A self-supervised isotropic resolution enhancement method for three-dimensional fluorescence microscopy imaging

Abstract

This invention discloses a self-supervised method for isotropic resolution enhancement in three-dimensional fluorescence microscopy. It acquires a stack of three-dimensional fluorescence images with axial resolution lower than lateral resolution, and performs intensity normalization and pixel size alignment. Synthetic degradation data is constructed using lateral slices, and the isotropic restoration network IsoNet is pre-trained under supervision. Encoder-decoder networks IsoEnc and IsoDec are constructed to model the latent space of the anisotropic degradation process. During the real data inference stage, the parameters of IsoEnc and IsoDec are fixed to ensure that the restored result, after degradation reconstruction, is consistent with the input image, resulting in isotropic resolution enhancement. The axial slices are then recombined to obtain the three-dimensional restored image. This invention does not require precise point spread function modeling and does not rely on lateral and axial structural similarity assumptions. It effectively improves axial blurring and structural elongation in three-dimensional fluorescence microscopy images, achieving isotropic resolution enhancement. It is applicable to confocal, light-sheet, and three-dimensional structured light imaging systems.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional fluorescence microscopy and computational imaging technology, specifically relating to a self-supervised method for enhancing isotropic resolution in three-dimensional fluorescence microscopy. Background Technology

[0002] Three-dimensional fluorescence microscopy is an important tool in life science research, enabling three-dimensional visualization of cells, subcellular structures, and tissue samples, and providing fundamental data for the quantitative analysis of dynamic processes in living organisms. However, in actual imaging, due to the diffraction-limited characteristics of the microscopic optical system, the fact that the axial point spread function is usually significantly wider than the lateral point spread function, and the fact that the axial sampling step size is usually larger than the lateral pixel size, the resulting three-dimensional image stacks generally exhibit significant resolution anisotropy, meaning that the axial resolution is significantly lower than the lateral resolution. This problem directly causes elongation, blurring, and fragmentation of three-dimensional fine structures in the axial direction, which is detrimental to the accurate reconstruction of complex biological structures and subsequent applications such as segmentation, tracking, and morphological analysis.

[0003] To address the resolution anisotropy problem in three-dimensional fluorescence microscopy, existing technologies mainly fall into two categories: hardware improvement schemes and computational reconstruction schemes. Hardware improvement schemes typically enhance axial information acquisition capabilities through multi-view acquisition, interferometric illumination, dual-view joint reconstruction, or axial scanning enhancement. Although these methods can improve axial resolution to some extent, they generally suffer from drawbacks such as complex system structure, high alignment requirements, reduced acquisition speed, increased hardware costs, and cumbersome subsequent reconstruction procedures.

[0004] Computational restoration schemes attempt to improve axial imaging quality based on existing single-view data through image deconvolution or deep learning. For example, while point spread function-based deconvolution methods can improve image contrast and reduce defocus blur to some extent, the axial optical transfer function bandwidth of a single-view 3D fluorescence microscopy system is inherently narrower than that of the lateral optical transfer function. Axial high-frequency information is significantly compressed or even lost during acquisition. Therefore, conventional deconvolution methods can usually only restore the image within a limited bandwidth, making it difficult to reconstruct accurate axial high-frequency information and thus failing to fundamentally achieve isotropic resolution enhancement of 3D images. Supervised learning methods typically rely on manually constructed pairwise training data, and the distribution of training samples often differs from that of real test data [Weigert, 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]. Unsupervised methods based on adversarial networks [Ning, K., Lu, B., Wang, X. et al. Deep self-learning enables fast, high-fidelity isotropic resolution restoration for volumetric fluorescence microscopy. Light Sci Appl12, 204 (2023).] https: / / doi.org / 10.1038 / s41377-023-01230-2 While adversarial learning can alleviate the problem of insufficient paired data to some extent, it usually implicitly assumes that transverse slices and axial slices have high similarity in structural distribution. However, this assumption often does not hold true for biological samples with obvious directionality, complex branching, or layered structures. At the same time, adversarial learning itself also suffers from problems such as training instability, hyperparameter sensitivity, and susceptibility to pseudostructures [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 / s41467-022-30949-6Further semi-blind restoration methods attempt to transfer pre-trained models from natural images to microscopic images, but still face the problem of inconsistency between the microscopic imaging degradation model and the blurring degradation of natural images, resulting in insufficient axial detail recovery [Han, J., Liu, K., Isaacson, KB et al. System- and sample-agnostic isotropic three-dimensional microscopy by weakly physics-informed, domain-shift-resistant axialdeblurring. Nat Commun 16, 745 (2025). https: / / doi.org / 10.1038 / s41467-025-56078-4].

