Neural Net Cell Segmenter

Goal: develop a tool using machine learning to segment images of packed epithelial cell layers (for future use in pycellfit)

Project started in May 2020 by Nilai Vemula

Development:

• Status: Planning and Brainstorming

To-do List:

• Background Research
• Make a FCN or other type of neural network in TensorFlow
• Write scripts to pre-process and post-process images
• Create a training set of images from the ECAD dataset (provided by James White)
• Train neural network
• Evaluate accuracy of model and continue training
• Collect code as a package

Brainstorming and Background Research

• CSML: Paper pre-print & GitHub
• Instance Segmentation: DeepMask Review
• Semantic Segmentation: FCN Review
• Tensorflow: FCN in Tensorflow
• FCN in Tensorflow for Segmentation: GitHub

Requirements

This project should require numpy and tensorflow for developing the neural network as well as Pillow and opencv for some image processing. Additional requirements include matplotlib, scikit-learn, and scipy. A full list of requirements is present in requirements.txt and should be used with a virtual environment based on Python 3.8.

Test Data

For testing data, we are using a variety of images (some from James White). Each raw tif image is found in data/raw . Each raw image is then loaded into SeedWaterSegmenter and is manually segmented using a watershed method. The output files from SeedWaterSegmenter are located in data/ground_truth/<name_of_raw_tif>. The watershed-segmented file is located in data/ground_truth/<name_of_raw_tif>/Segments/Segment_0_000.tif. This file is then converted to a black and white mesh by the neural_net_cell_segmenter/ground_truth_preprocess.py script. The mesh is saved in data/ground_truth/<name_of_raw_tif>_mask.tif.