AxonSynth: Domain-Randomized Synthetic Data for Zero-Shot 3D Axon Segmentation in Light-Sheet Microscopy

Edward Gaibor; Kyriaki-Margarita Bintsi; Chiara Mauri; Carmen Luz Leiva Ureta; Zayneb Bellatif; Chiara Maffei; Wenze Li; Elizabeth Hillman

cs.CVSep 25, 2026

AxonSynth: Domain-Randomized Synthetic Data for Zero-Shot 3D Axon Segmentation in Light-Sheet Microscopy

Authors: Edward Gaibor, Kyriaki-Margarita Bintsi, Chiara Mauri, Carmen Luz Leiva Ureta, Zayneb Bellatif, Chiara Maffei, Wenze Li, Elizabeth Hillman, +2 more

Organizations: Department of Computer Science, University of Massachusetts Boston, Boston, MA, USA · Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School, Charlestown, MA, USA · Universidad San Sebastián, Chile · Université Claude Bernard Lyon 1, Université de Lyon, Lyon, France · Department of Imaging Sciences, St. Jude Children’s Research Hospital, Memphis, TN, USA · Department of Biomedical Engineering, Columbia University in the City of New York, New York, NY, USA · Department of Experimental Psychology, University College London, London, United Kingdom

Abstract

Accurate segmentation of axons in 3D microscopy data is important for analyzing white-matter organization, but dense ground truth labels are expensive to obtain. Existing supervised axon segmentation methods rely on target-domain annotations and can be brittle when tissue type, species, modality, or acquisition conditions change. We present AxonSynth, a domain-randomized synthetic-data framework for training 3D axon segmentation models without manually annotated real training volumes. AxonSynth generates dense synthetic axon labels with orientation priors that reflect realistic fiber configurations and renders them with randomized density, contrast, bias fields, blur, and noise. A three-class 3D U-Net is trained to predict background, axon sheath and intra-axonal space. We evaluate zero-shot transfer on 10 held-out light-sheet microscopy (LSM) patches from macaque and human brain samples labeled with one of three axonal markers, comparing against calibrated thresholding and Frangi filtering using overlap, corrected detection, false-positive, and topology metrics. On macaque samples, AxonSynth achieved the best corrected Dice and corrected precision (0.826 and 0.851), compared with 0.765 and 0.754 for thresholding and 0.685 and 0.762 for Frangi. On human samples, corrected Dice was comparable to thresholding (0.857 vs. 0.868), while component-count error decreased from 22,504 to 3,377. Across all held-out patches, AxonSynth reduced component-count error in 10/10 patches and Euler-characteristic error in 8/10. These results show that synthetic-label domain randomization can reduce dependence on manual axon annotation while supporting synthetic-to-real 3D segmentation.

Figures & tables

Figure 1: AxonSynth training pipeline. Dense synthetic 3D axon instance labels are sampled as $128^{3}$ patches and converted into three-class supervision (background, axon sheath, intra-axonal space). A stochastic renderer applies domain randomization to generate diverse synthetic image–label pairs. A 3D U-Net is trained only on synthetic data; at inference on real LSM, axon sheath and intra-axonal space probabilities are summed to obtain the binary axon foreground.

Component	Setting	Component	Setting
Label count	500 volumes; 400 train, 100 validation	Patch and voxel	$128^{3}$ voxels; 0.8 $\mu$ m
Axon-tree count	$10^{U(3.13,3.63)}$ trees/patch ( $\sim 1350$ – $4300$ )	Orientation prior	Fixed $z$ -axis Bingham; one component
Orientation variance	$\mathrm{LN}(0.1,0.05)$	Radius	$\mathrm{LN}(10^{-3},5{\times}10^{-4})$ mm; clip $2{\times}10^{-3}$ mm
Radius change	$\mathrm{LN}(1,0.1)$	Branch radius ratio	$\mathrm{LN}(1,0.1)$
Branch count	$\mathrm{LN}(0.5,1)$	Tortuosity	$\mathrm{LN}(1.1,2)$
Levels per tree	Max. 2 (zero or one branch point)	Rasterization	B-spline curves; cosine partial volume

Table 1: Synthetic label geometry. A tree denotes one synthetic axon instance before rasterization; the two-level maximum allows zero or one branch point.

U(a,b)

denotes uniform sampling, and

\mathrm{LN}(\mu,\sigma)

denotes LogNormal(mean, standard deviation).

Human
Method	Dice $\uparrow$	Corr. Dice $\uparrow$	cPrec. $\uparrow$	cRec. $\uparrow$	cFPR $\downarrow$	$\|\Delta\beta_{0}\|$ $\downarrow$	$\|\Delta\chi\|$ $\downarrow$
Thresholding	0.576	0.868	0.910	0.837	0.0088	22,504	19,499
AxonSynth	0.465	0.857	0.878	0.839	0.0109	3,377	3,692
Frangi	0.512	0.769	0.970	0.639	0.0017	1,458	1,006
Macaque
Thresholding	0.550	0.765	0.754	0.852	0.0409	14,081	9,914

Table 2: Held-out LSM segmentation results on five human and five macaque annotated patches. Calibrated thresholds (human vs. macaque): thresholding 0.47 vs. 0.46; AxonSynth 0.94 vs. 0.88; Frangi 0.05 vs. 0.01.

\uparrow

higher is better;

\downarrow

lower is better.