[0005] Therefore, current technologies still lack an isotropic resolution enhancement method for three-dimensional fluorescence microscopy that does not require complex hardware modifications, precise point spread function modeling, additional high-quality paired samples, or rely on the assumption of strong lateral and axial structural similarity. To address these issues, it is necessary to propose a novel method capable of directly learning degradation features from the measured data and performing adaptive recovery. Summary of the Invention

[0006] In view of the above-mentioned shortcomings in the existing technology, this invention provides a self-supervised method for isotropic resolution enhancement in three-dimensional fluorescence microscopy. This method employs a two-stage internal learning framework combining pre-training and adaptive inference to establish a reconstruction model from anisotropic three-dimensional microscopic images to isotropic three-dimensional microscopic images. Unlike traditional methods that rely on precise point spread function modeling or lateral-axial structural similarity assumptions, this invention explicitly decouples structural reconstruction from degradation modeling by constructing an isotropic reconstruction backbone network IsoNet and degradation latent space encoders / decoders IsoEnc and IsoDec. During the pre-training stage, stable isotropic structural representations and degradation characteristics are learned; during the inference stage, only the reconstruction backbone network is optimized within samples using in-sample self-supervised methods. This allows the model to directly adapt to the specific degradation characteristics of real-world acquired data, achieving high-fidelity axial resolution enhancement and three-dimensional isotropic reconstruction.

[0007] The method of the present invention includes the following steps:

[0008] This invention discloses a self-supervised method for isotropic resolution enhancement in three-dimensional fluorescence microscopy, comprising the following steps:

[0009] 1) In a three-dimensional fluorescence microscopy imaging system, fluorescently labeled biological samples are scanned layer by layer along the axis to obtain three-dimensional image stack data. The three-dimensional images have anisotropic characteristics where the axial resolution is lower than the lateral resolution. The three-dimensional fluorescence microscopy image stack data is preprocessed to obtain preprocessed three-dimensional image data.

[0010] 2) Extract horizontal slices from the preprocessed 3D image data, and construct synthetic degradation training data based on the horizontal slices;

[0011] 3) The isotropic restoration network IsoNet is pre-trained using synthetic degradation training data to learn the restoration mapping from anisotropic images to isotropic images. At the same time, encoder-decoder IsoEnc and IsoDec are constructed to model the latent space of the anisotropic degradation process, resulting in the pre-trained IsoNet, IsoEnc and IsoDec.

[0012] 4) Input the real-collected anisotropic axis data into the pre-trained IsoNet to obtain the initial recovery result. While keeping the IsoEnc and IsoDec parameters fixed, perform degradation reconstruction on the initial recovery result using IsoEnc and IsoDec. Then, perform self-supervised optimization on IsoNet based on the consistency between the degradation reconstruction result and the original input to obtain the self-supervised optimized IsoNet.

[0013] 5) The axial slices in the stacked data of three-dimensional fluorescence microscopy images were restored using IsoNet after self-supervised optimization, and the corresponding axial restoration results were obtained;

[0014] 6) The axial recovery results are reconstructed in three dimensions to obtain an isotropic resolution-enhanced three-dimensional fluorescence microscopy image. The method does not require precise point spread function modeling, additional paired training data, or structural similarity assumptions between transverse and axial slices. The method can be used to recover axial high-frequency information in three-dimensional biological sample images composed of point-like structures, linear structures, membrane structures, filamentous structures, branching structures, or combinations thereof, so as to improve the blurring, elongation, and breakage phenomena in the axial direction of three-dimensional biological sample images.

[0015] As a further improvement, the preprocessing in step 1) of the present invention includes intensity normalization, pixel size alignment and axial interpolation of the three-dimensional fluorescence microscopy image stack data, so that the axial sampling interval matches the lateral pixel size on a numerical scale and meets the input requirements for subsequent network training and inference.

