Organizations: MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, China. · Anhui Province Key Laboratory of Biomedical Imaging and Intelligent Processing, Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, Hefei, China. · Institute for Brain and Intelligence, Fudan University, Shanghai, China. · Department of Biomedical Engineering, School of Biological Sciences and Medical Engineering, Southeast University, Nanjing, China.
Serial electron microscopy (EM) is essential for large-volume biological imaging and connectomics, yet its throughput remains a major bottleneck, with cubic-millimeter-scale brain imaging requiring months or years. Existing deep learning-based EM super-resolution methods are limited to small upscaling factors or rely on synthetic data, offering modest acceleration and limited practical impact. We introduce ZoomSEM, a deep learning-based super-resolution method to dramatically accelerate SEM imaging. Trained on real low- and high-resolution scanning electron microscopy (SEM) pairs, ZoomSEM enables reliable recovery of 4×, 8×, and up to 16× higher-resolution images from fast low-resolution scans while preserving fine neuronal structures critical for connectomic analysis, achieving up to 30-fold effective acceleration. Extensive experiments across datasets of diverse biological samples demonstrate robust generalization under realistic low-resolution acquisition. This advance enables high-resolution, large-volume serial EM imaging at unprecedented speed, poised to transform connectomics and related fields.
Figures & tables
Figure 1: Overview of the ZoomSEM network and speed advantages. a , Workflow of SEM super-resolution using ZoomSEM. Given a low-resolution SEM volume, the ZoomSEM network performs 4 × , 8 × , or 16 × super-resolution without requiring high-resolution acquisition. b , Training pipeline and schematic structure of the ZoomSEM network. c , Estimated improvement in imaging speed when using ZoomSEM. Considering both acquisition and inference time, the overall speedup is 10.0-fold from 16 nm to 4 nm, 21.0-fold from 32 nm to 4 nm, and 29.2-fold from 64 nm to 4 nm. Time estimates are based on a 60 μ m × 60 μ m imaging area, corresponding to an image resolution of 15000 × 15000 pixels at 4 nm voxel size. The inference time is measured using 8 A100 GPUs.
Figure 2: ZoomSEM achieves high-fidelity performance across various upscaling factors. a,d,g , Quantitative evaluation of ZoomSEM, CFAT [ 23 ] , IART [ 22 ] , EMDiffuse [ 14 ] , and bicubic-interpolated inputs (Bic) for super-resolution tasks from 16 nm to 4 nm (4 × ), from 32 nm to 4 nm (8 × ), and from 64 nm to 4 nm (16 × ), assessed by PSNR, SSIM [ 24 ] , and FSIM [ 25 ] . All models are trained under their original training-data pipelines. b,e,h , Qualitative comparison for super-resolution tasks from 16 nm to 4 nm (4 × ), from 32 nm to 4 nm (8 × ), and from 64 nm to 4 nm (16 × ). Arrows in b highlight fine structural details that are restored by our method but not by other methods. From 16 nm to 4 nm, ZoomSEM closely matches the high-resolution ground truth (GT). From 32 nm to 4 nm, ZoomSEM shows a pronounced advantage, successfully resolving fine structures that other methods fail to reconstruct while maintaining high fidelity. From 64 nm to 4 nm, other methods largely fail to recover meaningful content. Despite minor errors in the internal cellular structures, ZoomSEM recovers the overall cell morphology, even when the input retains almost no discernible details. c,f,i , pixel-wise intensity error comparison along the blue lines in panels b, e, and h. ZoomSEM is compared with baseline methods by measuring the pixel intensity differences along the blue lines, with GT as reference.
Figure 3: Training on the RealSEM-Pair dataset and employing the ZoomSEM network substantially improve EM super-resolution performance. a , Schematic of the RealSEM-Pair dataset creation pipeline, including data preparation and processing. b , Samples from the RealSEM-Pair dataset, showing pixel-level alignment of 3D SEM volumes at different resolutions, which demonstrates the feasibility of our registration strategy. c-e , Quantitative comparisons from 16 nm to 4 nm (4 × ), from 32 nm to 4 nm (8 × ), and from 64 nm to 4 nm (16 × ) super-resolution on the same models trained with the RealSEM-Pair dataset, the bicubic downsampling (BD) pipeline (synthetic), and the high-order downsampling pipeline (HOD) (synthetic) [ 27 ] . Bicubic interpolation (Bic) is also provided for reference. The models used here are CFAT [ 23 ] and IART [ 22 ] , evaluated by PSNR, SSIM [ 24 ] , and FSIM [ 25 ] . f-h , Quantitative comparisons from 16 nm to 4 nm (4 × ), from 32 nm to 4 nm (8 × ), and from 64 nm to 4 nm (16 × ) super-resolution under training on the RealSEM-Pair dataset, evaluating ZoomSEM, CFAT+ [ 23 ] , IART+ [ 22 ] , EMDiffuse [ 14 ] , and Bic. “+” indicates that the model is trained on the RealSEM-Pair dataset, discarding its original synthetic data training pipeline. Since EMDiffuse is already trained on the RealSEM-Pair dataset, EMDiffuse+ is identical to EMDiffuse. i , Qualitative comparison from 32 nm to 4 nm (8 × ) super-resolution. Other methods exhibit noticeable artifacts, whereas ZoomSEM demonstrates superior fidelity. j , Qualitative comparison from 64 nm to 4 nm (16 × ) super-resolution. At this extreme upscaling factor, while other methods largely fail, ZoomSEM recovers the overall cell morphology. Arrows in panel i highlight fine structural details that are restored by our method but not by other methods.
