SegNeuron

Official implementation, datasets and trained models of "SegNeuron: 3D Neuron Instance Segmentation in Any EM Volume with a Generalist Model" （MICCAI 2024)

Tip

SegNeuron is now available as an Agent Skill in EM-Skills.

EM-Skills is a growing collection of reusable Agent Skills for EM analysis, currently including:

• 🧠 Neuron Segmentation Skill — SegNeuron-based neuron reconstruction
• 🧬 Mitochondria Segmentation Skill — MitoNet-based mitochondria segmentation
• 🎯 Annotation Selection Skill — informative region selection for efficient annotation
• 🎥 EM Visualization Skill — CloudVolume-based visualization and video generation

👉 See EM-Skills for installation and usage. Contributions are welcome—feel free to submit issues, propose new EM Skills, or improve existing ones.

How does SegNeuron speed up neuron segmentation in EM volumes?

The general-purpose model achieves outstanding reconstruction performance on entirely unseen 3D EM datasets (x/y resolution: 5–10 nm). Human experts only need to perform connectivity corrections on the coarse segmentation results, which can then be directly used to fine-tune SegNeuron or to train new lightweight models.

Environments

We have packaged all the dependencies into Connect.tar.gz, which can be directly downloaded for easy access here.

Datasets and Models

The datasets required for model development and validation are available here. The trained models can be download here. If you use any of the following vEM datasets in your work, please also cite the corresponding original publications:

• vEM1: MiRA-ADWT
  Ultrastructural Alterations of Dendritic Morphology in the Prefrontal Cortex of Alzheimer’s Disease Model Rats link

• vEM2: MiRA-ZF
  Multiplexed Neuromodulatory-Type-Annotated EM-Reconstruction of Larval Zebrafish link

• vEM3: MiRA-SCN
  Connectomic Organization of the Suprachiasmatic Nucleus link

• vEM4: MiRA-PIB
  PIB: Parallel ion beam etching of sections collected on wafer for ultra large-scale connectomics link

Table: Details of EMNeuron

Dataset	Modality	Res.($nm$) ($x/y,z$)	Total voxels (M)	Labeled voxels (M)	Dataset	Modality	Res.($nm$) ($x/y,z$)	Total voxels (M)	Labeled voxels (M)
ZFinch	SBF-SEM	9, 20	3635	131	HBrain	FIB-SEM	8, 8	3072	844
Layer4	SBF-SEM	9, 20	1674	-	FIB25	FIB-SEM	8, 8	312	312
vEM1 (adwt)	ATUM-SEM	8, 50	1205	157	Minnie	ssTEM	8, 40	2096	-
vEM2 (zfish)	ATUM-SEM	8, 30	1329	281	Pinky	ssTEM	8, 40	1165	117
vEM3 (scn)	ATUM-SEM	8, 40	1301	253	FAFB	ssTEM	8, 40	2625	577
MitoEM	ATUM-SEM	8, 30	1048	-	Basil	ssTEM	8, 40	23	23
H01	ATUM-SEM	8, 30	1166	118	Harris	others	6, 50	30	30
Kasthuri	ATUM-SEM	6, 30	1526	478	vEM4 (ionsem)	others	8, 20	45	-

Training

1. Pretraining

cd Pretrain

python pretrain.py

2. Supervised Training

cd Train_and_Inference

python supervised_train.py

Inference

1. Affinity Inference

cd Train_and_Inference

python inference.py

2. Instance Segmentation

cd Postprocess

python FRMC_post.py

3. Zero-shot Segmentation Examples on MitoEM and Wildenberg (scale bar: 2 um)

Acknowledgement

This code is based on SSNS-Net (IEEE TMI'22) by Huang Wei et al. The postprocessing tools are based on constantinpape/elf. Should you have any further questions, please let us know. Thanks again for your interest.