[0016] As a further improvement, step 2) of the present invention, which involves constructing synthetic degradation training data based on transverse slices, specifically involves: establishing a one-dimensional fuzzy kernel library with different scales and directions; applying directional fuzziness to the transverse slices to simulate the axial degradation effect; and performing downsampling and upsampling processing on at least one spatial dimension after fuzzing to simulate the information loss caused by insufficient axial sampling.

[0017] As a further improvement, the one-dimensional blur kernel library described in this invention consists of multiple one-dimensional Gaussian blur kernels, each with a different standard deviation and rotation direction, used to generate pseudo-anisotropic input images with different degrees of degradation and different directions.

[0018] As a further improvement, in step 3) of the present invention, IsoNet is used to learn the restoration mapping from anisotropic image to isotropic image, IsoEnc is used to encode the anisotropic input image into low-dimensional latent features, and IsoDec is used to perform degradation reconstruction on the image based on the latent features to establish a latent degradation representation of the anisotropic observation data.

[0019] As a further improvement, the IsoNet described in this invention is implemented using a residual channel attention network, a convolutional neural network, a Transformer network, or a combination thereof; IsoEnc and IsoDec are implemented using an encoder-decoder structure for data-driven modeling of the anisotropic degradation process.

[0020] As a further improvement, the self-supervised optimization in step 4) of the present invention is achieved in the following way: the real anisotropic axial data collected is input into IsoNet to obtain the restoration result, and then the restoration result is input into IsoEnc and IsoDec with fixed parameters for degradation reconstruction. The difference between the degradation reconstruction result and the original input is used as the loss function to iteratively update the parameters of IsoNet so that IsoNet can adapt to the real degradation features of the current sample.

[0021] As a further improvement, the restoration of axial slices in the three-dimensional fluorescence microscopy image stack data in step 5) of the present invention specifically includes performing restoration processing on the XZ direction slices and YZ direction slices respectively, and rearranging the restored two-dimensional slices according to the original spatial order to form a three-dimensional restoration result.

[0022] As a further improvement, the present invention performs multi-angle rotation inference on each axial slice when performing restoration processing on XZ and YZ direction slices, and fuses the restoration results corresponding to different angles after reverse rotation, so as to reduce directional bias and improve the consistency of restoration results on different orientation structures.

[0023] As a further improvement, the fusion described in this invention adopts any one of frequency domain maximum value fusion, weighted average fusion, or other image fusion methods. The frequency domain maximum value fusion includes performing Fourier transform on the recovery results corresponding to different angles, taking the maximum value in the frequency domain, and then performing inverse Fourier transform to obtain the fused axial recovery result.

[0024] This invention discloses a self-supervised method for isotropic resolution enhancement in three-dimensional fluorescence microscopy. The method includes: acquiring a stack of three-dimensional fluorescence images with axial resolution lower than lateral resolution, and performing intensity normalization and pixel size alignment; constructing synthetic degradation data using lateral slices, and performing supervised pre-training on the isotropic restoration network IsoNet; constructing encoder-decoder networks IsoEnc and IsoDec to model the latent space of the anisotropic degradation process; during the real data inference stage, fixing the parameters of IsoEnc and IsoDec, and performing self-supervised optimization only on IsoNet to ensure that the restoration result, after degradation reconstruction, is consistent with the input image, thus obtaining the isotropic resolution enhancement result; and recombining the axial slices to obtain the three-dimensional restored image. This invention does not require precise point spread function modeling, additional paired data, or lateral-axial structural similarity assumptions. It can effectively improve axial blurring and structural elongation in three-dimensional fluorescence microscopy images, achieving isotropic resolution enhancement, and is applicable to confocal, light-sheet, and three-dimensional structured light imaging systems.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] First, this invention eliminates the need for precise point spread function (PSF) modeling, thus avoiding restoration errors caused by model adaptation. Traditional microscopic image restoration techniques largely rely on precise PSF construction and parameter fitting. This not only demands extremely high accuracy in PSF measurement and modeling but is also highly susceptible to various interference factors such as imaging equipment parameters, experimental environment, and sample optical characteristics, leading to model adaptation deviations and consequently various restoration errors such as image distortion, loss of detail, and pixel shift. In contrast, this invention abandons the traditional reliance on precise PSF modeling, eliminating the need for cumbersome function modeling and parameter calibration processes. It fundamentally avoids various restoration defects caused by model mismatch and insufficient modeling accuracy, significantly improving the accuracy and stability of image restoration results while effectively simplifying the imaging restoration process and lowering the technological implementation threshold.

