noise of time step t-1 to the conditional image y; x t-1 is the final two-dimensional image at time step t-1.

[0033] As a further improvement, the neural network training process in step 7) of the present invention is expressed as:

[0034] x0=Θ(y0);

[0035] Where x0 is the two-dimensional axial slice of the preprocessed data, which serves as the output of the neural network Θ; y0 is the corresponding image conditionally generated in step 6), which serves as the input of the neural network Θ.

[0036] As a further improvement, the neural network Θ in step 7) of the present invention is a CNN-based two-dimensional image super-resolution neural network, or a Transformer-based two-dimensional image super-resolution neural network.

[0037] As a further improvement, step 8) of the present invention is specifically expressed as follows:

[0038] If there are M angles in total, for the i-th angle,

[0039] o i =iRot i (Θ(Rot i (x0)))

[0040] Where x0 is the two-dimensional axial slice of the preprocessed data; Rot i () is a rotation operator that rotates the two-dimensional image clockwise by angle i; Θ() is the neural network trained in step (7); iRot i () is a rotation operator, which rotates the two-dimensional image counterclockwise by angle i; o i is the output corresponding to the i-th angle.

[0041] As a further improvement, step 9) of the present invention can be specifically expressed as follows:

[0042] o=ifft(Max(fft(o1),...,fft(o i ),...,fft(o M ))))

[0043] Among them, o1 is the output corresponding to the first angle; o i is the output corresponding to the i-th angle; o M is the output corresponding to the Mth angle; fft() is the two-dimensional discrete fast Fourier transform; ifft() is the two-dimensional discrete inverse fast Fourier transform; Max() is the element-wise maximum operator, which takes the maximum value of all matrix elements affected by the operator at each position; o is the isotropic resolution axial slice of the output.

[0044] The beneficial effects of the present invention are as follows:

[0045] The present invention realizes three-dimensional isotropic resolution fluorescence microscopy through a pre-trained unconditional diffusion model, a conditional generation strategy, a paired training neural network strategy, and a multi-angle inference-Fourier maximum fusion strategy. It solves the problems of insufficient biological structure analysis caused by anisotropic resolution in existing three-dimensional fluorescence microscopy technology, complex hardware improvement solutions, and the strong assumption dependence of existing deep learning methods on the structural similarity of biological samples in the axial and lateral directions.

[0046] In summary, the present invention provides a diffusion-based self-supervised (zero-sample) isotropic resolution enhancement method for three-dimensional fluorescence microscopy to address the problems of insufficient biological structure resolution due to anisotropic resolution in existing three-dimensional fluorescence microscopy techniques, complex hardware improvement solutions, and the strong assumptions of existing deep learning methods. This method only requires axial slices to complete model training, completely breaking away from the strong assumptions of axial-transverse structural similarity in traditional methods and significantly improving the universality of biological samples. Compared with solutions based on adversarial learning, the training process of the present invention is more stable and easier to train. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the process of the self-supervised (zero-sample) three-dimensional fluorescence microscopy isotropic resolution enhancement method based on the diffusion model of the present invention;

[0048] Figure 2 A comparison diagram of the results of using the present invention to perform isotropic resolution enhancement on anisotropic resolution simulation data;

[0049] The first row is the slice image (Slice) and maximum intensity projection (MIP) of the intensity distribution phantom with points, lines, shells, spheres, cylinders, etc. in the YZ direction;

[0050] The second row is the slice image (Slice) and maximum intensity projection image (MIP) in the YZ direction of the three-dimensional fluorescence microscopy simulation image stack with anisotropic resolution;

[0051] The third row is a slice image (Slice) and a maximum intensity projection image (MIP) of the isotropic resolution enhancement result after processing by the present invention.

[0052] Figure 3 The present invention is used to compare the maximum intensity projection images in the XY, XZ and YZ directions and the ROI image comparison of the thy1 mouse neural confocal microscopy imaging data before and after isotropic resolution enhancement. DETAILED DESCRIPTION

[0053] In order to describe the present invention more specifically, the technical solution of the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] The excitation light path of the three-dimensional fluorescence microscopy system emits excitation light to illuminate biological samples labeled with fluorescent proteins or containing fluorescent proteins themselves, causing them to emit fluorescence. The signal collection light path of the microscopy system collects fluorescence data, and the focal plane of the collection light path is continuously moved to traverse the signal of the entire biological sample. The images collected each time are stacked to form three-dimensional biological sample data.

