Source-linked AI summary
AxonSynth: Domain-Randomized Synthetic Data for Zero-Shot 3D Axon Segmentation in Light-Sheet Microscopy
Edward Gaibor, Kyriaki-Margarita Bintsi, Carmen Luz Leiva Ureta, Zayneb Bellatif, Chiara Maffei, Wenze Li, Elizabeth Hillman, Yaël Balbastre, Anastasia Yendiki
TL;DR
Dense annotations are costly, and few light-sheet microscopy axon-segmentation models support zero-shot 3D segmentation without target-domain labels. AxonSynth trains a 3D model on domain-randomized synthetic axon data; across 10 held-out patches, it reduced component-count error in all 10 and Euler-characteristic error in 8.
Problem
Few off-the-shelf LSM axon-segmentation models address zero-shot 3D segmentation without target-domain labels, despite the high cost of manual annotation.
Source: §1 · p. 2Method
AxonSynth uses domain-randomized synthetic axon labels to train a three-class 3D segmentation model without manually annotated real training data.
Source: §1 · pp. 2–3Results
10/10 held-out patches showed lower |∆β0| error, and 8/10 showed lower |∆χ| error for AxonSynth.
Source: §4 · p. 7Takeaways & Limitations
AxonSynth is a promising first step toward automatic large-scale 3D axon segmentation from synthetic training data.
Source: §5 · p. 9Depth-wise degradation suggests modeling attenuation and blurring; broader LSM validation, other microscopy modalities, and comparison with real-annotation-trained 3D networks remain future work.
Source: §5 · p. 9
Abstract
from arXiv · showhide
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.
1 Introduction
AxonSynth addresses zero-shot 3D axon segmentation in LSM without target-domain labels by training on domain-randomized synthetic data and evaluating on annotated real patches across species and axonal markers.
1 Introduction
- Axon-specific EM methods address segmentation across modalities and species, but still depend on labeled EM data or target-domain examples and may require substantial proofreading [6].Dense connectomics pipelines have also advanced through supervised learning and large-scale reconstruction.Source: §1 · p. 2
- LSM can cover larger tissue volumes than EM, but available axon-segmentation methods often require target-domain adaptation, while manual annotation remains costly.This leaves zero-shot 3D axon segmentation in LSM without target-domain labels unaddressed.Source: §1 · p. 2
- Domain randomization varies appearance so models learn robust structure, and synthetic-to-real transfer depends on covering plausible appearances rather than making simulations realistic.Axons’ directional coherence and dense bundles make separating neighboring fibers difficult, alongside small calibers and partial-volume effects.Source: §1 · p. 2
- AxonSynth trains a 3D U-Net on synthetic data with three-class supervision, then sums sheath and intra-axonal probabilities for binary foreground segmentation at inference.Its zero-shot evaluation uses annotated LSM patches spanning two species and three axonal markers.Source: §1 · pp. 2–3, Figure 1 · p. 3
2 Methods
AxonSynth creates synthetic axon labels and domain-randomized images, then trains a three-class 3D U-Net for zero-shot segmentation. Its methods specify the synthetic data, rendering pipeline, segmentation targets, and training setup.
Synthetic labels.
- 500 dense 128×128×128 synthetic volumes at 0.8 µm isotropic resolution are split into 400 training and 100 validation volumes.Each volume includes per-axon object IDs for on-the-fly masking and a partial-volume occupancy map; axons are sampled around a shared preferred direction. Table 1 summarizes the geometry distributions, whose values are not provided here.Source: Synthetic labels · p. 3, Table 1 · p. 4
Domain-randomized image synthesis.
- Synthetic axons are rendered with randomized local density, shape perturbations, intensity, background structures, bias fields, smoothing, noise, and normalization.Whole axons are retained according to keep probabilities averaged over their voxels; the renderer spans appearances beyond realistic microscopy while preserving brighter axons than surrounding tissue [4].Source: Domain-randomized image synthesis · pp. 3–4
Segmentation targets.
- Training labels assign voxels to background, axon sheath, or intra-axonal space, with sheath voxels formed by 3D face-adjacent random erosion.For evaluation against binary annotations, sheath and intra-axonal probabilities are summed; these training classes do not imply hollow axon morphology in rendered images.Source: Segmentation targets · p. 4
Network and training.
- Experiments use a residual 3D U-Net with five encoder widths, dropout, and three softmax outputs, trained with equally weighted Dice–cross-entropy supervision branches [3].Dense labels are stored on disk, while image intensities are rendered from sampled label patches during training; one loss branch separates foreground from background and the other separates sheath from intra-axonal space.Source: Network and training · pp. 4–5
3 Experiments
The experiments evaluate AxonSynth on annotated human and macaque light-sheet microscopy patches, comparing it with calibrated thresholding and Frangi filtering using corrected segmentation and topology metrics.
