Cell candidate classification#

To change how the cell candidate classification step runs, you can change these options:

  • --trained-model To use your own network (not the one supplied with cellfinder) specify the model file.

  • --model-weights To use pretrained model weights. Ensure that this model matches the --network-depth parameter.

  • --network-depth. Resnet depth (based on He et al. (2015)) Default: 50

  • --batch-size Batch size for classification. Can be adjusted depending on GPU memory. This can often be increased on high-memory modern GPUS (e.g. 128 works well on a Titan RTX). Default: 64

You shouldn’t need to change these:

  • --x-pixel-um-network The pixel size (in microns, in the first dimension) that the machine learning network was trained on. Set this to adjust the pixel sizes of the extracted cubes. Default 1

  • --y-pixel-um-network The pixel size (in microns, in the second dimension) that the machine learning network was trained on. Set this to adjust the pixel sizes of the extracted cubes. Default 1

  • --z-pixel-um-network The pixel size (in microns, in the third dimension) that the machine learning network was trained on. Set this to adjust the pixel sizes of the extracted cubes. Default 5

  • --cube-width The width of the cubes to extract in pixels (must be even). Default 50

  • --cube-height The height of the cubes to extract in pixels (must be even). Default 50

  • --cube-depth The depth (z)) of the cubes to extract in pixels(must be even). Default 20

  • --save-empty-cubes If a cube cannot be extracted (e.g. to close to the edge of the image), save an empty cube instead. Useful to keep track of all cell candidates.