[0055] like Figure 1 As shown, the present invention provides a self-supervised three-dimensional fluorescence microscopy isotropic resolution enhancement method based on a diffusion model, comprising the following steps: Figure 1 ):

[0056] S1. In a three-dimensional fluorescence microscopy system, excitation light is used to excite fluorescence of a fluorescently labeled biological sample, and data is collected through a detection objective lens. The process of collecting a three-dimensional image stack involves performing Z-axis axial layered scanning of the sample, combined with optical sectioning technology and high-precision mechanical control, to ultimately obtain three-dimensional biological sample data; the point spread function of the corresponding three-dimensional fluorescence microscopy system is obtained through simulation or actual measurement, and then the point spread function is normalized and its back projection operator is calculated; the axial direction of the three-dimensional biological sample data is interpolated to make the axial pixel size consistent with the lateral pixel size, and the collected biological sample image data is preprocessed by non-blind deconvolution using the Richardson-Lucy deconvolution algorithm or the Wiener-Butterworth filter deconvolution algorithm to obtain a two-dimensional axial slice of the preprocessed data.

[0057] S2. Pre-train the unconditional generative diffusion model using only the 2D axial slices of the preprocessed data to learn the distribution characteristics of the 2D axial slices of the data, thereby obtaining a pre-trained unconditional diffusion model. It is important to note that the 2D axial slices of the data are large and need to be cropped into smaller square slices of equal size to form a small dataset for pre-training the unconditional generative diffusion model.

[0058] S3. The two-dimensional axial slices of the preprocessed data have orthogonal high-resolution and low-resolution directions. A theoretically estimated one-dimensional Gaussian blur kernel is used to perform one-dimensional blurring on the high-resolution direction of the two-dimensional axial slices of the preprocessed data. The high-resolution direction of the two-dimensional axial slices is then downsampled and interpolated along this direction to a pixel size consistent with the original axial pixel size. The two-dimensional axial slices are then upsampled and interpolated to a pixel size consistent with the lateral pixel size to obtain the two-dimensional axial slices of the processed data. The processed two-dimensional axial slices are used as conditional images and conditionally generated using the pre-trained unconditional diffusion model in step S2. The resolutions of the two orthogonal dimensions of the generated two-dimensional slice images are both consistent with the original axial low-resolution.

[0059] S4. Pair the two-dimensional axial slices of the preprocessed data in S1 with the two-dimensional low-resolution slices generated in step S3, and then use the paired data to train a neural network, with the latter as input and the former as output.

[0060] S5. Rotate the two-dimensional axial slices of the preprocessed data at several different angles at equal intervals, then input the rotated images into the neural network trained in S4 to obtain several corresponding output images. The output images are then reversely rotated by corresponding angles to obtain the final output image. The obtained output images are Fourier transformed, and their Fourier spectra are maximized for fusion. They are then inverse Fourier transformed to obtain isotropic resolution axial slices. It should be noted that the two-dimensional axial slices of the data are large in size and need to be cropped into smaller square slices of equal size. The smaller square slices are then input into the network. The output smaller square slices are reassembled into larger two-dimensional axial slices in the order in which they were cropped.

[0061] S6. Repeat the operation of step S5 for all axial slices in the XZ direction and the YZ direction, and then average the results in the two directions to finally obtain isotropic high-resolution three-dimensional fluorescence microscopy stack data.

[0062] Taking the isotropic resolution enhancement of three-dimensional stack data of confocal fluorescence microscopy system as an example, a detailed introduction is given.

[0063] After the data preprocessing in step S1, the unconditional generative diffusion model is trained using two-dimensional axial slices of the data:

[0064]

[0065] right Perform gradient descent optimization, and repeat the above process until convergence, where x0 is the two-dimensional axial slice of the preprocessed data; q(x0) is the distribution characteristics of the two-dimensional axial slice of the data; t is the number of time steps; Uniform() is a uniform distribution; T is the total number of time steps; noise ∈ is a standard Gaussian distribution with a variance of 1; ∈ θ A neural network for predicting noise distribution; is a coefficient and decreases as t increases.

