FastCellpose

Citations

Han Y, Zhang Z, Li Y, et al. FastCellpose: A Fast and Accurate Deep-Learning Framework for Segmentation of All Glomeruli in Mouse Whole-Kidney Microscopic Optical Images[J]. Cells, 2023, 12(23): 2753.

Dependencies

FastCellpose uses nearly the same environment configuration as Cellpose project, and has been heavily tested on Windows 10 system. Here are the packages they rely on(which can be automatically installed with conda/pip if missing):

• pytorch
• pyqtgraph
• PyQt5
• numpy (>=1.16.0)
• numba
• scipy
• natsort
• skimage
• cv2

You can check Cellpose/README.md for more Cellpose's local installation details.

Demo

Before you start your own data training process, we offer a demo project to ensure you have configured the proper environment.

In folder demo_infer we prepare a fine-trained model and 2 2048×2048-pixel images. You can run 2_finalspeed_whole_inference.py and check the inputs/results in:

└─demo_infer
    ├─inference_out

Own data training

FastCellpose requires a sufficient amount of data to complete its supervised training process. Here we introduce the folders and process to deal with your data(every folder has a demo image, please check them out if you have any problem).

1. Split your data into train and test set, then move them to the corresponding folders: origin and whole_mask.

2. Transfer your images and GT into small-size patches, and move them to the corresponding folders: input and mask. Remember to save some data under folder test_while_train to track the training results.

3. Transfer your mask patches into annotated form, which means background's gray value is 0, while different segmented objects' are 1,2,etc.

4. Transfer the annotated form into flow using 0_flow_production.py .

5. Start your training process using 1_train_glomeruli.py.

6. Comment out the demo part and start your inference using 2_finalspeed_whole_inference.py.

   we complete 1-3 using Matlab with the code attatched in folder preprocess_matlab. Before you start 4-6, it's recommended to skim over super_params_set.py and set the correct super params including data path, training/testing params and more.

└─data
    ├─flow
    │  ├─test
    │  │      
    │  └─train
    ├─test
    │  ├─annotate
    │  │      
    │  ├─input
    │  │      
    │  ├─mask
    │  │      
    │  ├─origin
    │  │      
    │  ├─whole_mask
    ├─test_while_train
    │  ├─input
    │  │      
    │  ├─mask
    └─train
        ├─annotate
        │      
        ├─input
        │      
        ├─mask
        │      
        ├─origin
        │      
        └─whole_mask

Advanced model settings

LR iteration

Change in core.py line 851

base feature channels num of U-Net

Change in core.py. There are [16, 32, 64, 128] and [32, 64, 128, 256] to choose in python class: UnetModel

conv num in every encoder/decoder layer of U-Net

Change in resnet_torch.py. There are 2 or 4 to choose by commenting or uncommenting out those python class: convup, convdown, resup and resdown