PaperlayerView exact revision on arXiv

Source-linked AI summary

A Fully Automatic Pipeline for 3D Dendrite Instance Segmentation in SBF-SEM

Zewen Zhuo, Ilya Belevich, Eija Jokitalo, Alejandra Sierra, Jussi Tohka

arXiv:2610.03332v1cs.CV

TL;DR

Dense, weakly bounded SBF-SEM volumes make consistent 3D dendrite reconstruction difficult, especially when dense 3D annotations are impractical. This paper combines automatic detection, prompted segmentation, inter-slice linking, and high-resolution refinement; it reports high semantic accuracy on both tissue types, with dense-tissue instance recognition the principal remaining limitation.

  • Problem

    SBF-SEM volumes contain densely packed structures with weak boundaries, making distinct 3D dendrite tracking difficult to scale manually.

    Source: §I · p. 1

  • Method

    A fully automatic pipeline combines YOLOv6-prompted SAM masks, iterative 2D refinement, random-forest 3D linking, and instance-aware native-resolution nnU-Net refinement.

    Source: §I · p. 2, E. 3D Instance Linking · p. 3

  • Results

    Dice was 0.93 on control tissue and 0.91 on epileptic tissue, while instance recognition and separation were the dominant errors in dense epileptic tissue.

    Source: B. Quantitative Results · p. 4, §IV · p. 5

  • Takeaways & Limitations

    Native-resolution refinement recovers thin protrusions attenuated by downsampling, supporting reconstruction of dendritic structures beyond the downsampled masks.

    Source: §I · p. 2

    In denser epileptic tissue, detection and instance-level scores fall, with errors dominated by recognizing and separating dendrites rather than delineating them.

    Source: B. Quantitative Results · p. 4

Abstract

from arXiv · show

Accurate three-dimensional (3D) reconstruction of individual dendrites in serial block-face scanning electron microscopy (SBF-SEM) is essential for quantifying structural plasticity in the brain, yet manual annotation at scale is infeasible. We present a fully automatic pipeline for 3D dendrite instance segmentation that unifies YOLOv6-guided Segment Anything Model (SAM) prompting on downsampled slices, iterative two-dimensional mask refinement, random forest 3D instance linking, and instance-aware high-resolution refinement using nnU-Net at native resolution into a single system requiring no manual prompting at inference. Applied to hippocampal CA1 SBF-SEM datasets from a control rat and a pilocarpine- induced epileptic rat, our pipeline reconstructs coherent, well- separated dendrites with high semantic accuracy (Dice 0.93 and 0.91) and strong instance-level performance on control tissue, while analysis of the more challenging epileptic tissue identifies instance recognition in dense regions as the principal remaining limitation. The high-resolution refinement stage recovers thin dendritic protrusions, providing a basis for downstream spine- level analysis. Code is available at https://github.com/ ZE-WEN/dendrite-3d-instance-seg.

I. INTRODUCTION

Dense, weakly bounded SBF-SEM volumes make consistent 3D dendrite instances difficult to recover, while dense 3D ground truth is scarce for thin, branching dendrites. The paper proposes a fully automatic pipeline combining detector-prompted segmentation, inter-slice linking, and native-resolution refinement.

I. INTRODUCTION

  • SBF-SEM packs neuronal and non-neuronal structures into large, anisotropic volumes with weak boundaries, making distinct 3D dendrite tracking difficult to scale manually.Epileptic CA1 dendrites can also show spine loss, beading, and irregular profiles.Source: §I · p. 1
  • For thin, branching dendrites, scarce dense 3D ground truth leaves reconstruction from modular pretrained components without dense annotation or joint end-to-end training an open problem.Source: §I · p. 2
  • The proposed pipeline resolves inter-slice identity with an explicit, interpretable classifier while composing pretrained stages without joint end-to-end optimization.Source: §I · p. 2
  • YOLOv6 supplies bounding-box prompts to a promptable segmenter, producing fully automatic per-slice masks.Source: §I · p. 2
  • An iterative 2D refinement loop and random-forest inter-slice linker assemble slice-wise masks into coherent 3D instances.Source: §I · p. 2
  • An instance-aware nnU-Net branch refines linked instances at native resolution, recovering thin protrusions attenuated by downsampling.Source: §I · p. 2
  • Evaluation on control and epileptic hippocampal CA1 data identifies instance recognition in dense tissue, rather than delineation, as the dominant remaining failure mode.Source: §I · p. 2

A. Datasets and Annotations

The study uses two hippocampal CA1 rat datasets, sparse annotations for training and evaluation, and aligned downsampled and native-resolution image stacks.

