Lightweight Deep Learning Framework for Astrocyte-Centric 3D Reconstruction across Normal and Stress-associated Ultrastructural States
Authors: Bowen Deng, Gwang-Jin Jeong, Seulgi Noh, Ji Young Mun, Na Young Kim
Organizations: Institute for Quantum Computing, Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, Canada · Department of Neuroscience, Korea Brain Research Institute, Daegu, Republic of Korea · Department of New Biology, Daegu Gyeongbuk Institute of Science and Technology, Daegu, Republic of Korea · Waterloo Institute for Nanotechnology, University of Waterloo, Waterloo, Canada
The morphology of astrocytes is closely associated with their ability to support and modulate neural circuit functions; however, its reconstruction from volume electron microscopy (vEM) data remains labor-intensive and computationally resource-intensive, placing substantial barriers on its broader adoption and scalability despite the high spatial resolution of vEM. We present AstroSeg-vEM, a lightweight, astrocyte-centric framework for morphology-preserving astrocyte segmentation, 3D reconstruction, and quantitative morphological analysis in vEM images. AstroSeg-vEM combines transformer-based 2D segmentation with dedicated post-processing methods and operates directly on high-resolution EM sections, making it well suited to anisotropic vEM data and downstream proofreading workflows while preserving the full in-plane resolution of the original EM data. Across four mouse hippocampal datasets representing normal and morphologically defined stress-associated astrocyte ultrastructural states, AstroSeg-vEM demonstrates robust performance and generalization, achieving average Dice and weighted Dice values of 0.92 and 0.95, respectively, on a held-out test dataset. Its lightweight architecture enables efficient, high-fidelity reconstruction on consumer-grade hardware, providing a scalable and accessible platform for large-scale astrocyte ultrastructure analysis and future astrocyte-neuron interaction studies.
Figures & tables
Figure 1: Visualization and statistics of astrocyte morphological profiles across vEM datasets. A Representative EM sections from astrocytes under stress (S-AUS) datasets (D1, D4) featuring complex, hyper-ramified morphologies and normal (N-AUS) datasets (D2, D3) exhibiting standard baseline morphologies. Red overlays indicate astrocyte masks. B Summary of dataset statistics and imaging parameters across four vEM datasets. Panel A overlays are downsampled for display from the native 5 × 5 nm 2 in-plane acquisition resolution to effective in-plane pixel (px) spacings of 58.6 nm/px (D1), 29.3 nm/px (D2), 48.8 nm/px (D3), and 68.4 nm/px (D4).
Figure 2: Lightweight 2D astrocyte segmentation and morphology-preserving reconstruction in AstroSeg-vEM. A Patch-level dataset generation leveraging foreground-aware sampling. B Hierarchical transformer-based U-Net with a MiT-B2 encoder capturing multi-scale morphological features through self-attention. C Overlapping sliding-window inference strategy ensuring in-plane spatial continuity via probabilistic aggregation (See Methods). D Morphology-preserving post-processing by sequentially performing soma-guided component refinement, volumetric structural consolidation, and in-plane morphological correction to filter spatially inconsistent artifacts and refine the connectivity of astrocytic processes across vEM sections. Full-section overviews and post-processing renders in A, C, and D are downsampled for display to an effective in-plane pixel (px) spacing of 29.3 nm/px.
Experiment
Objective
Training Set
Validation Set
Testing Set
Downstream Tasks
Exp 1
Cross-domain (S-AUS → N-AUS)
D1, D4 (first 80%)
D1, D4 (last 20%)
D2, D3
Segmentation Accuracy
Exp 2
Cross-domain (N-AUS → S-AUS)
D2, D3 (first 80%)
D2, D3 (last 20%)
D1, D4
Segmentation Accuracy
Exp 3
Full Pipeline & Morphology
D1, D2, D3 (first 80%)
D1, D2, D3 (last 20%)
D4
3D Reconstruction & Analysis
Table 1: Experimental design and dataset allocation
Architecture
Dice
wDice
Precision
Recall
Inference Time
Parameters
GFLOPs
(Exp 3)
(Exp 3)
(Exp 3)
(Exp 3)
per Section (s)
AstroSeg-vEM (Ours)
0.70 ± 0.02
0.94 ± 0.02
0.58 ± 0.03
0.89 ± 0.04
101.96
27.60M
10.80
SegFormer 39
0.68 ± 0.02
0.94 ± 0.03
0.55 ± 0.03
0.89 ± 0.04
110.23
24.72M
8.39
U-Net 28
0.36 ± 0.05
0.91 ± 0.02
0.23 ± 0.04
0.87 ± 0.04
207.37
34.53M
100.07
Swin-Unet 37
0.65 ± 0.06
0.90 ± 0.10
0.54 ± 0.03
0.84 ± 0.12
118.53
41.38M
17.82
Table 2: Segmentation Architectures Performance and Efficiency Comparison Dice, wDice, precision, and recall are mean ± standard deviation across n = 83 sections from a single held-out volume (Exp 3 D4), not independent biological replicates. Values are computed at the native 5 nm resolution and are not directly comparable with the 20 nm common-grid values of the 3D nnU-Net comparison (see Discussion).
