ecSeg: Semantic Segmentation of Metaphase Images containing Extrachromosomal DNA

Reference implementation of methods outlined in ecSeg: Semantic Segmentation of Metaphase Images containing Extrachromosomal DNA.

ecSeg: Semantic Segmentation of Metaphase Images containing Extrachromosomal DNA
Utkrisht Rajkumar, Kristen Turner, Jens Luebeck, Viraj Deshpande, Manmohan Chandraker, Paul Mischel, and Vineet Bafna

Installation

This platform requires a modern processor with support for AVX instructions and python 3.5 or greater.

We highly recommend installing ecseg through a conda environemnt.

git clone https://github.com/ucrajkumar/ecSeg
cd ecSeg
conda create -n ecseg python=3.7
source activate ecseg
conda install Pillow tqdm scikit-image
conda install -c menpo opencv 
pip install keras=2.2.5
pip install matplotlib=2.2.4

Run ecSeg

To produce segmentations, run ecSeg.py:

python3 ecSeg.py -i "input_path"

Input specifications

1. input_path (must be enclosed by double quotes "").

For example: python3 ecSeg.py -i "C:/Users/Utkrisht/path/to/images" 2. Software will only read the .tif images 3. Input images must be 1040x1392x3 (RGB images)

Note: The flag, -i, must be provided.

Output

1. Coordinates folder will be created which will contain a coordinate file for each image. Each coordinate file will have the coordinates of all the ecDNA present in the corresponding image in the form (x, y).
2. Labels folder will contain the RGB version of the post-processed segmentation. It will also contain raw values saved as a .npy file.

Segmentation details

The segmented images (.npy) will be in dimension 1024x1280.

To extract individual classes (seg.npy can be found in "example_results" folder):

seg_I = np.load('example_results/seg.npy')
background = (seg_I==0)
nuclei = (seg_I==1)
chromosome = (seg_I==2)
ecDNA = (seg_I==3)

Run ecSeg_fish

To analyze fish interaction run ecSef_fish.py:

python3 ecSeg_fish.py -i "input_path"

Input specifications

Arguments	Description
-h	Displays argument options
-i	Path to folder containing images. Must be wrapped in double quotes. See example above on how to run ecSeg.py.
-c	Fish color (optional). Must be 'green' or 'red'
-t	Threshold value (optional). Threshold values must be [0, 255]. Indicates sensitivity of FISH interaction. 0 and 255 are the least and highest sensitivity, respectively
-p	Segment boolean (optional). Must be 'True' or 'False'. Indicates whether to re-segment images. Enter 'False' if you have already segmented the images

Training/test Dataset

Training and test dataset can be downloaded from:

https://data.mendeley.com/datasets/m7n3zvg539/draft?a=30ace699-6d6a-4c49-a770-29b09f759795

Dataset	Description
train_im	RGB patches used to train neural network
train_mask	mask for train_im patches
test_im	RGB patches used to evaluate neural network
test_mask	mask for test_im patches
full images	full sized images used for evaluation