A. Datasets and Annotations

  • Two CA1 rat datasets comprise control tissue and tissue from a rat after pilocarpine-induced status epilepticus.Both were imaged at 15 × 15 nm in-plane resolution with a 40 nm sectioning interval.Source: A. Datasets and Annotations · p. 2
  • Dataset A’s 100 annotated slices (2,168 2D dendrite instances) trained YOLOv6; 30 slices of sparse 3D annotations (27 instances) trained the RF 3D linking classifier and nnU-Net tissue classifier.Source: A. Datasets and Annotations · p. 2
  • Evaluation used sparse 3D instance annotations from separate slices of both datasets, with no Dataset B evaluation slice used for training.Source: A. Datasets and Annotations · p. 2

B. Preprocessing

  • Aligned stacks were retained at 1024 × 1024 for YOLO-guided SAM segmentation and at original resolution for high-resolution refinement.StackReg transformations estimated on downsampled stacks were scaled and applied to original-resolution slices.Source: B. Preprocessing · p. 2

C. YOLO-Guided SAM Segmentation

YOLOv6 detections prompt SAM independently on downsampled slices; subsequent iterative processing addresses artifacts in the resulting masks.

C. YOLO-Guided SAM Segmentation

  • YOLOv6 detects candidate dendrites on downsampled, aligned slices, and each predicted box independently prompts SAM to produce a mask for the slice instance map.Source: C. YOLO-Guided SAM Segmentation · p. 2

D. Iterative 2D Mask Refinement

  • Because SAM processes boxes independently, initial masks can contain spurious components, holes, missed detections, and split labels.Source: D. Iterative 2D Mask Refinement · p. 2
  • Three refinement cycles remove small components, fill holes, harmonize nearby labels, and create supplementary prompts for regions found in the next two slices but absent in the current slice.Components smaller than 500 pixels are removed; harmonization considers proximity, contact strength, dilated overlap, and background gap thickness.Source: D. Iterative 2D Mask Refinement · p. 2

E. 3D Instance Linking

A random-forest classifier scores candidate links between components on adjacent slices, and retained links are assembled into volumetric dendrite instances.

E. 3D Instance Linking

  • Training links are candidate pairs of components on adjacent slices whose pixel overlap exceeds a minimum threshold, labeled positive when both belong to the same annotated instance.Source: E. 3D Instance Linking · p. 3
  • Each candidate link is represented by eight features covering overlap, component size, area ratio, centroid displacement, and centroid distance.The feature vector also includes intersection over union.Source: E. 3D Instance Linking · p. 3
  • At inference, the classifier scores adjacent-slice candidate edges; thresholded edges are resolved by union-find so branches remain one connected 3D instance.Source: E. 3D Instance Linking · p. 3

F. High-Resolution Refinement

The high-resolution branch uses nnU-Net tissue evidence to restore thin protrusions at native resolution while retaining the linked instances’ identities.

F. High-Resolution Refinement

  • Downsampling weakens thin spine necks and fine protrusions, so nnU-Net supplies tissue evidence at original resolution after 3D linking.Nearest-neighbor upscaling preserves discrete instance identities, which assign recovered voxels to the correct dendrite.Source: F. High-Resolution Refinement · p. 3
  • The binary 2D nnU-Net is trained on 30 annotated original-resolution slices, using a self-configured 1280 × 1024 patch and batch size 2.Training uses Z-score normalization.Source: F. High-Resolution Refinement · p. 3
  • Recovery thresholds the probability volume at 0.98 and adds connected components that overlap each linked instance on that slice.Because recovery is additive, it retains instance identity while reattaching high-confidence protrusions.Source: F. High-Resolution Refinement · p. 3

III. EXPERIMENTS AND RESULTS

The evaluation uses sparse, manually refined references and reports instance, panoptic, and sparse semantic measures for control and epileptic datasets. Its figures show the pipeline and qualitative reconstructions, while Table I covers quantitative results across the two datasets.

III. EXPERIMENTS AND RESULTS

  • Each dataset has 10 sparsely annotated slices; pipeline output initialized references that were manually refined, though ambiguous-boundary bias cannot be fully excluded.Missed dendrites were annotated from scratch.Source: A. Evaluation Protocol · p. 3
  • Predictions entirely outside the union of annotated reference voxels were ignored because they could be valid but unannotated structures.This sparse-region strategy underlies the reported metrics.Source: A. Evaluation Protocol · p. 3
  • Instance quality uses one-to-one Hungarian matching of pairs meeting IoU thresholds and summarizes instance Jaccard across thresholds.The matching maximizes total IoU.Source: A. Evaluation Protocol · p. 3
  • With sparse foreground-only annotations, semantic precision is not identifiable, and reported IoU measures recovery of annotated dendrite voxels.Predicted voxels outside the annotated region are unpenalized.Source: A. Evaluation Protocol · p. 3
  • Figure 1 depicts the linked-mask and high-resolution branches, Figure 2 pairs slice overlays with 3D reconstructions, and Table I reports results for both datasets.Figure 2 compares control and epileptic tissue; Table I is identified as quantitative results for Dataset A and Dataset B.Source: Figure 1 · p. 4, Figure 2 · p. 4, Table I · p. 4

B. Quantitative Results

Semantic recovery remained high across control and epileptic tissue, but detection and instance scores declined more sharply on epileptic tissue, pointing to recognition and separation—not delineation—as the dominant difficulty.