[0066] Next, the theoretically estimated one-dimensional Gaussian blur kernel is used to perform one-dimensional blurring on the high-resolution direction of the two-dimensional axial slice of the preprocessed data. Then, the image is downsampled and interpolated along this direction until the pixel size is consistent with the original axial pixel size, and then upsampled and interpolated until the pixel size is consistent with the lateral pixel size. Specifically,

[0067] y=U(D(Blur(x0)))

[0068] Where x0 is the two-dimensional axial slice of the preprocessed data; Blur() is the theoretically estimated one-dimensional Gaussian blur kernel; D() is a one-dimensional downsampling operator, which downsamples and interpolates along a single dimension to a pixel size consistent with the original axial pixel size; U() is a one-dimensional upsampling operator, which upsamples and interpolates along a single dimension to a pixel size consistent with the horizontal pixel size; y is the processed two-dimensional axial slice.

[0069] Then, the pre-trained unconditional diffusion model in step S2 is used for conditional generation. The resolution of the two orthogonal dimensions of the generated two-dimensional slice image is consistent with the original axial low resolution, which can be specifically expressed as follows:

[0070] Given the conditional image y, sample x T ~N(0,I), the time step number t traverses T,…,1 from large to small, repeating the following process, and finally obtaining a new image x0, which is denoted as t0 for differentiation:

[0071] z~N(0,I)

[0072] x′ t-1 ~p θ (x′ t-1 ∣x t )

[0073] y′ t-1 ~q(y′ t-1 ∣y)

[0074] x t-1 ←(y′ t-1 +x′ t-1 ) / 2

[0075] Among them, the noise z obeys the standard Gaussian distribution with a variance of 1 to ensure the randomness of the generation process; x t is a two-dimensional image at time step t; p θ (∣) is the process of generating using the pre-trained unconditional generative diffusion model; x' t-1 is a two-dimensional image at time step t-1, generated by the pre-trained unconditional generative diffusion model; q(|) is a two-dimensional image after adding noise of time step t-1 to the conditional image y; x t-1 is the final two-dimensional image at time step t-1.

[0076] Next, the neural network in step S4 is trained. The training process can be described as follows:

[0077] x0=Θ(y0)

[0078] Where x0 is the two-dimensional axial slice of the preprocessed data, which serves as the output of the neural network Θ; y0 is the corresponding image conditionally generated in step (6), which serves as the input of the neural network Θ. The neural network Θ uses the RCAN neural network.

[0079] Finally, the multi-angle inference-Fourier maximum fusion process is performed.

[0080] The multi-angle inference process can be expressed as:

[0081] If there are M angles in total, for the i-th angle,

[0082] o i =iRot i (Θ(Rot i (x0)))

[0083] Where x0 is the two-dimensional axial slice of the preprocessed data; Rot i () is a rotation operator that rotates the two-dimensional image clockwise by angle i; Θ() is the neural network trained in step (7); iRot i () is a rotation operator, which rotates the two-dimensional image counterclockwise by angle i; o i is the output corresponding to the i-th angle.

[0084] The Fourier maximum fusion process can be expressed as:

[0085] o=ifft(Max(fft(o1),...,fft(o i ),...,fft(o M )))

[0086] Among them, o1 is the output corresponding to the first angle; o i is the output corresponding to the i-th angle; o M is the output corresponding to the Mth angle; fft() is the two-dimensional discrete fast Fourier transform; ifft() is the two-dimensional discrete inverse fast Fourier transform; Max() is the element-wise maximum operator, which takes the maximum value of all matrix elements affected by the operator at each position; o is the isotropic resolution axial slice of the output.

[0087] The following simulated data is used to verify the effectiveness of the present invention. A phantom with intensity distributions such as points, lines, shells, spheres, and cylinders is randomly generated. Anisotropic point spread functions are used for forward projection to simulate the excitation and acquisition process of three-dimensional fluorescence microscopy imaging, and a set of simulated data is obtained. The present invention is used to perform isotropic resolution enhancement on a stack of three-dimensional fluorescence microscopy images with anisotropic resolution, and the processed results are recorded ( Figure 2The simulation results show that the isotropic resolution enhancement method of the present invention has a good effect and does not rely on the strong assumption that the biological sample data has structural similarity in the axial and lateral directions.