[0027] Secondly, this invention does not rely on the assumption of lateral and axial structural similarity, thus making it more suitable for complex biological samples with significant differences in axial and lateral perspectives. Existing mainstream 3D microscopic image restoration algorithms are generally based on the prior assumption of high similarity in the lateral and axial structural features of the sample, and can only adapt to simple biological samples with regular structures and small differences in features across dimensions. However, in actual biological imaging scenarios, many complex biological samples, such as cell tissues, microbial structures, and biological vascular tissues, generally exhibit significant differences in axial and lateral morphology, texture, and structural features. The inherent assumptions of traditional algorithms become completely invalid, leading to a significant decrease in image restoration results or even restoration failure. This invention breaks through the prior constraint of structural similarity, does not rely on the precondition of consistent lateral and axial structures, and can perfectly adapt to complex biological samples with large differences in structures across dimensions, greatly expanding the applicable sample range of the technology.

[0028] Thirdly, this invention constrains the restoration result through a degradation modeling module to ensure it can interpret the original input, thereby effectively suppressing artifacts in unconstrained restoration. Traditional unconstrained image restoration techniques lack effective result constraint mechanisms during iterative optimization and image reconstruction, easily leading to over-optimization of image details and the generation of various artifacts such as false textures, edge noise, blurring, ghosting, and structural distortion. These artifacts severely interfere with the true information of the image, reduce imaging reliability, and greatly affect subsequent biological analysis results. This invention innovatively sets up a degradation modeling constraint module, forcibly constraining the final restored image result to accurately reverse-engineer and interpret the feature information of the original input image, eliminating unfounded autonomous optimization by the algorithm, and fundamentally curbing the generation of various artifacts, ensuring the authenticity, integrity, and reliability of the restored image.

[0029] Fourth, this invention is directly applicable to various three-dimensional fluorescence microscopy imaging modalities, exhibiting good versatility and scalability. Currently, most dedicated microscopic image restoration technologies are designed for a single imaging modality, with narrow algorithm architectures and parameter systems, making them incompatible with mainstream three-dimensional fluorescence imaging modalities such as confocal microscopy, wide-field microscopy, structured light microscopy, and two-photon microscopy. Different imaging devices and scenarios require separate development of adaptable algorithms, resulting in high R&D costs and poor reusability. This invention adopts a universal and modular technical architecture, eliminating the need for specific modifications and parameter fine-tuning for different imaging modalities. It can directly adapt to various mainstream three-dimensional fluorescence microscopy imaging devices and modes, while reserving expansion space to accommodate subsequent iterative upgrades of new imaging modalities. It possesses extremely strong engineering versatility, scenario adaptability, and technical scalability.

[0030] Fifth, this invention can recover richer axial high-frequency information, improve the continuity and clarity of filamentous, membrane-like, and branched structures in the axial direction, and further enhance the reliability of subsequent three-dimensional segmentation, tracking, and quantitative analysis. In traditional three-dimensional fluorescence microscopy, the axial high-frequency detail information of biological samples is easily lost due to the limitation of the imaging system's axial resolution, leading to problems such as axial discontinuity, blurred edges, structural breaks, and missing details in delicate biological structures such as filamentous fibers, biomembrane structures, and branched cell veins. This invention can efficiently mine and restore the axial high-frequency detail features of images, accurately repair the axial integrity of various delicate biological structures, and significantly improve the clarity and structural continuity of axial imaging. High-quality restored images can provide accurate, complete, and realistic image data support for subsequent downstream scientific research and engineering tasks such as three-dimensional image segmentation, dynamic target tracking, quantitative measurement of biological structures, and feature data analysis, effectively reducing analytical errors caused by image detail defects and comprehensively improving the accuracy, reliability, and scientific rigor of subsequent data analysis and research conclusions. Attached Figure Description

[0031] Figure 1 This is a schematic flowchart of the self-supervised three-dimensional fluorescence microscopy isotropic resolution enhancement method of the present invention;

[0032] Figure 2 This is a schematic diagram illustrating the isotropic resolution enhancement effect of the method of the present invention on simulated three-dimensional volume data. Detailed Implementation

[0033] To enable those skilled in the art to more clearly understand the technical solution of the present invention, the present invention will be further described in detail below with reference to specific embodiments. It should be understood that the following embodiments are only for illustrating the present invention and are not intended to limit the scope of protection of the present invention.