Figure 4: Accurate downstream segmentation of ZoomSEM demonstrated via SAM- [ 28 ] and Otsu-based [ 29 ] segmentation. a , Quantitative evaluation of ZoomSEM, CFAT [ 23 ] , IART [ 22 ] , EMDiffuse [ 14 ] , and bicubic-interpolated inputs (Bic) using SAM-based segmentation against SAM-generated pseudo-labels from the ground truth (GT). The results from 16 nm to 4 nm (4 × ), from 32 nm to 4 nm (8 × ), and from 64 nm to 4 nm (16 × ) super-resolution tasks are assessed by VOI [ 30 ] . b , Quantitative evaluation of the same models using Otsu-based segmentation against Otsu-generated pseudo-labels from the GT, across the same three EM super-resolution settings, assessed by ACC. c-d , Qualitative comparison of SAM- and Otsu-based segmentations from 32 nm to 4 nm (8 × ) super-resolution task. ZoomSEM produces a clearer, more complete cellular structure with significantly higher semantic consistency than other methods. Arrows in all panels (c-d) indicate fine structural details that are accurately super-resolved and segmented by our method but are either missing or misrepresented in the results of other methods. In panel c, the SAM result is further refined by supplying an identical foreground prompt to every input patch, compensating for slight segmentation error in GT. e-f , Qualitative comparison of SAM- and Otsu-based segmentations from 64 nm to 4 nm (16 × ) super-resolution task. At this extreme scale, other methods fail to recover reliable structures, while ZoomSEM reconstructs coherent cellular morphology, yielding segmented results close to the GT. All segmentation was performed fully automatically without manual correction.
Figure 5: ZoomSEM demonstrates excellent generalization performance. a , ZoomSEM super-resolves images from 32 nm to 4 nm (8 × ). The network is trained on the RealSEM-Pair dataset (sampled from mouse entorhinal cortex) and evaluated on large-scale mouse striatum data. The comparison shows the low-resolution input and the high-resolution output. b , Qualitative comparison of Otsu-based segmentation results between ZoomSEM and bicubic-interpolated inputs (Bic) on large-scale mouse striatum data from 32 nm to 4 nm (8 × ). ZoomSEM can super-resolve clear cell boundaries and intracellular organelle structures without any fine-tuning. c-d , Quantitative evaluation of ZoomSEM-FT, ZoomSEM, CFAT [ 23 ] , IART [ 22 ] , and Bic on the acini dataset from 32 nm to 4 nm (8 × ) and from 64 nm to 4 nm (16 × ) super-resolution tasks, measured by PSNR, SSIM [ 24 ] , and VOI [ 30 ] . ZoomSEM-FT denotes the network fine-tuned on the acini dataset, while the other models are trained from scratch on the acini dataset. e , Qualitative comparison from 32 nm to 4 nm (8 × ) super-resolution task and SAM-based segmentation on the acini dataset. ZoomSEM-FT outperforms ZoomSEM in both fidelity and segmentation consistency relative to the ground truth (GT), while ZoomSEM notably surpasses CFAT and IART. f , Qualitative comparison from 64 nm to 4 nm (16 × ) super-resolution task on the acini dataset. Under this more challenging setting, all other methods, including ZoomSEM, fail to preserve the overall structure, whereas ZoomSEM-FT achieves a marked improvement in fidelity and segmentation consistency with respect to the GT. Arrows in all panels (b, e-f) indicate fine structural details that are accurately super-resolved and segmented by our method but are either missing or misrepresented in the results of other methods.