Figure 2: Qualitative error comparison on two held-out LSM patches (human NEFH white matter; macaque NEFH gray matter). Ground truth (GT) axons are shown in blue. For each method, green denotes true positives, yellow false positives, and pink false negatives. Thresholding tends to over-segment bright background in macaque tissue, producing many fragmented false positives, whereas AxonSynth yields fewer spurious components. In the human example, AxonSynth recovers additional faint axons but introduces some extra false positives relative to thresholding.

Explore similar work

Jun 25, 2026cs.CV

Tractography-Driven Synthetic Data Generation for Fiber Bundle Segmentation in Tracer Histology

Diffusion MRI (dMRI) tractography enables non-invasive reconstruction of white-matter pathways, but its accuracy is fundamentally limited by indirect, low-resolution measurements of axonal organization. Tracer injection studies in non-human primates provide a gold standard for validating dMRI tractography. This, however, requires time-consuming manual annotation of fiber bundles in histology sections. We propose a synthetic-data augmented framework for automated fiber bundle segmentation in macaque tracer histology. Our approach uses ex vivo dMRI tractography as a generative prior to synthesize 2D image patches for training. This provides us with sufficiently realistic foreground texture, which we compose with backgrounds from blockface photos and diversify via domain randomization. A 2D U-Net is trained on mixed real and synthetic patches. Experiments on held-out brains demonstrate improved generalization across brains and fiber bundle densities compared to training with real data only. Training with synthetic data only leads to poor performance, underscoring the need for real supervision. Overall, our approach achieves performance comparable to the state-of-the-art while requiring 3x less manually annotated data.

Kyriaki-Margarita Bintsi, Sparsh Makharia, Yaël Balbastre +4

Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School, Charlestown, MA, United States · Krea University, Sri City, Andhra Pradesh, India · Department of Experimental Psychology, University College London, London, United Kingdom +3

Medical Image Generation Synthetic Data Digital Pathology

Ask PDF

Apr 16, 2026cs.CV

Data Synthesis Improves 3D Myotube Instance Segmentation

Myotubes are multinucleated muscle fibers serving as key model systems for studying muscle physiology, disease mechanisms, and drug responses. Mechanistic studies and drug screening thereby rely on quantitative morphological readouts such as diameter, length, and branching degree, which in turn require precise three-dimensional instance segmentation. Yet established pretrained biomedical segmentation models fail to generalize to this domain due to the absence of large annotated myotube datasets. We introduce a geometry-driven synthesis pipeline that models individual myotubes via polynomial centerlines, locally varying radii, branching structures, and ellipsoidal end caps derived from real microscopy observations. Synthetic volumes are rendered with realistic noise, optical artifacts, and CycleGAN-based Domain Adaptation (DA). A compact 3D U-Net with self-supervised encoder pretraining, trained exclusively on synthetic data, achieves a mean IPQ of 0.22 on real data, significantly outperforming three established zero-shot segmentation models, demonstrating that biophysics-driven synthesis enables effective instance segmentation in annotation-scarce biomedical domains.

David Exler, Nils Friederich, Martin Krüger +5

Institute for Automation and Applied Informatics, Karlsruhe Institute of Technology, Hermann-von-Helmholtz-Platz 1, 76344 Eggenstein-Leopoldshafen, Germany. · Institute of Biological and Chemical Systems, Karlsruhe Institute of Technology, Hermann-von-Helmholtz-Platz 1, 76344 Eggenstein-Leopoldshafen, Germany. · CeMOS Research and Transfer Center, Technische Hochschule Mannheim, Paul-Wittsack-Straße 10, 68163 Mannheim, Germany.

Synthetic-To-Real Domain Gap Semi-Supervised Medical Image Segmentation U-Net 3D Generation

Ask PDF

May 4, 2026cs.CV

SIAM: Head and Brain MRI Segmentation from Few High-Quality Templates via Synthetic Training

Synthetic training has recently advanced brain MRI segmentation by enabling contrast-agnostic models trained entirely on generated data. However, most existing approaches rely on hundreds of automatically labeled templates, introducing systematic biases and limiting their flexibility to incorporate new anatomical structures. We present the Segment It All Model (SIAM), a 3D whole-head segmentation framework for 16 anatomical structures, trained using only six high-quality, manually annotated templates. SIAM extends domain randomization to both intensity and shape domains: synthetic image generation ensures contrast variability, while high-resolution spatial transformations model anatomical differences in cortical thickness and deep nuclei morphology. Unlike prior synthetic models, SIAM simultaneously segments brain as well as extra-cerebral tissues, including cerebrospinal fluid, vessels, dura mater, skull, and skin, enabling fully automated, preprocessing-free analysis. Evaluation across eight heterogeneous datasets (N=301), that include multiple contrasts (T1-weighted, T2-weighted, CT) and span a wide range of ages, demonstrates that SIAM matches or outperforms state-of-the-art methods for brain structures, in addition to extending automated segmentation to non-brain structures. The model also exhibits superior consistency across contrasts and repeated acquisitions, together with improved sensitivity to subtle gray matter atrophy. We openly release the model and the label templates at https://github.com/romainVala/SIAM.

Romain Valabregue, Ines Khemir, Eric Badinet +3

Sorbonne Universit´e, Institut du Cerveau - Paris Brain Institute ICM, CNRS, Inria, Inserm, AP-HP, Hˆopital de la Piti´e Salpˆetri`ere, F-75013, Paris, France · IMT Atlantique, LaTIM INSERM U1101, Brest, France · Aix-Marseille Universit´e, CNRS, Institut de Neurosciences de la Timone, UMR 7289, Marseille, France

Segment Anything Model Synthetic Training Data Contrastive Learning

Ask PDF