[0034] To describe the present invention in more detail, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] like Figure 1 As shown, the self-supervised three-dimensional fluorescence microscopy isotropic resolution enhancement method of the present invention includes the following steps:

[0036] S1. First, a three-dimensional fluorescence microscopy system (such as a light-sheet fluorescence microscopy system) is used to acquire volumetric data from fluorescently labeled biological samples. During acquisition, the sample is scanned layer by layer along the axis, and the microscopy system records two-dimensional images layer by layer and stacks them to form three-dimensional volumetric data. Since the axial point spread function of the microscopy system is usually wider than the lateral point spread function, and the axial step size is usually larger than the lateral pixel size, the resolution of the obtained three-dimensional volumetric data in the axial direction is significantly lower than that in the lateral direction. To facilitate subsequent network training and inference, the original three-dimensional volumetric data is first subjected to intensity normalization, and the axial sampling is interpolated and aligned to ensure that the axial pixel interval is consistent with the lateral pixel interval on a numerical scale. After this preprocessing, three-dimensional volumetric data suitable for network processing is obtained.

[0037] S2. In the pre-training phase, lateral slices are extracted from the three-dimensional volume data as relatively high-resolution reference images. Since lateral slices typically have a higher spatial resolution than axial slices, they can serve as a source of approximately isotropic structural representation. To construct anisotropic to isotropic training pairs, this invention first establishes a library of one-dimensional blur kernels with different orientations and scales. Preferably, one-dimensional Gaussian kernels with different standard deviations and rotation directions are used as the basic degradation kernels. Oriented blurring is applied to the lateral slices to simulate axial degradation effects of different degrees and directions. To further simulate the information loss caused by axial sampling sparsity in real data, downsampling can be performed on one spatial dimension after blurring, and then restored to the original size to obtain pseudo-anisotropic input images. The original lateral slices serve as the supervision targets, and the pseudo-anisotropic slices serve as the network input, thereby forming the paired samples required for pre-training.

[0038] S3. Using the training pairs described above, construct the isotropic restoration backbone network IsoNet and perform supervised pre-training. IsoNet's role is to learn a mapping relationship from the degraded image to the isotropic image, enabling the network to recover compressed axial high-frequency information from directionally blurred and undersampled images. In a preferred embodiment, IsoNet employs a residual channel attention network structure, enhancing its ability to model structural features and texture details through multi-level residual groups and attention modules. To improve the stability and fidelity of the restoration results, Charbonnier loss, L1 loss, or other pixel-level reconstruction losses can be used as supervision targets, ensuring the network output approximates the reference lateral image as closely as possible.

[0039] S4. While performing IsoNet pre-training, this invention further introduces the degradation latent space encoder-decoder networks IsoEnc and IsoDec. Specifically, IsoEnc receives pseudo-anisotropic image input and encodes it into a low-dimensional latent feature representation; IsoDec receives this latent feature and the image information to be degraded, and reconstructs the corresponding anisotropic image. By minimizing the difference between the reconstructed result and the pseudo-anisotropic input, IsoEnc and IsoDec can learn a set of stable latent space representations that can describe the distribution of anisotropic degradation observations. Thus, IsoEnc and IsoDec no longer rely on explicit point spread functions, but instead learn, through a data-driven approach, what kind of restoration result should be reverted to the input image after degradation. This design prevents the degradation process in this invention from being simply viewed as the action of a fixed convolutional kernel, but rather models it as a learnable degradation mapping modulated by latent features.

[0040] S5. After pre-training, the adaptive inference stage with real data begins. Unlike traditional single-pass forward inference, this invention keeps the IsoEnc and IsoDec parameters fixed in this stage, performing only in-sample self-supervised optimization of the IsoNet parameters. Specifically, the process is as follows: First, the real-collected anisotropic axial slices are input into IsoNet to obtain preliminary restoration results; then, the fixed IsoEnc and IsoDec are used to perform degradation reconstruction on the restoration results, obtaining a potential degradation output corresponding to the input image; finally, the difference between this degradation output and the real input image is used as the self-supervised loss to iteratively update the IsoNet parameters. As optimization progresses, IsoNet gradually sheds the general restoration bias learned solely from synthetic degradation data and becomes more adapted to the real degradation pattern of the current sample. Since this stage of optimization does not rely on any additional annotations and does not require an assumption of structural similarity between lateral and axial slices, this invention belongs to a self-supervised adaptation process based on internal learning. The IsoNet output updated through this process can then be used as the final isotropic restoration result for the current data.

