isi-segmentation 0.1.0

Supervised ISI segmentaion using tensorflow

Welcome!

This is a repository for segmenting visual cortex areas for sign map. The model was trained on about 2000 isi-experiment data using UNet and TensorFlow.

The sign map will be segmented into different regions and 14 cortex areas could be identified. The output label map will be saved as '.png' file with different values (i.e., 1, 2, 3 ...) corresponding to different visual cortex areas (i.e., VISp, VISam, VISal ...). The class definition is as follows:

Class	acronym	name
1	VISp	Primary visual area
2	VISam	Anteromedial visual area
3	VISal	Anterolateral visual area
4	VISl	Lateral visual area
5	VISrl	Rostrolateral visual area
6	VISpl	Posterolateral visual area
7	VISpm	posteromedial visual area
8	VISli	Laterointermediate area
9	VISpor	Postrhinal area
10	VISrll	Rostrolateral lateral visual area
11	VISlla	Laterolateral anterior visual area
12	VISmma	Mediomedial anterior visual area
13	VISmmp	Mediomedial posterior visual area
14	VISm	Medial visual area

Installation

To use isi-segmentation library, clone this repository and install the requirements.

1. First, ensure git is installed:

git --version

If git is not recognized, install git.

2. Move into the directory where you want to place the repository folder, and then download it from GitHub:

cd <SOME_FOLDER>
git clone https://github.com/AllenNeuralDynamics/isi_segmentation.git
cd isi_segmentation
pip install -e .

The script should take four inputs:

• hdf5_path (str): the hdf5 file which contains the sign map
• sign_map_path (str): the sign map extracted from .hdf5 file for prediction
• label_map_path (str): the output label map for the given sign map
• model_path (str): path to trained model (to download it, follow here)

Download trained model

mkdir -p model
gdown 'https://drive.google.com/uc?id=1X5C0avuOcjnbZDcS0hG6yujd2bY1hrK1' -O ./model/isi_segmentation_model.h5

Usage

To predict the label map for the sample sign map with the download model, run:

python run_predict.py \
    --hdf5_path ./sample_data/661511116_372583_20180207_processed.hdf5\
    --sign_map_path ./sample_data/661511116_372583_20180207_sign_map.jpg\
    --label_map_path ./sample_data/661511116_372583_20180207_label_map.png\
    --model_path ./model/isi_segmentation_model.h5\
    --plot_segmentation True

Or you could directly run

sh run.sh

Please make sure you have already downloaded the trained model (follow here) and update model_path.

Model output directory structure

After running prediction, a directory will be created with the following structure

    /path/to/outputs/
      ├── <experiment_name>.png
      └── <experiment_name>_visualize.png

• <experiment_name>.png: prediction from the sign map, the filename is set to label_map_path

• <experiment_name>_visualize.png: visualize the sign map and its resulting label map if plot_segmentation is set to true.

An example of isi segmentation outputs is ./sample_data/

Visualization

If you would like to visualize the output label map, set plot_segmentation to True. The plot will be saved as <experiment_name>_visualize.png and stored in the same folder as the label map.