A light and robust tool to do 3D cell instance segmentation for cell membrane images. It is useful when the cells clump together.

Official implementation of A novel deep learning-based 3D cell segmentation framework for future image-based disease detection. (However, we think the title should be 3DCellSeg - a robust deep learning-based 3D cell instance segmentation pipeline. We could not understand why the corresponding author J changed the title. It was so wired. Anyway, the work itself is what it is.)

• Check the required python packages in requirements.txt.

3DCellSeg was trained and tested on four datasets: HMS (private), ATAS, LRP, Ovules.

For HMS dataset, please send email to the authors to get data access.

• Put the data filefolder under a given path (e.g., \data\CellSeg_dataset in my case)

• The data structure should be like:

For HMS

\data\CellSeg_dataset\HMS
    \raw
        \100.mha
        ...
    \segmentation
        \100.mha
        ...

For ATAS

\data\CellSeg_dataset\ATAS
    \plant1
        \processed_tiffs
            \0hrs_plant1_trim-acylYFP.tif
            ...
        \segmentation_tiffs
            \0hrs_plant1_trim-acylYFP_hmin_2_asf_1_s_2.00_clean_3.tif
            ...
        ...
    ...

For LRP

\data\CellSeg_dataset\LateralRootPrimordia
    \train
        \Movie1_t00003_crop_gt.h5
        ...
    \test
        \Movie1_t00006_crop_gt.h5
        ...
    ...

For Ovules

\data\CellSeg_dataset\Ovules
    \train
        \N_404_ds2x.h5
        ...
    \test
        \N_294_final_crop_ds2.h5
        ...
    ...

• Run prepare_datasets.py (which calls py files in folder prepare_dataset) to pre-process the dataset. (Here to download the samples of pre-processed files (passwd: cellseg).)

• You may use prepare_datasets.py to pre-crop the 3D images if you find the training is slow.

• [IMPORTANT] prepare_datasets.py generates a python dict (stored in dataset_info) for each processed dataset. The dict contains file paths to each train and test files. All dataloading operations during training and testing depend on the dict. You should create your own dict. See dict samples in dataset_info.

Run train_HMS.py and train_ATAS.py to train the model on the corresponding dataset. They were implemented by PyTorch. You can easily adjust the hyperparameters.

Run notebook test_HMS.ipynb, test_ATAS.ipynb, test_LRP.ipynb, and test_Ovules.ipynb to test the model. The notebooks show details of every step.

Run notebook test_Ovules_cal_acc.ipynb, Ovules_test _cal_multi_img_acc.ipynb, test_LRP_cal_acc.ipynb, and LRP_test _cal_multi_img_acc.ipynb to calculate accuracy on Ovules and LRP datasets. (Note that the checkpoints are not the ones used in paper because I could not find the original ones.)

See folder output.

Please contact [email protected].