REN

REN is the acronym for ROI Extraction Net and it offers the automatic extraction of ROIs as binary mask in (microscopic) images using a CNN implemented in Tensorflow.

Usage

from ren.preprocessing import TileGenerator

gen = TileGenerator()
gen.add_files("mixed/", img_type = "tif", roi_type = "zip")
gen.generate_tiles(tiles_per_file = 8, tile_size = (600, 600), max_border_fill_fac = 0.1, \
                    rotate_tile = True, mirror_tile = True, improve_tile = True, \
                    num_threads = 4)
# yields tf.data.TensorSliceDataset with the image and roi mask tiles

Requirements

• python >= 3.5
• tensorflow >= 1.9
• https://github.com/hadim/read-roi (in order to read ImageJ's ROI files)
• scikit-image >= 0.16.2 (support for TIFFs with 16-bit color depth)
• tifffile as plugin for skimage.io.imread
• tqdm (fancy progress bar)
• joblib (easy parallelization)
• matplotlib (visualization)

Installation

First download the zipped repository from here or clone the repository, then install the module with

pip install -e path_to_package/

Notes

It is recommened to use the module tifffile (in combination with scikit-image) to read TIFF files with 16 bit color depth. Pillow (which is integrated in matplotlib) only supports 8 bit color depth which could mean an information loss. For really big TIFFs (approx. 1 GB) used in geoinformatics e.g. satellite imagery, the module osgeo.gdal is recommened.

License

All source code is under the Artistic License