[0041] S6. To obtain complete three-dimensional isotropic volumetric data, this invention can process axial slices in the XZ and YZ directions separately. For each slice, it can be fed into the network for reconstruction using a sliding window or block-based method to adapt to large-size images and limited video memory. The reconstructed two-dimensional axial slices are then reassembled into three-dimensional volumetric data in their original order. In some embodiments, to further improve the stability of the reconstruction results in different directions, multi-angle rotation inference can be performed on each axial slice. That is, the slices are first rotated by several preset angles and then input into the network to obtain reconstructed images. These reconstructed images are then rotated in the opposite direction and fused. The fusion method can employ frequency domain maximum fusion, weighted average fusion, or other fusion methods commonly used in the art. Through the above methods, directional bias can be further reduced, and the consistency of the reconstruction results across different oriented structures can be enhanced.

[0042] The technical effects of the present invention are illustrated below with reference to simulation embodiments. A three-dimensional simulated body membrane containing point-like, line-like, shell-like, spherical, and cylindrical structures is constructed, and input body data is generated using anisotropic imaging degradation processes. Some structures exhibit significant differences in appearance between lateral and axial views to simulate the dissimilarity between lateral and axial structures in real biological samples. After inputting the simulated anisotropic body data into the method of the present invention, a restored three-dimensional isotropic result can be obtained. Figure 2 This diagram illustrates the isotropic resolution enhancement effect of the method of this invention on simulated 3D volume data. The diagram includes the original anisotropic input data, the isotropic result restored using the method of this invention, and a comparison with a reference knitted surface. Observation of the YZ-direction slices and the maximum intensity projection map shows that this invention can effectively restore the structural extension caused by axial blur, improve the continuity and true shape of point-like, linear, shell-like, and cylindrical structures in the axial direction, and reduce the introduction of stripe artifacts and unrealistic details, thereby achieving high-fidelity 3D isotropic resolution enhancement. This invention can significantly improve axial blur and axial elongation phenomena, making the restored results closer to the real structure in both the YZ slices and the maximum intensity projection map. This indicates that this invention not only improves axial resolution but also does not rely on the assumption of structural similarity between the lateral and axial directions.

[0043] It should be noted that the above embodiments are merely preferred embodiments of the present invention. Those skilled in the art can make various modifications and substitutions to the network structure, loss function, preprocessing method, synthesis degradation strategy, block inference method, and multi-angle fusion strategy without departing from the concept of the present invention, and all such equivalent modifications and substitutions should fall within the protection scope of the present invention.

Claims

1. A self-supervised three-dimensional fluorescence microscopy isotropic resolution enhancement method, characterized in that, Includes the following steps: 1) In a three-dimensional fluorescence microscopy imaging system, fluorescently labeled biological samples are scanned layer by layer along the axis to obtain three-dimensional image stack data, wherein the three-dimensional images have anisotropic characteristics with axial resolution lower than lateral resolution, and the three-dimensional fluorescence microscopy image stack data is preprocessed to obtain preprocessed three-dimensional image data. 2) Extract horizontal slices from the preprocessed 3D image data, and construct synthetic degradation training data based on the horizontal slices; 3) The isotropic recovery network IsoNet is pre-trained using the synthetic degradation training data to learn the recovery mapping from anisotropic images to isotropic images. At the same time, encoder-decoder IsoEnc and IsoDec are constructed to model the latent space of the anisotropic degradation process, and the pre-trained IsoNet, IsoEnc and IsoDec are obtained. 4) Input the real-collected anisotropic axial data into the pre-trained IsoNet to obtain the initial recovery result. While keeping the IsoEnc and IsoDec parameters fixed, perform degradation reconstruction on the initial recovery result using the IsoEnc and IsoDec. Then, perform self-supervised optimization on IsoNet based on the consistency between the degradation reconstruction result and the original input to obtain the self-supervised optimized IsoNet. 5) The axial slices in the stacked data of the three-dimensional fluorescence microscopy image are restored using the self-supervised optimized IsoNet to obtain the corresponding axial restoration results; 6) Perform three-dimensional reconstruction on the axial recovery results to obtain an isotropic resolution-enhanced three-dimensional fluorescence microscopy image. The method does not require precise point spread function modeling, additional paired training data, or structural similarity assumptions between transverse and axial slices. The method can be used to recover axial high-frequency information in three-dimensional biological sample images composed of point-like structures, linear structures, membrane-like structures, filamentous structures, branching structures, or combinations thereof, so as to improve the blurring, elongation, and breakage phenomena of the three-dimensional biological sample images in the axial direction.