Figure 3: Quantitative robustness and pixel-level morphological fidelity of the astrocyte segmentation pipeline. A Quantitative segmentation performance. Boxplots (median and interquartile ranges) and individual section scores (scatter) summarize metrics across three experiments. B Segmentation visual comparison. Representative pixel-level results for selected sections are collected across Exp 1–3. Yellow = true positives (overlap), red = false negatives (missed), and blue = false positives (over-segmentation). Sections span a range of Dice and wDice scores and morphological regions, not only the highest-scoring examples. Panel B overlays are downsampled for display from the native 5 × 5 nm 2 in-plane acquisition resolution to effective in-plane pixel (px) spacings of 58.6 nm/px (D1), 29.3 nm/px (D2), 48.8 nm/px (D3), and 68.4 nm/px (D4).
Post-processing
Exp 1 D2
Exp 1 D3
Exp 2 D1
Exp 2 D4
Exp 3 D4
SCR
VSC
IMC
Dice
wDice
Dice
wDice
Dice
wDice
Dice
wDice
Dice
wDice
0.57 ± 0.19
0.91 ± 0.06
0.50 ± 0.13
0.87 ± 0.05
0.38 ± 0.09
0.91 ± 0.05
0.71 ± 0.02
0.94 ± 0.02
0.70 ± 0.02
0.94 ± 0.02
✓
0.72 ± 0.09
0.90 ± 0.07
0.52 ± 0.12
0.87 ± 0.05
0.43 ± 0.10
0.89 ± 0.08
0.71 ± 0.03
0.92 ± 0.03
0.70 ± 0.02
0.93 ± 0.02
✓
✓
0.88 ± 0.07
0.90 ± 0.07
0.75 ± 0.10
0.86 ± 0.05
0.67 ± 0.12
0.81 ± 0.18
0.85 ± 0.05
0.87 ± 0.06
0.89 ± 0.02
0.93 ± 0.03
✓
✓
✓
0.90 ± 0.06
0.91 ± 0.06
0.76 ± 0.10
0.88 ± 0.05
0.68 ± 0.12
0.82 ± 0.18
0.87 ± 0.05
0.89 ± 0.06
0.92 ± 0.02
0.95 ± 0.02
Table 3: Ablation study of the post-processing pipeline for astrocyte segmentation Dice and wDice are mean ± standard deviation across sections within each held-out dataset (Exp 1 D2, n = 88; Exp 1 D3, n = 53; Exp 2 D1, n = 110; Exp 2 D4, n = 83; Exp 3 D4, n = 83), not independent biological replicates. All values are computed at the native 5 nm resolution and are not directly comparable with the 20 nm common-grid values of the 3D nnU-Net comparison (see Discussion).
Figure 4: Morphology-preserving 3D reconstruction (downsampled ten times and anisotropically stretched back to the original voxel scale) and quantitative analysis of astrocytes using AstroSeg-vEM. AstroSeg-vEM reconstruction (left) and the ground-truth masks (right) of N-AUS D3 from Exp 1 ( A ) and S-AUS D4 from Exp 3 ( B ). C Sholl plots (left) and histograms of mean intersections (right) between manual masks and AstroSeg-vEM-predicted masks of S-AUS D4 from Exp 3 for original, thinning, and manual methods. D Visualization of Sholl analysis results with AstroSeg-vEM (top) and manual (bottom) masks for three Fiji SNT plugin approaches. Voxel spacing after stretching back to native scale: 5 × 5 × 100 nm 3 (D3 in A, D4 in B).
