Cell labeling scripts

This program works with the cell detection process. Currently, the program works with one animal and one task at a time. All models are stored in /net/birdstore/Active_Atlas_Data/cell_segmentation/models/ The models are named by round: models_round_X_threshold_2000.pkl where ‘X’ is the round number. The program can be run with the following commands:

  • python sr/labeling/scripts/create_labels.py –animal DKXX –task detect

  • python sr/labeling/scripts/create_labels.py –animal DKXX –task extract

  • python sr/labeling/scripts/create_labels.py –animal DKXX –task train

  • python sr/labeling/scripts/create_labels.py –animal DKXX –task fix

Explanation for the tasks:

  • detect - This is the 1st task to run and will create the cell_labels/detection_XXX.csv files. This task will run the cell detection model on the images and create the detections.

  • extract - This task will extract the predictions from the detection files and create the cell_labels/all_predictions.csv file.

  • train - This task creates the detection_XXX.csv files created above and trains the model. The features are taken from the detection_XXX.csv files. The model is saved in the cell_segmentation/models dir. This new model can then be used to rerun the detection process. Repeat as necessary.

  • fix - This is only needed when the images have the extra tissue and skull present. You will need to create the rotated and aligned masks for the images.

Plan for implementation:

  • Detect cells on available brains.

  • Some of the brains have too many points to easily display in Neuroglancer. DK59 has about 75MB of points. This won’t display and will crash the browser. We can take the points and display them as a precomputed data format, similar to the way we display large images.

  • Once we have the display of the predicted points along with the image stacks of the dye and the virus channels, we can create two more layers. A ‘bad’ layer where the user marks as ‘bad’ the predictions that are bad. And another layer ‘sure’ where the user creates annotations that the prediction process has missed.

  • These ‘bad’ and ‘sure’ new annotations are then saved to the database.

  • We then create features from these ‘bad’ and ‘sure’ coordinates.

  • These features are then fed back into the training process and a new model is created which we then use to repeat the process.