3.1 Dataset
- Ten held-out patches (five each from human and macaque) sampled diverse gray/white matter regions and NEFH-, NEFL-, or PV-labeled axons at approximately 0.4 µm in-plane resolution.Expert ground truth covered 12 patches; two patches, one per species, were reserved for threshold calibration and excluded from held-out results.Source: §3.1 · p. 5
3.2 Baselines.
- Fixed foreground cutoffs were 0.47 human and 0.46 macaque for thresholding, 0.94 and 0.88 for AxonSynth, and 0.05 and 0.01 for Frangi.Frangi parameters were selected on the same validation patches by maximizing mean Dice; its filter targeted bright ridges.Source: §3.2 · p. 5
3.3 Evaluation metrics.
- Evaluation averaged per-patch corrected Dice, precision, recall, and false-positive rate, alongside absolute connected-component and Euler-characteristic errors.A two-round 6-connected face-neighbor correction modifies predictions for corrected metrics; topology metrics use the original binarized masks.Source: §3.3 · pp. 5–6
3.4 Training details.
- Training used AdamW, mixed precision, and warmup followed by cosine learning-rate decay, selecting the best epoch on synthetic validation data for real-data evaluation.Each cache contained 800 training and 200 validation samples, was reused for three epochs, and then rebuilt; the selected checkpoint was used for all real-data evaluations.Source: §3.4 · p. 6
4 Results
On held-out macaque LSM patches, AxonSynth led corrected overlap and precision, while human overlap favored thresholding; AxonSynth improved topology overall. The three-class model performed better on corrected Dice, corrected precision, and cFPR than the two-class ablation, and both methods degraded with imaging depth.
4 Results
- On macaque LSM, AxonSynth led corrected Dice (0.826) and precision (0.851), while on human LSM thresholding led Dice and corrected Dice (0.868 vs. 0.857 for AxonSynth).On human LSM, AxonSynth had similar recall to thresholding (0.839 vs. 0.837), but lower precision.Source: §4 · p. 7
- AxonSynth reduced macaque |∆β0| from 14,081 to 808 and |∆χ| from 9,914 to 1,564 versus thresholding, improving |∆β0| in 10/10 patches and |∆χ| in 8/10.In human LSM, it reduced |∆β0| from 22,504 to 3,377 and |∆χ| from 19,499 to 3,692 versus thresholding; Frangi had the lowest human topology errors. Thresholding predicted about 13.0 times the target component count, compared with 4.7 times for AxonSynth.Source: §4 · p. 7
- In the macaque example, bright-background thresholding produced many fragmented false positives, while AxonSynth recovered faint human axons but added some false positives relative to thresholding.Average human overlap metrics still favored thresholding, while AxonSynth improved topology over thresholding in both species.Source: §4 · p. 7, Figure 2 · p. 8
4.1 Ablation and diagnostic analyses.
- The three-class model improved corrected Dice from 0.785 to 0.842, corrected precision from 0.742 to 0.865, and lowered cFPR from 0.0285 to 0.0125 versus the two-class ablation.The two-class model had higher recall (0.863 vs. 0.825) and lower mean topology errors (|∆β0| 1,852 vs. 2,093; |∆χ| 1,369 vs. 2,628), but lower precision; the three-class model was selected as primary.Source: Two-class versus three-class targets · p. 8
LSM depth analysis.
- Across 12 annotated LSM patches, corrected Dice declined with depth for both thresholding and AxonSynth, with slopes of −0.135 and −0.141 per normalized-depth unit.The deepest slab's corrected Dice was lower than the most superficial slab's by 0.124 for thresholding and 0.133 for AxonSynth; normalized intensity fell with depth (slope −0.0566). Declines were stronger in macaque patches (slopes −0.215 and −0.214) than human patches (−0.055 and −0.068).Source: LSM depth analysis · pp. 8–9
5 Discussion & Conclusion
AxonSynth shows that broadly randomized synthetic axon labels can support zero-shot segmentation on real human and macaque LSM without real training labels. Its advantage over thresholding is clearest in reduced topology errors, rather than uniformly higher overlap.
Discussion & Conclusion
- Separating sheath from intra-axonal space improved real-data precision and corrected Dice, making the three-class target the conservative default.The binary-target model had higher recall and lower mean topology errors, but lower precision and higher cFPR; the three-class target may prevent neighboring axons from merging in dense bundles.Source: §5 · p. 9
- Depth-wise degradation motivates synthesis that models attenuation and blur, alongside broader LSM validation, other microscopy modalities, and comparisons with networks trained on real annotations.Source: §5 · p. 9
- AxonSynth supports zero-shot 3D segmentation on real human and macaque LSM, reducing topology errors versus thresholding without consistently improving overlap.The discussion identifies lower fragmented-component and Euler-characteristic errors as the principal advantage over thresholding.Source: §5 · p. 9