[0088] In addition, the effectiveness of the present invention is further verified using thy1 mouse neural confocal microscopy imaging data. The maximum intensity projection images in the XY, XZ and YZ directions and the ROI images are compared before and after isotropic resolution enhancement using the present invention on thy1 mouse neural confocal microscopy imaging data ( Figure 3 ) It can be concluded that the present invention effectively improves the axial resolution of thy1 mouse neuronal confocal microscopy imaging data, reveals more high-frequency details, and achieves isotropic resolution enhancement.

[0089] The above description of the embodiments is intended to facilitate understanding and application of the present invention by those skilled in the art. It will be apparent that those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without requiring inventive effort. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention should fall within the scope of protection of the present invention.

Claims

1. A method for isotropic resolution enhancement in self-supervised three-dimensional fluorescence microscopy based on a diffusion model, characterized by: The steps include: 1) In a 3D fluorescence microscopy system, excitation light is used to excite fluorescently labeled biological samples, and data is collected through a detection objective. The process of acquiring a 3D image stack involves performing Z-axis layered scanning of the sample, combined with optical sectioning technology and high-precision mechanical control, to ultimately obtain 3D biological sample data; 2) obtaining the point spread function of the corresponding three-dimensional fluorescence microscopy imaging system through simulation or actual measurement, then normalizing the point spread function and calculating its back-projection operator; 3) performing axial interpolation processing on the three-dimensional biological sample data so that the axial pixel size is consistent with the lateral pixel size, and performing non-blind deconvolution preprocessing on the collected biological sample image data using the Richardson-Lucy deconvolution algorithm or the Wiener-Butterworth filter deconvolution algorithm to obtain a two-dimensional axial slice of the preprocessed data; 4) pre-training the unconditional generative diffusion model using only the two-dimensional axial slices of the pre-processed data to learn the distribution characteristics of the two-dimensional axial slices of the data and obtain a pre-trained unconditional diffusion model; 5) The two-dimensional axial slices of the preprocessed data in step 3) have orthogonal high-resolution and low-resolution directions. A theoretically estimated one-dimensional Gaussian blur kernel is used to perform one-dimensional blur processing on the high-resolution direction of the two-dimensional axial slices of the preprocessed data. Then, downsampling and interpolation are performed along this direction to a pixel size consistent with the original axial pixel size. Then, upsampling and interpolation are performed to a pixel size consistent with the lateral pixel size to obtain a two-dimensional axial slice of the processed data. 6) Using the two-dimensional axial slice processed in step 5) as a conditional image, conditionally generate the image using the pre-trained unconditional diffusion model in step 4), so that the resolution of the two orthogonal dimensions of the generated two-dimensional slice image is consistent with the original axial low resolution; 7) Pairing the two-dimensional axial slices of the preprocessed data in step 3) with the two-dimensional low-resolution slices generated in step 6), and then using the paired data to train a neural network, with the latter as input and the former as output; 8) rotating the two-dimensional axial slices of the preprocessed data in step 3) at equal intervals by several different angles, and then inputting the rotated images into the neural network trained in step 7) to obtain several corresponding output images, and then reversely rotating the output images by corresponding angles to obtain a final output image; 9) Fourier transforming the output images obtained in step 8) respectively, taking the maximum value of their Fourier spectra for fusion, and then performing inverse Fourier transform to obtain isotropic resolution axial slices; 10) Repeat step 9) for all axial slices in the XZ and YZ directions, and then average the results in the two directions to finally obtain isotropic high-resolution three-dimensional fluorescence microscopy stack data.

2. The method for isotropic resolution enhancement of three-dimensional fluorescence microscopy according to claim 1, wherein: The three-dimensional biological sample data collected in step 1) is a volume image, which is obtained by stacking two-dimensional slice images, and the axial resolution is lower than the lateral resolution.

3. The method for isotropic resolution enhancement of three-dimensional fluorescence microscopy according to claim 1, wherein: The point spread function in step 2) has three spatial dimensions and is a three-dimensional matrix structure.

