Sophon

Paper | Dataset

Abstract

Sophon is a robust model for nanocrystal segmentation in transmission electron microscopy (TEM) images. The model was trained on large-scale mixed datasets along with large amount of unlabeled TEM image data. Besides, with our carefully designed weak label generation pipeline and weakly supervised learning process, our segmentation model achieves superior performance for robust nanocrystal segmentation.

We show the learning process in the following:

Dataset preparation

Our training, testing, and additional datasets from 10 other publications can be downloaded from the link above. Due to the large size of the weakly labeled dataset, it is not included in the shared folder. If you would like access to it, please contact us at: liangyingping@bit.edu.cn.

Datasets	Type	Num of Images	Number of Instances	Link
Ours (labeled, OD)	Nanocrystals	523	49,976	Manual
Ours (unlabeled, W)	Nanocrystals	7344	918,531	Automatic
Extra Data 1	Nanocrystals	80	8,881	Workflow towards automated segmentation of agglomerated, non-spherical particles from electron microscopy images using artificial neural networks
Extra Data 2	Cell	540	67,724	Cellpose: a generalist algorithm for cellular segmentation
Extra Data 3	Cell	493	1,885	An Instance Segmentation Dataset of Yeast Cells in Microstructures
Extra Data 4	Cell	200	23,615	Evaluation of Deep Learning Strategies for Nucleus Segmentation in Fluorescence Images
Extra Data 5	Nanocrystals	93	1,763	nNPipe: a neural network pipeline for automated analysis of morphologically diverse catalyst systems
Extra Data 6	Virus	622	3,192	VISN: virus instance segmentation network for TEM images using deep attention transformer
Extra Data 7	Cell	665	30,704	Nuinsseg: A fully annotated dataset for nuclei instance segmentation in h&e-stained histological images
Extra Data 10	Cell	30	8,251	CryoNuSeg: A dataset for nuclei instance segmentation of cryosectioned H&E-stained histological images
Extra Data 11	Cell	1169	81,915	Blood Cell Segmentation Dataset
Extra Data 12	Nanocrystals	465	11,541	Bayesian Particle Instance Segmentation for Electron Microscopy Image Quantification

The dataset for evaluation contains 80 images with manual labels.

Performance

We show the performance on our validation dataset. "OD" indicates our labeled dataset, "W" indicates weakly supervised learning on unlabeld data, and "LD" indicates extra datasets:

Category	AP_50	AP_75	AP_90	mIoU
OD	68.5	51.6	12.1	69.3
OD+W	80.1	66.3	18.1	81.9
OD+LD	81.3	72.8	22.9	81.9
OD+LD+W	82.5	70.1	25.3	84.5

Results

Train or Finetune your own model

Please refer to Training Installation for installation instructions.

How to use our software

If you use Sophon, please cite our paper:

@article{DLMCNS2025,
    author  = {Kai Gu and Yingping Liang and Jiaming Su and Peihan Sun and Jia Peng and Naihua Miao and Zhimei Sun and Ying Fu and Haizheng Zhong and Jun Zhang},
    title   = {Deep Learning Models for Colloidal Nanocrystal Synthesis},
    journal = {ACS Nano},
    year    = {2025},
    volume  = {19},
    number  = {45},
    pages   = {39025--39034},
    doi     = {10.1021/acsnano.5c09134}
}

Software with our segmentation model will be available once the paper is accepted.

Please refer to Software Usage for software usage.

License and Citation

This repository can only be used for personal/research/non-commercial purposes. Please cite the following paper if this model helps your research:

@article{DLMCNS2025,
    author  = {Kai Gu and Yingping Liang and Jiaming Su and Peihan Sun and Jia Peng and Naihua Miao and Zhimei Sun and Ying Fu and Haizheng Zhong and Jun Zhang},
    title   = {Deep Learning Models for Colloidal Nanocrystal Synthesis},
    journal = {ACS Nano},
    year    = {2025},
    volume  = {19},
    number  = {45},
    pages   = {39025--39034},
    doi     = {10.1021/acsnano.5c09134}
}

Contact

If you find any problem, please feel free to contact me (liangyingping@bit.edu.cn). A brief self-introduction (including your name, affiliation and position) is required, if you would like to get an in-depth help from me. I'd be glad to talk with you if more information (e.g. your personal website link) is attached.

Acknowledgments

The GUI code is borrowed from Cellpose, we thank the authors for their great effort.