Dataset
Number of Sections
Inference Time per Section (s)
Total Inference Time (s)
D1 S-AUS
110
96.18
12598.69
D2 N-AUS
88
23.63
2443.60
D3 N-AUS
53
66.55
4152.01
D4 S-AUS
83
102.47
10148.54
Table 4: Segmentation Runtime Performance on Exp 1 and Exp 2 Leveraging Transformer-based model Note: ‘Inference Time per Section’ is the wall-clock time for loading, inferring and saving a single section and ‘Total Inference Time’ is the wall-clock time of the whole evaluation run and additionally includes disk I/O and data-transfer overhead that is not attributed to individual sections.
Dataset
Number of Sections
Total Inference Time (s)
D1 S-AUS
110
520.81
D2 N-AUS
88
67.88
D3 N-AUS
53
111.47
D4 S-AUS
83
377.19
Table 5: Morphology-preserving Post-processing Run Time Performance on Exp 1 and Exp 2
Exp 1 D2
Exp 1 D3
Exp 2 D1
Exp 2 D4
Exp 3 D4
Metric
Median
IQR
Median
IQR
Median
IQR
Median
IQR
Median
IQR
Dice
0.91
0.07
0.77
0.11
0.70
0.23
0.90
0.08
0.92
0.02
wDice
0.92
0.07
0.88
0.07
0.94
0.36
0.93
0.10
0.96
0.02
Precision
0.97
0.03
0.77
0.19
0.70
0.17
0.95
0.03
0.93
0.06
Recall
0.86
0.13
0.79
0.11
0.89
0.50
0.87
0.16
0.93
0.05
Table 6: Quantitative robustness and pixel-level morphological fidelity of the astrocyte segmentation pipeline Values are median and IQR across sections within each held-out dataset (Exp 1 D2, n = 88; Exp 1 D3, n = 53; Exp 2 D1, n = 110; Exp 2 D4, n = 83; Exp 3 D4, n = 83), not independent biological replicates.
Figure S1 : Comparison of per-section segmentation quality before and after post-processing across the D1 and D4 volumes. A Per-section raw (before post-processing) and final (post-processed) wDice across the D1 volume. B Per-section raw and final wDice across the D4 volume, shown for comparison. C Overlays for the four post-processed outlier sections from D1. D Overlays for four representative sections from D4. Yellow = correctly predicted overlap, red = mask missed by the prediction, and blue = false positive prediction.
Figure S2 : Post-processing error decomposition into false discovery rate and miss rate across experiments and datasets. A False discovery rate (FDR = FP / (FP+TP) = 1 − precision) before and after post-processing for each experiment and test dataset. B Miss rate (FN / (FN+TP) = 1 − recall) before and after post-processing for each experiment and test dataset. The two rates have different denominators (predicted-positive and ground-truth-positive pixels) and are shown as complementary measures; bars show the mean across the sections of each held-out volume. TP, FP and FN are the numbers of true-positive, false-positive and false-negative pixels in a section.
Figure S3 : Paired, per-section comparison of AstroSeg-vEM and SegFormer Dice on Exp 3 D4. A Per-section paired Dice ( n = 83 sections, same D4 volume evaluated by both models). Points above the diagonal are sections where AstroSeg-vEM scores higher. B Distribution of per-section paired differences (mean paired difference +0.021, moving-block bootstrap 95% confidence interval 0.018–0.024 (block length 5–10 sections; 0.018–0.023 if serial dependence is ignored); positive in 81 of 83 sections; lag-1 autocorrelation of the paired differences 0.26, AR(1) effective sample size ≈ 49). This is a descriptive comparison within one volume; the sections are not independent replicates.
