unicell 0.0.1.post0

Universal cell segmentation

UniCell

UniCell is a universal cellular segmentation tool for multi-modality microscopy images. It has three main features

• It works for various microscopy images, especially for the RGB images, e.g., bone marrow slides.
• It works for various image format (e.g., png, bmp, jpg, tif, tiff) without format converting and does not require users to manually select segmentation models and image channels.
• The inference speed is fast (~0.07s for 256x256 image and ~0.33s for 512x512 image on NVIDAI 2080Ti).

Installation

pip install git+https://github.com/bowang-lab/unicell.git

Train UniCell

unicell_train -dir <path to training set> --model_folder <unicell> --batch_size 32

Training set folder structure

training_set/
|----images
|--------img1.png
|--------img2.jpg
|--------img3.bmp
|--------img4.tif
|--------img5.tiff
|----labels
|--------img1_label.tiff
|--------img2_label.tiff
|--------img3_label.tiff
|--------img4_label.tiff
|--------img5_label.tiff

UniCell does not have limitation on the image format. The corresponding labels should have a suffix _label.tiff.

Inference

unicell_predict -i <input path> -o <output path> --pretrain_model unicell --contour_overlay

Compute metrics

We provide a interface to compute various metrics for cell segmentation results, including F1 score, precision, recall, the number of missing cells, the number of false-positive cells, and dice

com_metric -g <path to ground truth folder> -s <path to segmentation folder> -thre 0.5 0.7 0.9 -o <path to save folder> -n <csv name>

Graphical User Interface (GUI)

We develop a GUI plugin based on napari, which enables users who may not have coding experience to analyze their microscopy images visually in real time.

Install GUI: pip install napari-unicell

Online demo

We deploy an online demo on huggingface, which enables users to directly upload the cell images to get the segmentation results.

Remark: huggingface provides 2 free CPU for the deployment. So the inference can only use CPU, which is a little bit slow for large images (e.g., 1000x1000). We recommend using the command line interface or GUI to analyze large images if GPU is available on your local desktop.