4. The method for isotropic resolution enhancement of three-dimensional fluorescence microscopy according to claim 1, wherein: The pre-training process of the unconditional generative diffusion model in step 4) is as follows: x0~q(x0) t~Uniform({1,…,t}) right Perform gradient descent optimization, and repeat the above process until convergence, where x0 is the two-dimensional axial slice of the preprocessed data; q(x0) is the distribution characteristics of the two-dimensional axial slice of the data; t is the number of time steps; Uniform() is a uniform distribution; T is the total number of time steps; noise ∈ is a standard Gaussian distribution with a variance of 1; ∈ θ A neural network for predicting noise distribution; is a coefficient and decreases as t increases.

5. The method for isotropic resolution enhancement of three-dimensional fluorescence microscopy according to claim 1, wherein: In step 5), the one-dimensional Gaussian blur kernel estimated theoretically is used to perform one-dimensional blur processing on the high-resolution direction of the two-dimensional axial slice of the preprocessed data, and then downsamples and interpolates along this direction until the pixel size is consistent with the original axial pixel size, and then upsamples and interpolates until the pixel size is consistent with the lateral pixel size, specifically: y=U(D(Blur(x0))) Where x0 is the two-dimensional axial slice of the preprocessed data; Blur() is the theoretically estimated one-dimensional Gaussian blur kernel; D() is a one-dimensional downsampling operator, which downsamples and interpolates along a single dimension to a pixel size consistent with the original axial pixel size; U() is a one-dimensional upsampling operator, which upsamples and interpolates along a single dimension to a pixel size consistent with the horizontal pixel size; y is the processed two-dimensional axial slice.

6. The method for isotropic resolution enhancement of three-dimensional fluorescence microscopy according to claim 1, wherein: The step 6) can be specifically expressed as follows: Given the conditional image y, sample x T ~N(0,I), the time step number t traverses T,…,1 from large to small, repeating the following process, and finally obtaining a new image x0, which is denoted as y0 for differentiation: z~N(0,I); x′ t-1 ~p θ (x′ t-1 ∣x t ); and' t-1 ~q(y′ t-1 ∣y); x t-1 ←(y′ t-1 +x′ t-1 ) / 2; Among them, the noise z obeys the standard Gaussian distribution with a variance of 1 to ensure the randomness of the generation process; x t is a two-dimensional image at time step t; p θ (∣) is the process of generating using the pre-trained unconditional generative diffusion model; x' t-1 is a two-dimensional image at time step t-1, generated by the pre-trained unconditional generative diffusion model; q(|) is a two-dimensional image after adding noise of time step t-1 to the conditional image y; x t-1 is the final two-dimensional image at time step t-1.

7. The method for isotropic resolution enhancement of three-dimensional fluorescence microscopy according to claim 1, wherein: The neural network training process in step 7) is expressed as: x0=Θ(y0); Where x0 is the two-dimensional axial slice of the preprocessed data, which serves as the output of the neural network Θ; y0 is the corresponding image conditionally generated in step 6), which serves as the input of the neural network Θ.

8. The method for isotropic resolution enhancement of three-dimensional fluorescence microscopy according to claim 7, wherein: The neural network Θ in step 7) is a CNN-based two-dimensional image super-resolution neural network, or a Transformer-based two-dimensional image super-resolution neural network.

9. The method for isotropic resolution enhancement of three-dimensional fluorescence microscopy according to claim 1, wherein: The step 8) can be specifically expressed as: If there are M angles in total, for the i-th angle, o i =iRot i (Θ(Rot i (x0))) Where x0 is the two-dimensional axial slice of the preprocessed data; Rot i () is a rotation operator that rotates the two-dimensional image clockwise by angle i; Θ() is the neural network trained in step (7); iRot i () is a rotation operator, which rotates the two-dimensional image counterclockwise by angle i; o i is the output corresponding to the i-th angle.

10. The method for isotropic resolution enhancement of three-dimensional fluorescence microscopy according to claim 1, wherein: The step 9) can be specifically expressed as: o=ifft(Max(fft(o1),...,fft(o i ),...,fft(o M ))) Among them, o1 is the output corresponding to the first angle; o i is the output corresponding to the i-th angle; o M is the output corresponding to the Mth angle; fft() is the two-dimensional discrete fast Fourier transform; ifft() is the two-dimensional discrete inverse fast Fourier transform; Max() is the element-wise maximum operator, which takes the maximum value of all matrix elements affected by the operator at each position; o is the isotropic resolution axial slice of the output.

Images