Source-linked AI summary
ZoomSEM: accelerating large-scale brain connectomics with intelligent SEM super-resolution
Jiateng Shou, Zhiwei Xiong, Ruikang Xu, Zhicheng Wang, Peiyao Shi, Te Shi, Haiqun Jin, Chunying Yin, Jun Guo, Manfred Auer, Hanchuan Peng, Ruobing Zhang, Feng Wu
TL;DR
Serial EM provides nanometer-scale views of brain circuits, but imaging large volumes is slow and existing super-resolution approaches have limitations in scale and realistic training data. ZoomSEM uses real paired SEM images and a recurrent multi-section architecture to reconstruct higher-resolution images from coarser scans. The reported speedups reach 29.2× at 64 nm imaging, and the method preserves fine structures across diverse samples.
Problem
Large-volume serial EM imaging is slow, while existing super-resolution methods are often limited to small upscaling factors or use synthetic training data that poorly reflects real acquisitions.
Source: Introduction · p. 1Method
ZoomSEM trains on real low- and high-resolution SEM pairs and uses recurrent propagation across serial sections with ADMA to align axial features.
Source: Workflow of using ZoomSEM for SEM… · p. 2, Discussion · p. 9Results
29.2× speedup is reported for 64 nm imaging combined with ZoomSEM, accounting for acquisition and inference time.
Source: Workflow of using ZoomSEM for SEM… · p. 2Takeaways & Limitations
ZoomSEM can bring conventional single-beam SEM imaging performance closer to costly multi-beam instruments and support large-volume EM applications.
Source: Introduction · p. 2, Discussion · p. 10The method is demonstrated on SEM data, and generalization from mouse entorhinal cortex to striatum remains inferior to a model fine-tuned on striatal data.
Source: Discussion · p. 10
Abstract
from arXiv · showhide
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$\times$, 8$\times$, and up to 16$\times$ 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.
Introduction
Serial EM can reveal synaptic-scale brain structure, but its slow acquisition limits large-volume imaging. ZoomSEM combines real-world paired SEM data with a multi-section network to reconstruct higher-resolution images from faster, coarser scans.
Introduction
- Cubic-millimeter brain imaging at nanometer resolution can take months or years, while reducing pixel dwell time raises throughput at the expense of signal-to-noise ratio.Mechanical overhead limits effective throughput of state-of-the-art multi-beam SEMs, despite their higher theoretical capture rate.Source: Introduction · p. 1
- Prior EM super-resolution methods often use small upscaling factors, synthetic degradation for large-factor training, or multi-frame optical flow that overlooks serial EM structural-scale variance.Earlier work also tended to prioritize algorithm design and small datasets over practical applicability to large-scale data.Source: Introduction · p. 1
- ZoomSEM supports 4×, 8×, and 16× lateral-resolution enhancement, using axial feature alignment and recurrent aggregation across sections to improve continuity and fine detail in elongated neurites and synapses.The ADMA module aligns axial features from adjacent sections, while the recurrent transformer aggregates aligned features across depth.Source: Introduction · p. 1
- RealSEM-Pair contains precisely registered low- and high-resolution brain SEM pairs that capture authentic noise, structural variation, and imaging artifacts from practical acquisitions.The dataset is presented as a resource for benchmarking and developing computational methods on real imaging conditions.Source: Introduction · pp. 1–2
- Nearly 30-fold shorter imaging duration is reported, with 16×16 coarser sampling yielding comparable quality to standard-resolution imaging at roughly 256× less acquisition time, neglecting overhead.The experiments also report accurate reconstruction of delicate brain structures at 16× super-resolution and generalization across samples, species, and a breast-cancer organoid system.Source: Introduction · p. 2
Results
ZoomSEM outperforms baseline methods across 4×, 8×, and 16× SEM super-resolution tasks, while real paired training data and its network design support high-fidelity reconstruction. It also accelerates acquisition and shows downstream segmentation and cross-sample and cross-species performance.