B. Quantitative Results

  • Dice was 0.93 on Dataset A and 0.91 on Dataset B, while IoU was 0.87 and 0.83, respectively.Segmentation quality was SQ 0.82 on Dataset A and 0.72 on Dataset B.Source: B. Quantitative Results · p. 4
  • Detection quality fell from 0.86 to 0.51, panoptic quality from 0.70 to 0.37, and Sinst from 0.63 to 0.38 between Dataset A and Dataset B.Because SQ declined less than DQ, the dominant errors involved recognizing and separating dendrites rather than delineating them.Source: B. Quantitative Results · p. 4
  • Table II presents an ablation of the inter-slice linking model.The supplied caption identifies the ablation’s scope but gives no comparative outcome.Source: Table II · p. 5
  • Instances detected only on their largest cross-sections could miss thin leading and trailing slices, lowering DQ and Sinst while their voxels still contributed to semantic scores.This mechanism helps explain the divergence between instance-level and semantic results.Source: B. Quantitative Results · p. 5
  • Sinst@0.75 was 0.44 and 0.11; averaging instance Jaccard up to IoU 0.90 makes these instance scores conservative for thin branching structures.A few-voxel disagreement can push an otherwise correct instance below a threshold.Source: B. Quantitative Results · p. 5

C. Ablation of the Inter-Slice Linking Model

The RF linker outperformed fixed-threshold geometric matching on both datasets, with the largest margin on control tissue. Its advantage is bounded by resolution-specific features, while the comparison does not establish that no geometric rule could match it.

C. Ablation of the Inter-Slice Linking Model

  • Sinst was 0.63 against 0.33 on Dataset A and 0.38 against 0.25 on Dataset B, favoring RF linking over geometric overlap matching in both.The margin was largest on Dataset A; upstream detection failures constrained both linkers on Dataset B.Source: C. Ablation of the Inter-Slice Linking… · p. 5
  • Learned scoring helped most where recognition was reliable by sustaining tracks through slice-to-slice variation rejected by a fixed overlap criterion.On control tissue, the geometric baseline scored 0.29 at both Sinst@0.5 and Sinst@0.75, with losses from unassembled tracks rather than imprecise ones.Source: C. Ablation of the Inter-Slice Linking… · p. 5
  • Six of the RF linker's eight features use absolute pixel units, requiring retraining to apply a classifier trained at one working resolution to another.The geometric baseline matches components one to one, splitting dendrites at branch points, and uses a single untuned overlap threshold; the result shows an advantage at the tested operating point, not that no geometric rule could match it.Source: C. Ablation of the Inter-Slice Linking… · p. 5

D. Qualitative Results

Qualitative reconstructions show coherent, separated control dendrites, while epileptic tissue contains denser, more irregular profiles that are harder to detect and track across slices.

D. Qualitative Results

  • Control-tissue reconstructions form coherent, clearly separated dendritic shafts.These are representative qualitative results for Dataset A.Source: D. Qualitative Results · p. 5
  • Epileptic-tissue profiles are denser and more irregular, with a larger proportion of small, fragmented profiles.These profiles are harder to detect and track consistently across slices.Source: D. Qualitative Results · p. 5
  • The qualitative comparison covers both control Dataset A and epileptic Dataset B.The contrast is between coherent, separated shafts in control tissue and denser, more irregular profiles in epileptic tissue.Source: D. Qualitative Results · p. 5

IV. DISCUSSION AND CONCLUSION

The discussion concludes that the pipeline reconstructs dendrite instances without manual prompting, with strong control-tissue results and remaining errors concentrated in recognizing and separating dendrites. Results are limited to one animal per condition, and generalization to other animals and acquisition protocols remains to be established.

IV. DISCUSSION AND CONCLUSION

  • The fully automatic pipeline combines detection, promptable 2D segmentation, iterative refinement, RF 3D linking, and instance-aware high-resolution refinement without manual prompting at inference.On control tissue it produced coherent, well-separated dendrites with Dice 0.93 and strong instance-level performance.Source: §IV · p. 5
  • On epileptic tissue, detection and instance-level scores fell while sparse semantic recovery remained high and segmentation quality was moderate, localizing dominant errors to recognition and separation.The linking ablation also found that fixed-threshold overlap matching roughly halved the control-tissue instance score relative to the learned edge scorer.Source: §IV · p. 5
  • Results come from one animal per condition, and the RF linker and nnU-Net were trained on only 30 annotated Dataset A slices.Generalization to other animals and acquisition protocols remains to be established.Source: §IV · p. 5
  • Native-resolution refinement reattaches thin protrusions that provide a substrate for identifying and separating individual dendritic spines.This supports the intended downstream spine-level morphological analysis.Source: §IV · p. 5
Loading 2610.03332v1…