Cellpose models for Label Prediction from Brightfield and Digital Phase Contrast images

Name: Cellpose models for Brightfield and Digital Phase Contrast images

Data type: Cellpose models trained via transfer learning from the ‘nuclei’ and ‘cyto2’ pretrained model with additional Training Dataset . Includes corresponding csv files with 'Quality Control' metrics(§) (model.zip).

Training Dataset: Light microscopy (Digital Phase Contrast or Brightfield) and automatic annotations (nuclei or cyto) (https://doi.org/10.5281/zenodo.6140064)

Training Procedure: The cellpose models were trained using cellpose version 1.0.0 with GPU support (NVIDIA GeForce K40) using default settings as per the Cellpose documentation . Training was done using a Renku environment (renku template).

Command Line Execution for the different trained models

nuclei_from_bf:

cellpose --train --dir 'data/train/' --test_dir 'data/test/' --pretrained_model nuclei  --img_filter _bf --mask_filter _nuclei --chan 0 --chan2 0 --use_gpu --verbose

cyto_from_bf:

cellpose --train --dir 'data/train/' --test_dir 'data/test/' --pretrained_model cyto2 --img_filter _bf --mask_filter _cyto --chan 0 --chan2 0 --use_gpu --verbose

nuclei_from_dpc:

cellpose --train --dir 'data/train/' --test_dir 'data/test/' --pretrained_model nuclei  --img_filter _dpc --mask_filter _nuclei --chan 0 --chan2 0 --use_gpu --verbose

cyto_from_dpc:

cellpose --train --dir 'data/train/' --test_dir 'data/test/' --pretrained_model cyto2 --img_filter _dpc --mask_filter _cyto --chan 0 --chan2 0 --use_gpu --verbose

nuclei_from_sqrdpc:

cellpose --train --dir 'data/train/' --test_dir 'data/test/' --pretrained_model nuclei --img_filter _sqrdpc --mask_filter _nuclei --chan 0 --chan2 0 --use_gpu --verbose

cyto_from_sqrdpc:

cellpose --train --dir 'data/train/' --test_dir 'data/test/' --pretrained_model cyto2 --img_filter _sqrdpc --mask_filter _cyto --chan 0 --chan2 0 --use_gpu --verbose

NOTE (§): We provide a notebook for Quality Control, which is an adaptation of the "Cellpose (2D and 3D)" notebook from ZeroCostDL4Mic .

NOTE: This dataset used a training dataset from the Zenodo entry(https://doi.org/10.5281/zenodo.6140064) generated from the “HeLa “Kyoto” cells under the scope”  dataset Zenodo entry(https://doi.org/10.5281/zenodo.6139958) in order to automatically generate the label images.

NOTE: Make sure that you delete the “_flow” images that are auto-computed when running the training. If you do not, then the flows from previous runs will be used for the new training, which might yield confusing results.