Skip to content

Latest commit

 

History

History
executable file
·
80 lines (62 loc) · 2.38 KB

README.md

File metadata and controls

executable file
·
80 lines (62 loc) · 2.38 KB
Raw

CellUNet

Runs a semantic segmentation of microscopy images with a U-Net based deep learning architecture and on-the-fly data augmentation.

Example command-line usage:

cd cellunet
python train.py -i data/nuc0.png -l data/labels0.tif -o output -n 100 -e 5 -p 256
python predict.py -i data/nuc1.png -w output/cnn_model_weights.hdf5 -o output

In brief, the model accepts as training data a series of images segmented into regions, with each region being given 1 of 3 possible labels. In the training data (where the model is trying to learn how to segment and identify cell nuclei), label 0 corresponds to the background, label 1 to the boundary, and label 2 to the interior of cell nuclei.

A few functions in utils are adopted from https://github.com/carpenterlab/unet4nuclei.

Tested with keras (2.0.0), tensorflow (1.8.0). For GPU, use Cuda 9.0, tensorflow (1.8.0), tensorflow-gpu (1.8.0), keras (2.0.0) Avoid Keras==2.2.

Parameters

Training

  • i - image file path
  • l - label file path
  • o - output directory
  • n - number of steps (default 100)
  • b - number of batches (default 16)
  • e - number of epochs (default 50)
  • p - pixel size of image patches, has to be divisible by 8 (default 256)
  • w - hdf5 weight file path
  • q - weights for loss function (default 1.0, 1.0, 1.0)

Prediction

  • i - image file path
  • w - hdf5 weight file path
  • o - output directory

Example Commands

Getting started

python train.py -i data/nuc0.png -l data/labels0.tif

Training with flexible weights (10x weights on Border)

python train.py -i data/nuc0.png -l data/labels0.tif -q 1 1 10 -o output

Resume training

python train.py -i data/nuc0.png -l data/labels0.tif -w data/weights.tests.hdf5

Training using multiple images

python train.py -i data/nuc0.png / data/nuc0.png -l data/labels0.tif / data/labels1.tif

Training using multiple color channels

Images provided may have any number of color channels; note that in the training data nuc_0.png and nuc_1.png have only 1 color channel while composite_nuc.tif has 2 color channels.

python train.py -i data/composite_nuc.tif -l data/labels0.tif -o output -n 100 -e 5 -p 256
python predict.py -i data/composite_nuc.tif -w output/cnn_model_weights.hdf5 -o output

Prediction with trained weights

python predict.py -i data/nuc0.png -w data/cnn_model_weights.hdf5