Classification

The segmentation model provided in CyPose has been hand-developed to work well with a variety of cyanobacterial cells, but requires more work to train a model on your data. You will need data containing your respective strains / cell types of interest as follows:

  • An original movie in nd2 format to pull data from
  • For each cell type, a tif format mask file containing masks specific to each cell type.

Files should be labeled as follows:

  • A base name for all files from a given movie, used for the nd2. For example, my_movie.nd2
  • Files with matching cell type labels added to the masks for that nd2. For example, my_movie_cell1.tif

So, for example, you might have the following:

  • my_movie.nd2, my_movie_cell1.tif, my_movie_cell2.tif, my_movie_cell3.tif
  • my_movie2.nd2, my_movie2_cell1.tif, my_movie2_cell2.tif, my_movie2_cell3.tif
  • my_movie3.nd2, my_movie3_cell1.tif, my_movie3_cell2.tif, my_movie3_cell3.tif

Warning

You must have the same set of channels in the same order in each movie used for training.

Once you have your data in this format, you can train a custom classifier model using the train_classification_model.py script, after making the following changes:

  • Update the root_dir variable in the FastND2DataSet loader to point to the directory containing your data.
  • In the ConvNetClassifier class, update the num_channels variable to match the number of channels in your movie and the num_classes variable to match the number of cell types you are training on.

Once that's done, you can run the script to train a model on your data. The script will save the model to a file, which you can then use to classify new data. To do this, you can use the run_classification.py script, which will classify all the cells in a given movie and save the results to a CSV file. This script can be run as follows:

python run_classification.py --model_path /path/to/model --data_dir /path/to/data

This will save the results to a file called classification_results.csv in the data directory. The results will contain the following columns: 

frame,centroid_x,centroid_y,cell_type

This can then be used in any downstream analysis you wish to perform.