Raw
Post-processed
Protocol
Training
Validation
Test set
n
Dice
wDice
Dice
wDice
Exp 1
D1, D4 (first 80%)
D1, D4 (last 20%)
D2 (N-AUS)
88
0.57 ± 0.19
0.91 ± 0.06
0.90 ± 0.06
0.91 ± 0.06
D3 (N-AUS)
53
0.50 ± 0.13
0.87 ± 0.05
0.76 ± 0.10
0.88 ± 0.05
Exp 2
D2, D3 (first 80%)
D2, D3 (last 20%)
D1 (S-AUS)
110
0.38 ± 0.09
0.91 ± 0.05
0.68 ± 0.12
0.82 ± 0.18
D4 (S-AUS)
83
0.71 ± 0.02
0.94 ± 0.02
0.87 ± 0.05
0.89 ± 0.06
Exp 3
D1–D3 (first 80%)
D1–D3 (last 20%)
D4 (S-AUS)
83
0.70 ± 0.02
0.94 ± 0.02
0.92 ± 0.02
0.95 ± 0.02
Table S1 : Segmentation metrics before and after post-processing for all data-split protocols Exp 1–3 are the main-text experiments, trained on the first 80% and validated on the last 20% of the sections of the listed volumes, whereas the other protocols use fewer training volumes. Values are mean ± standard deviation across the sections of the held-out test volume. Post-processed values for 2-fold-1 D2 vary strongly across sections.
Section set
Connected components per section
Spurious components
Missed GT components
Split GT components
Merged predicted components
Ground truth
29.1 ± 2.7
–
–
–
–
Raw prediction
655.7 ± 39.1
599.9 ± 37.6
0.9 ± 1.0
10.0 ± 2.0
0.5 ± 0.7
Post-processed prediction
66.3 ± 17.7
30.3 ± 13.6
3.4 ± 2.7
5.0 ± 2.3
0.5 ± 0.7
Table S2 : Connected-component analysis of the Exp 3 D4 predictions before and after post-processing on the common 20 nm grid Spurious components overlap no ground truth, missed components are ground-truth components without an overlapping prediction, and split and merged counts are ground-truth components covered by several predictions and predictions covering several ground-truth components. Values are mean ± standard deviation across the 83 D4 sections.
2D visual foundation models, such as DINOv3, a self-supervised model trained on large-scale natural images, have demonstrated strong zero-shot generalization, capturing both rich global context and fine-grained structural cues. However, an analogous 3D foundation model for downstream volumetric neuroimaging remains lacking, largely due to the challenges of 3D image acquisition and the scarcity of high-quality annotations. To address this gap, we propose to adapt the 2D visual representations learned by DINOv3 to a 3D biomedical segmentation model, enabling more data-efficient and morphologically faithful neuronal reconstruction. Specifically, we design an inflation-based adaptation strategy that inflates 2D filters into 3D operators, preserving semantic priors from DINOv3 while adapting to 3D neuronal volume patches. In addition, we introduce a topology-aware skeleton loss to explicitly enforce structural fidelity of graph-based neuronal arbor reconstruction. Extensive experiments on four neuronal imaging datasets, including two from BigNeuron and two public datasets, NeuroFly and CWMBS, demonstrate consistent improvements in reconstruction accuracy over SoTA methods, with average gains of 2.9% in Entire Structure Average, 2.8% in Different Structure Average, and 3.8% in Percentage of Different Structure. Code: https://github.com/yy0007/NeurINO.
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
Volume Electron Microscopy (VEM) is crucial for 3D tissue imaging but often produces anisotropic data with poor axial resolution, hindering visualization and downstream analysis. Existing methods for isotropic reconstruction often suffer from neglecting abundant axial information and employing simple downsampling to simulate anisotropic data. To address these limitations, we propose VEMamba, an efficient framework for isotropic reconstruction. The core of VEMamba is a novel 3D Dependency Reordering paradigm, implemented via two key components: an Axial-Lateral Chunking Selective Scan Module (ALCSSM), which intelligently re-maps complex 3D spatial dependencies (both axial and lateral) into optimized 1D sequences for efficient Mamba-based modeling, explicitly enforcing axial-lateral consistency; and a Dynamic Weights Aggregation Module (DWAM) to adaptively aggregate these reordered sequence outputs for enhanced representational power. Furthermore, we introduce a realistic degradation simulation and then leverage Momentum Contrast (MoCo) to integrate this degradation-aware knowledge into the network for superior reconstruction. Extensive experiments on both simulated and real-world anisotropic VEM datasets demonstrate that VEMamba achieves highly competitive performance across various metrics while maintaining a lower computational footprint. The source code is available on GitHub: https://github.com/I2-Multimedia-Lab/VEMamba
Longmi Gao, Pan Gao
College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics Nanjing 211106, China