2. The self-supervised three-dimensional fluorescence microscopy isotropic resolution enhancement method according to claim 1, characterized in that: The preprocessing in step 1) includes intensity normalization, pixel size alignment, and axial interpolation of the stacked three-dimensional fluorescence microscopy image data to match the axial sampling interval with the lateral pixel size on a numerical scale and to meet the input requirements for subsequent network training and inference.

3. The self-supervised three-dimensional fluorescence microscopy isotropic resolution enhancement method according to claim 1, characterized in that: In step 2), constructing synthetic degradation training data based on the transverse slices specifically involves: establishing a one-dimensional fuzzy kernel library with different scales and directions; applying directional fuzziness to the transverse slices to simulate the axial degradation effect; and performing downsampling and upsampling processing on at least one spatial dimension after fuzzing to simulate the information loss caused by insufficient axial sampling.

4. The self-supervised three-dimensional fluorescence microscopy isotropic resolution enhancement method according to claim 3, characterized in that: The one-dimensional blur kernel library consists of multiple one-dimensional Gaussian blur kernels, each with a different standard deviation and rotation direction, used to generate pseudo-anisotropic input images with different degrees of degradation and different directions.

5. The self-supervised three-dimensional fluorescence microscopy isotropic resolution enhancement method according to claim 1, 2, 3, or 4, characterized in that: In step 3), IsoNet is used to learn the restoration mapping from anisotropic image to isotropic image, IsoEnc is used to encode the anisotropic input image into low-dimensional latent features, and IsoDec is used to perform degradation reconstruction on the image based on the latent features to establish a latent degradation representation of the anisotropic observation data.

6. The self-supervised three-dimensional fluorescence microscopy isotropic resolution enhancement method according to claim 5, characterized in that: The IsoNet is implemented using a residual channel attention network, a convolutional neural network, a Transformer network, or a combination thereof; the IsoEnc and IsoDec are implemented using an encoder-decoder structure and are used for data-driven modeling of the anisotropic degradation process.

7. The self-supervised three-dimensional fluorescence microscopy isotropic resolution enhancement method according to claim 6, characterized in that: The self-supervised optimization in step 4) is achieved as follows: the real anisotropic axial data is input into IsoNet to obtain the restoration result, and then the restoration result is input into IsoEnc and IsoDec with fixed parameters for degradation reconstruction. The difference between the degradation reconstruction result and the original input is used as the loss function to iteratively update the parameters of IsoNet so that IsoNet can adapt to the real degradation features of the current sample.

8. The self-supervised three-dimensional fluorescence microscopy isotropic resolution enhancement method according to claim 1, 2, 3, 4, 6, or 7, characterized in that: The restoration of the axial slices in the three-dimensional fluorescence microscopy image stack data in step 5) specifically includes performing restoration processing on the XZ direction slices and YZ direction slices respectively, and rearranging the restored two-dimensional slices according to the original spatial order to form a three-dimensional restoration result.

9. The self-supervised three-dimensional fluorescence microscopy isotropic resolution enhancement method according to claim 8, characterized in that: When performing restoration processing on the XZ and YZ direction slices, multi-angle rotation inference is performed on each axial slice, and the restoration results corresponding to different angles are fused after reverse rotation to reduce directional bias and improve the consistency of restoration results on different orientation structures.

10. The self-supervised three-dimensional fluorescence microscopy isotropic resolution enhancement method according to claim 9, characterized in that: The fusion adopts any one of frequency domain maximum value fusion, weighted average fusion, or other image fusion methods. Frequency domain maximum value fusion includes performing Fourier transform on the recovery results corresponding to different angles, taking the maximum value in the frequency domain, and then performing inverse Fourier transform to obtain the fused axial recovery result.

Images