嵌 入 物 理 模 型 的 无 监 督 鲁 棒 超 分 辨 率 显 微 成 像

Objective Optical microscopy faces inherent challenges in resolving fine subcellular structures with high clarity due to the diffraction limit. While super^-resolution techniques such as stimulated emission depletion microscopy (STED), stochastic optical reconstruction microscopy (STORM), photo^-activated localization microscopy (PALM), and structured illumination microscopy (SIM) have surpassed this limit, their broad application remain constrained by system complexity, high cost, and limited imaging speed. In recent years, deep learning-based super^-resolution microscopy has become a valuable tool for visualizing biological microstructures. However, prevailing supervised approaches depend heavily on high-quality paired data, which substantially raises experimental difficulty and expense. Unsupervised methods, such as the cycle^-consistent generative adversarial network (CycleGAN), eliminate the need for paired data but are often limited by inadequate reconstruction accuracy and unstable output. To overcome these drawbacks, we propose an unsupervised, physics-embedded super^-resolution imaging framework named PE-CycleGAN. The proposed method improves reconstruction quality and generation stability using only a small set of training samples, while exhibiting strong noise robustness, high training efficiency, and good generalization ability. This approach offers an efficient and accessible solution for high^-resolution biological microscopy. Methods In this study, we employ PE-CycleGAN to train unsupervised super^-resolution models for both 2D and 3D imaging. PE-CycleGAN incorporates a learnable point spread function (PSF) model into the network in place of the original deep learning^-based generator, thereby simulating the physical degradation process from high^-resolution image to low^-resolution image. The framework trains a generator to reconstruct super^-resolution images directly from low^-resolution inputs. This design reduces model complexity and computational demand, through the integration of physical constraints, ensures that the mapping from the high^-resolution domain to the low^-resolution domain adheres more closely to actual imaging physics. Experiments utilize publicly available datasets: microtubule and endoplasmic reticulum images from BioSR, and zebrafish embryo membrane data from the Bio^-LFSR dataset. Using these sources, we construct low^-resolution to super^-resolution mapping models for both 2D and 3D settings, where wide^-field and light^-field microscopy images serve as low^-resolution inputs, while SIM and scanned light^-field microscopy images provide the corresponding super^-resolution references. Building on the conventional CycleGAN architecture, PE-CycleGAN integrates learnable 2D/3D adaptive PSFs to realistically model image degradation. Reconstructed image resolution is quantified using the ImDecorr method, and image quality is assessed via peak signal-to^-noise ratio (PSNR) and structural similarity index (SSIM). To evaluate robustness, varying noise levels are added to low^-resolution training images, training efficiency is examined by adjusting the size of the training set, and generalization capability is validated through cross-sample testing on synthetically generated microtubule and mitochondrial data. Finally, the practical applicability of the model is further verified under known PSF conditions in a 3D simulation scenario. Results and Discussions Experimental results demonstrate that PE-CycleGAN outperforms traditional CycleGAN in unsupervised super^-resolution microscopy image reconstruction tasks in terms of both reconstruction accuracy and generation stability. In both 2D and 3D models, PE-CycleGAN achieves approximately twice as much as CycleGAN in resolution while delivering higher image quality. Specifically, for zebrafish embryonic membrane samples, the resolution of images reconstructed by PE-CycleGAN reaches 0.583 μm, with an SSIM of 0.7276 and a PSNR of 28.19 dB (Table 2), whereas CycleGAN only achieves a resolution of 0.716 μm, an SSIM of 0.6798, and a PSNR of 26.73 dB. Studies on noise resistance reveal that noise degrades the quality of generated images, causing structural blurring and artifacts, while PE-CycleGAN effectively suppresses the impact of noise (Figs. 6 and 7). When the training sample size is reduced, traditional CycleGAN exhibits limited resolution improvement and unclear cellular structures, whereas PE-CycleGAN still recovers intact microscopic structures (Figs. 8 and 9). Cross^-sample tests further demonstrate that, leveraging constraints from the physical model, PE-CycleGAN exhibits excellent generalization performance on synthetic microtubule and mitochondrial samples. Even under known PSF conditions, PE-CycleGAN maintains advantages in PSNR and SSIM metrics (Table 6), highlighting the generalization capability and practicality of the physics^-embedded approach. Conclusions This study proposes a PE-CycleGAN for computational super^-resolution microscopy imaging. To address the unknown high-to^-low resolution mapping relationship, the model incorporates learnable 2D/3D adaptive PSFs to emulate the physical process of image degradation, enabling super^-resolution reconstruction from wide^-field and light^-field microscopy images via a generator. Compared to traditional CycleGAN, this method replaces part of the generator structure with a physics^-driven PSF, significantly reducing the number of parameters and decreasing training time and memory consumption. Systematic experiments on wide^-field^-to-SIM and light^-field^-to^-scanned^-light^-field mapping models show that PE-CycleGAN achieves resolution approximately twice as much as CycleGAN on both microtubule and zebrafish embryonic membrane samples, reaching 124.2 nm and 0.583 μm, respectively. Compared to CycleGAN, the proposed model exhibits substantial improvements in reconstruction quality, noise resistance, and training efficiency, maintaining cellular structural continuity and clarity even with limited samples or severe noise. Furthermore, the model demonstrates strong generalization capability, performing well on synthetic microtubule and mitochondrial data even when trained on zebrafish data. PE-CycleGAN retains advantages in PSNR and SSIM metrics even under known PSF conditions, underscoring the versatility and practicality of the physics^-embedded approach. In summary, by incorporating physical priors, PE-CycleGAN effectively enhances the precision, efficiency, and robustness of super^-resolution reconstruction, offering a novel strategy for microscopic image processing.