Three-dimensional (3D) fluorescence microscopy provides crucial structural information about biological samples, surpassing the capabilities of 2D imaging. Advances in tissue-clearing techniques and light-sheet fluorescence microscopy (LSFM) have enabled high-speed, large-scale 3D visualization. However, achieving isotropic resolution (equal resolution in all directions) remains a challenge. Anisotropy, characterized by more blurriness in the axial plane, stems from factors like light diffraction, axial undersampling, and aberration correction. Even super-resolution techniques like 3D-SIM and STED microscopy struggle to match axial and lateral resolution. While LSFM improves axial resolution, truly isotropic point spread functions (PSFs) are difficult to achieve. Deep learning offers an alternative to classical deconvolution algorithms, capturing the statistical complexity of image mapping and enabling end-to-end transformation without manual parameter tuning. Existing deep-learning methods often require knowledge of the target data domain for network training, such as paired high-resolution and low-resolution images or an explicit PSF model. These assumptions limit flexibility and real-world applicability, particularly in high-throughput imaging where conditions fluctuate. Unsupervised learning using cycleGANs offers a promising approach, as it avoids the need for matched data pairs. However, obtaining high-resolution 3D reference volumes for training remains a challenge. This research addresses this by presenting an unsupervised deep learning framework that enhances axial resolution in volumetric fluorescence microscopy using only a single 3D input image, eliminating the need for additional high-resolution data.

Deep learning has emerged as a powerful tool for image restoration in fluorescence microscopy, offering advantages over traditional deconvolution methods. Several studies have used deep learning to improve resolution across different imaging modalities and numerical aperture sizes, and to achieve a degree of isotropy. However, most deep learning-based super-resolution methods for microscopy rely on supervised learning strategies, requiring paired high-resolution and low-resolution images for network training. This limitation makes them less practical for high-throughput imaging, where acquiring matched datasets is difficult. Some approaches use GANs, but often require a predefined image degradation model, which limits the method’s robustness. The use of unsupervised learning, specifically cycle-consistent GANs, has shown promise in solving ill-posed inverse problems in optics. This approach eliminates the need for paired training data, making it more practical for real-world applications. However, the challenge of obtaining high-resolution 3D reference data remains. This paper addresses this limitation by proposing a novel reference-free approach.

The proposed framework is based on an optimal transport-driven cycle-consistent generative adversarial network (OT-cycleGAN). Two 3D generative networks (G and F) learn to transform anisotropic 3D images to isotropic 3D images and vice versa. Two groups of 2D discriminative networks (Dx and Dy) guide the learning process. A key innovation is the sampling strategy of the discriminative networks. In the forward path (super-resolution), Dx compares 2D axial projections from the generated 3D image to 2D lateral images from the real 3D image. This encourages G to enhance only axial resolution. In the backward path (blurring), Dy compares 2D images from the reconstructed 3D image to 2D images from the real 3D image in each orthogonal plane, training F to revert the image restoration. Cycle consistency loss stabilizes the learning process, ensuring G and F are mutually inverse. The network is trained using a mini-max game, balancing the loss convergence of the generative and discriminative networks. The framework was tested on simulated data, confocal fluorescence microscopy (CFM) images of a mouse brain, and open-top light-sheet microscopy (OT-LSM) images. For simulation, a synthetic volume containing randomly placed and deformed tubular objects was created, then blurred axially. For CFM, a tissue-cleared mouse brain was imaged using optical sectioning. A second dataset was acquired after physically rotating the sample by 90 degrees to provide a high-resolution reference. For OT-LSM, 0.5-µm fluorescent beads and a tissue-cleared mouse brain were imaged. In the OT-LSM brain imaging experiment, a modified version of the framework was used with separate discriminators for the YZ and XZ planes due to non-identical image degradation across these planes. The generative networks are based on 3D U-Net architectures for most cases, and image volumes were processed in sub-regions with overlaps for large-scale inference. Image preprocessing included median filtering and normalization. Neuron tracing was performed using NeuroGPS-tree, followed by manual correction and verification against the reference image. Image quality was assessed using metrics such as PSNR, SSIM, MS-SSIM, and BRISQUE.

Simulation studies demonstrated the effectiveness of the proposed method in blind deconvolution, significantly improving PSNR and MS-SSIM. The method successfully resolved intricate structures masked by axial blurring, as quantified by improvements in PSNR, MS-SSIM, and FWHM mismatch. The method's performance was robust across varying levels of axial blurring and undersampling. In confocal fluorescence microscopy (CFM) experiments, the framework restored near-isotropic resolution in images of a mouse brain. The enhanced resolution allowed for more accurate reconstruction of neuronal morphology, demonstrated by comparing tracings from the enhanced images to a reference dataset acquired after physically rotating the sample. The quantitative analysis showed a mean PSNR improvement of 2.42 dB. In open-top light-sheet microscopy (OT-LSM) experiments using fluorescent beads, the framework demonstrated effective PSF deconvolution, reducing axial elongation and achieving near-isotropic resolution. This was confirmed by comparing FWHM values before and after deconvolution. Experiments on OT-LSM images of a mouse brain showed the framework's ability to correct various imaging artifacts beyond simple PSF blurring, such as image doubling and motion blur, resulting in significantly improved image quality and enabling more detailed reconstruction of neuronal structures. Comparison to Richardson-Lucy deconvolution highlighted the superior artifact correction capabilities of the deep learning approach. Experiments with artificially blurred ground truth datasets confirmed that the network is able to resolve axial blurring and correct other artifacts.

This research introduces a novel deep-learning-based super-resolution technique that significantly enhances axial resolution in volumetric fluorescence microscopy. The reference-free nature of the method, using only a single 3D input image for training, makes it highly practical for various microscopy applications. The use of an OT-cycleGAN architecture, along with a tailored sampling strategy for discriminative networks, enables the effective decoupling of resolution-relevant information from other image characteristics, enhancing the method’s robustness in the presence of variations in image degradation. The results from simulation, CFM, and OT-LSM experiments consistently demonstrate the method’s effectiveness in restoring suppressed details and correcting artifacts, leading to improved image quality and more accurate biological interpretations. The framework's ability to handle various types of image degradation and artifacts makes it suitable for various fluorescence microscopy modalities.

This study demonstrates a novel reference-free deep learning approach for isotropic super-resolution in volumetric fluorescence microscopy. The method successfully enhances axial resolution, restores suppressed details, and corrects imaging artifacts across different microscopy platforms. Its ease of use and broad applicability make it a valuable tool for advancing biological imaging research. Future work could explore optimizing the framework for specific microscopy techniques and expanding its application to other types of microscopy data. Investigating alternative network architectures and loss functions could further improve performance and robustness.

While the method demonstrates significant improvement in axial resolution and artifact correction, some minor visual artifacts may arise depending on the visualization method (2D vs. 3D) and intensity range. These artifacts are primarily due to the use of 2D MIPs in training the discriminative networks. Post-processing steps, such as local histogram matching, can mitigate these artifacts. The accuracy of neuronal tracings, while high, is still subject to potential errors in the automatic tracing algorithm and manual correction process. The ground truth data used for evaluation in certain experiments (such as the rotated CFM image) is not perfect due to variations in imaging conditions and potential registration errors. The results primarily showcase the method's effectiveness on particular biological samples. Further testing on diverse tissue types and imaging conditions is recommended to establish the method’s generalizability.