NuSeT

NuSeT: A Deep Learning Tool for Reliably Separating and Analyzing Crowded Cells

To run the code

The model was bulit on python3 and tensorflow 1, the following packages are needed for NuSeT to run:

tensorflow (currently only works for tensorflow 1) (pip install tensorflow==1.15 / pip install tensorflow-gpu==1.15)
PIL (pip install Pillow)
numpy (pip install numpy)
scikit-image (pip install scikit-image)
tqdm (pip install tqdm)

To install all packages, simply run python setup.py in the root directory

NuSeT works better for moderate sized nuclei/cells, please adjust the 'resize ratio' in 'Configuration' section under 'Predicting' module. For optimal nuclei/cells sizes, please refer to images under 'sample_image' folder.

Update: we have noticed that the new OS update on Macbook is conflicting with tkinter graphic user interface, if that is the case, please use the NuSeT_CL (command line version) instead

After finishing installing packages, download the 2 weight files from google drive:(https://drive.google.com/file/d/1fcs1F2lGPX0ejzEGPZ63YNF3AmUbdBcM/view?usp=sharing https://drive.google.com/file/d/1hythQfvD6kbaUClAPY96nHcXB7RXVmBx/view?usp=sharing), move those files to Network/ folder. Then navigate to the root folder of this repo (NuSeT/), in the command line run python3 NuSeT.py.

For the detailed user-guide, please see our paper: https://www.biorxiv.org/content/10.1101/749754v1

The motivation for this work

Tools for segmenting fluorescent nuclei need to address multiple features and limitations of biological images. Typical issues and limitations include:

Boundary assignment ambiguity: biological samples frequently have very high cell density with significant overlap between objects.
Signal intensity variation: Within one image, the signal can vary within each nucleus (e.g. due to different compaction states of the DNA in heterochromatin vs. euchromatin) and across nuclei (e.g. due to cell to cell differences in nuclear protein expression levels and differences in staining efficiency).
Non-cellular artifacts and contaminants: Fluorescence microscopy samples are often contaminated with auto-fluorescent cell debris as well as non-cellular artifacts.
Low signal to noise ratios (SNRs): Low SNRs typically result from lower expression levels of fluorescent targets and/or high background signal, such as sample auto fluorescence.

The highlights for this work

This work took the advantages of two state-of-the-art cell segmentation models, U-Net and Mask-RCNN. This work also incorporated other algorithms to specifically address issues cell biologists may encounter during imaging. The improvements include:

Fusing U-Net parallely with Region Proposal Network (RPN), following a watershed transform to achieve accurate cell-boundary assignment in dense environment.
Foreground normalization to improve detection on sample paration artifacts and signal variations
Sythetic images to further improve detection on sample paration artifacts and cell boundary assignment.
Graphic user interface for using pre-trained NuSeT models, and for training new models using custom training data.

Results using pre-trainined NuSeT model (on Kaggle 2018 data science bowl and images from our lab)

Reference and citation

This work is inspired from https://github.com/tryolabs/luminoth/tree/master/luminoth/models/fasterrcnn, https://github.com/endernewton/tf-faster-rcnn, https://www.kaggle.com/c/data-science-bowl-2018, https://github.com/matterport/Mask_RCNN.