Results
- 10.0×, 21.0×, and 29.2× speedups accompany 16 nm, 32 nm, and 64 nm imaging, respectively, accounting for acquisition and inference time.Inference estimates use 8 A100 GPUs; resolution-independent workflow delays remain a substantial portion of imaging time.Source: Workflow of using ZoomSEM for SEM… · p. 2, Figure 1 · p. 3
- Across 4×, 8×, and 16× tasks, ZoomSEM consistently outperforms all baselines on PSNR, SSIM, and FSIM and recovers fine structures where other methods fail.At 16×, it recovers most cellular features despite minor errors, while other approaches largely fail.Source: ZoomSEM excels in SEM super-resolution… · p. 2, Figure 2 · p. 4
- Synthetic-degradation training performs worse in real super-resolution scenarios, while ZoomSEM markedly outperforms baselines trained on RealSEM-Pair.The dataset provides pixel-level aligned SEM volume pairs spanning 4 nm, 16 nm, 32 nm, and 64 nm; ZoomSEM uses recurrent bidirectional propagation and ADMA for structural scale variance.Source: Workflow of using ZoomSEM for SEM… · p. 2, Advancing SEM super-resolution with… · p. 3
- ZoomSEM shows higher semantic consistency with ground-truth pseudo-labels than other methods under both SAM- and Otsu-based segmentation evaluations.Segmentation was automated without manual refinement, and the authors note possible error in pseudo-labels and mask selection.Source: Accurate downstream segmentation… · pp. 3, 7, Figure 4 · p. 6
- Without fine-tuning, ZoomSEM trained on mouse entorhinal cortex data reveals cell boundaries and organelles in large-scale mouse striatum images.The evaluation used 898 serial sections after excluding two with poor imaging quality.Source: Robust cross-sample and cross-species… · p. 7
- On acini data, ZoomSEM outperforms scratch-trained CFAT and IART, while ZoomSEM-FT improves further, especially at 16× where other methods fail to preserve overall structure.The fine-tuned model used 100 registered sections; the remaining 100 were reserved for testing.Source: Robust cross-sample and cross-species… · p. 7
Ultrastructure recovery with ZoomSEM
ZoomSEM recovers multiple neuronal and cellular ultrastructures across super-resolution scales, and a 32 nm-to-4 nm reconstruction supports tracing an axon bouton–dendritic spine pair.
Ultrastructure recovery with ZoomSEM
- Synaptic clefts become distinguishable at all three scales, while synaptic vesicles are recovered at 4× and 8× but remain unresolved at 16×.At 16×, the visible cleft still leaves the synapse recognizable.Source: Ultrastructure recovery with ZoomSEM · p. 7, Figure 6 · p. 9
- The ultra-thin dendritic spine neck is recovered at 4×, 8×, and 16× super-resolution.Manual tracing on a 32 nm-to-4 nm volume reconstructed an axon bouton–dendritic spine pair and short segments of the axon terminal and dendrite.Source: Ultrastructure recovery with ZoomSEM · p. 7, Figure 6 · p. 9
- Mitochondrial cristae are faithfully recovered from 32 nm to 4 nm super-resolution.Source: Ultrastructure recovery with ZoomSEM · p. 7, Figure 6 · p. 9
- ZoomSEM-FT reconstructs larger microvilli across scales, but finer structures remain unresolved at 64 nm-to-4 nm super-resolution.Finer microvillar structures are recovered at 16 nm-to-4 nm and 32 nm-to-4 nm.Source: Ultrastructure recovery with ZoomSEM · p. 7, Figure 6 · p. 9
- Manual tracing of a 32 nm-to-4 nm super-resolved volume produced a 3D reconstruction of an axon bouton–dendritic spine pair.Short segments of the axon terminal and dendrite were also traced.Source: Ultrastructure recovery with ZoomSEM · p. 7, Figure 6 · p. 9
Discussion
Serial EM's throughput and multi-beam costs constrain large-scale connectomics. ZoomSEM reconstructs high-resolution SEM images from faster, coarser scans, with reported speedups and preserved biological detail that support downstream analyses.
Discussion
- Approximately 10 to nearly 30 times acquisition speedup is reported for reconstructions from initial resolutions up to 16× coarser.The method combines real-world paired data with a transformer-based architecture.Source: Discussion · p. 9
- ZoomSEM is presented as bringing standard single-beam SEM performance toward costly multi-beam instruments and broadening access to high-throughput volumetric EM.The authors conclude that this could improve the accessibility and feasibility of ambitious connectomics and ultrastructural biology studies.Source: Discussion · p. 10
- Synapses, synaptic vesicles, spine necks, mitochondria, microvilli, and organelle morphologies are preserved across species and tissue types.The authors describe this as evidence of robust cross-domain generalization.Source: Discussion · p. 10
- ZoomSEM achieves higher consistency with ground-truth segmentation than baseline methods, particularly at larger upscaling factors.The authors state that preserved ultrastructural features can support reliable instance interpretation when other approaches fail to recover recognizable structures.Source: Discussion · p. 10
- The method is demonstrated on SEM data, and domain gaps, unhandled artifacts, and very low-signal scans constrain its current scope.Mouse entorhinal-cortex training generalizes reasonably to mouse striatum, but performs worse than a model fine-tuned on striatal data.Source: Discussion · p. 10