Figure 6: ZoomSEM facilitates ultrastructure recovery. a , Synapses under three EM super-resolution settings: from 16 nm to 4 nm, from 32 nm to 4 nm, and from 64 nm to 4 nm. Comparing ZoomSEM to input, synaptic clefts transform from blurred or imperceptible to clearly defined and detailed from 16 nm to 4 nm, become distinguishable again from 32 nm to 4 nm (after being virtually lost), and emerge from near-complete absence at from 64 nm to 4 nm to reveal a recognizable synaptic outline. b , Synaptic vesicles at the same three EM super-resolution scales. ZoomSEM faithfully reconstructs vesicles (yellow arrows) when super-resolving from 16 nm and 32 nm to 4 nm. Although vesicles are unresolved in the 64 nm to 4 nm setting, the synaptic cleft is still clearly visible (between red arrows), so the synapse remains easy to recognize. c , Spine neck at three EM super-resolution scales. The ultra-thin neck of the dendritic spine is successfully recovered in all three super-resolution settings. d , Mitochondria at three EM super-resolution scales. ZoomSEM faithfully recovers mitochondrial cristae from 32 nm to 4 nm super-resolution. e , Microvilli at three EM super-resolution scales using ZoomSEM-FT. Larger microvilli indicated by red arrows are accurately reconstructed across all scales, while finer microvillar structures (green arrows) are faithfully recovered from 16 nm to 4 nm and from 32 nm to 4 nm, but remain unresolved in the 64 nm to 4 nm setting. f , Manual tracing on the 32 nm to 4 nm super-resolved volume reconstructs an axon bouton-dendritic spine pair, and short segments of the axon terminal and dendrite are further traced. Left, a representative section with hand-drawn annotations. Right, the resulting 3D tracing.
Recent advances in high-resolution fluorescence and electron microscopy have enabled nanoscale imaging across increasingly large brain volumes, but the resulting terabyte- to petabyte-scale datasets make complete neuronal reconstruction prohibitively expensive in computation, data movement, and manual proofreading. Here, we present SkNeXt, a topology-first framework for scalable neuronal reconstruction from large volumetric microscopy datasets. Instead of densely processing entire image volumes, SkNeXt first converts neuronal morphology into compact SWC skeletons that preserve long-range connectivity. Proofreading is therefore focused on sparse neuronal trees, allowing branch, continuity, and connectivity errors to be corrected before high-resolution reconstruction. The corrected skeletons then serve as persistent structural priors for recovering detailed morphology while preserving neuronal identity and topology. Crucially, SkNeXt also uses neuronal skeletons as spatial indices for selective data access, retrieving high-resolution image regions only along reconstructed trajectories and bypassing most background and signal-free volumes. This substantially reduces I/O and computational overhead, allowing reconstruction cost to scale with neuronal morphology rather than total dataset size. Using SkNeXt, we reconstructed neurons from a petabyte-scale super-resolution fluorescence dataset of the mouse brain on a single GPU within one week, without requiring exhaustive dense inference across the complete imaging volume.
Jiayi Ding, Hu Zhao
Chinese Institute for Brain Research, Beijing, China
Vision foundation models have substantially advanced computer vision, enabling state-of-the-art performance in zero- and few-shot settings. They have been successfully applied to biomedical imaging tasks ranging from organ segmentation in computed tomography to cell segmentation in light microscopy. Electron microscopy (EM) is a central modality for analyzing cellular ultrastructure due to its nanometer-scale resolution. However, the application of foundation models in EM has so far been limited to specific organelles, such as mitochondria, largely due to the diversity of segmentation tasks and the scarcity of comprehensively annotated data. As a result, EM segmentation still predominantly relies on supervised learning, requiring extensive manual annotation and limiting ultrastructural analysis. To address this gap, we propose μMatch, a framework for semi-supervised learning and domain adaptation that leverages foundation models. We implement state-of-the-art student-teacher-based methods and evaluate multiple foundation models (SAM, SAM2, μSAM, DINOv2/v3) on challenging EM tasks, including mitochondrion, nucleus, and neurite segmentation. Our results demonstrate consistent improvements over strong baselines and highlight a path toward substantially reducing the annotation effort in EM.
Scanning Electron Microscopy (SEM) is indispensable in modern materials science, enabling high-resolution imaging across a wide range of structural, chemical, and functional investigations. However, SEM imaging remains constrained by task-specific models and labor-intensive acquisition processes that limit its scalability across diverse applications. Here, we introduce the first foundation model for SEM images, pretrained on a large corpus of multi-instrument, multi-condition scientific micrographs, enabling generalization across diverse material systems and imaging conditions. Leveraging a self-supervised transformer architecture, our model learns rich and transferable representations that can be fine-tuned or adapted to a wide range of downstream tasks. As a compelling demonstration, we focus on defocus-to-focus image translation-an essential yet underexplored challenge in automated microscopy pipelines. Our method not only restores focused detail from defocused inputs without paired supervision but also outperforms state-of-the-art techniques across multiple evaluation metrics. This work lays the groundwork for a new class of adaptable SEM models, accelerating materials discovery by bridging foundational representation learning with real-world imaging needs.
Sk Miraj Ahmed, Yuewei Lin, Chuntian Cao +9
Computing and Data Sciences, Brookhaven National Laboratory, Upton, 11973, New York, USA. · Center for Functional Nanomaterials, Brookhaven National Laboratory, Upton, 11973, New York, USA. · Department of Material Sciences and Engineering, The University of Texas at Dallas, Richardson, 75080, Texas, USA.