Home

Awesome

MorphConNet

Chi Zhang, Qihua Chen, Xuejin Chen. Self-Supervised Learning of Morphological Representation for 3D EM Segments with Cluster-Instance Correlations. In MICCAI 2022

Dataset: FAFB-CellSeg

cd fafb-cellseg
uzip *.zip 

The segment ids of corresponding classes are in glias/neurites/soma.txt. You can copy the ids and visualize them in Neuroglancer.

The point cloud data are in the folders, where each line records the coordinates of a point.

/*Note that we expanded the number of glia in the dataset to 331 (in the paper we have 110).
Feature embedding space of MorphConNet on FAFB-CellSeg.

<div align="center"><img src="https://github.com/zhangchih/MorphConNet/blob/main/feature_distribution.png" width="400px" />

The glia segments we annotated in the fly brain:

<div align="center"><img src="https://github.com/zhangchih/MorphConNet/blob/main/glia_331.png" width="800px" /> <div align="left">

Test the pre-trained self-supervised model

uzip ./code/pretrained/model_sslcontrast.pth.zip

Finetune the model with the classification task of fafb-CellSeg: finetune_eval.ipynb
Linear classification and t-sne visualization: linear_feature_eval.ipynb

Pretrain the model with your own dataset

Format your dataset as:

your_dataset_forlder  
├── neuron_dataset_shape_names.txt    # contains the class_name
├── neuron_dataset_test.txt   # contains the file names in folder class_name
├── class_name  
    ├── your_point_cloud_data1.txt  
    ├── your_point_cloud_data2.txt
    ├── ...

Modify the dataset path in ./main.py
Config the training parameters in ./config/config.yaml

python main.py