Automatic Cone Photoreceptor Localisation with MDRNNs

This repo contains an implementation of the method described in this paper. Please cite the paper if you use the code.

@Article{Davidson2018,
    author={Davidson, Benjamin
    and Kalitzeos, Angelos
    and Carroll, Joseph
    and Dubra, Alfredo
    and Ourselin, Sebastien
    and Michaelides, Michel
    and Bergeles, Christos},
    title={Automatic Cone Photoreceptor Localisation in Healthy and Stargardt Afflicted Retinas Using Deep Learning},
    journal={Scientific Reports},
    year={2018},
    volume={8},
    number={1},
    pages={7911},
    issn={2045-2322},
    doi={10.1038/s41598-018-26350-3},
    url={https://doi.org/10.1038/s41598-018-26350-3}
}

Getting Started

To install and use requires:

• Python 3.5.x or 3.6.x
• pip

Installing

1. Download the git repository to a folder of your choice, /path/to/code/ConeDetector

2. Install Python package using pip. Ubuntu: pip install /path/to/code/ConeDetector; Windowspython -m pip install /path/to/code/ConeDetector

3. • If you do not have a gpu, pip install tensorflow: Ubuntupip install tensorflow; Windowspython -m pip install tensorflow
   • If you do have a gpu, follow these instructions to install tensorflow-gpu

If you just want to apply the model from the paper, you only need tensorflow, not tensorflow-gpu. The gpu version is needed if you want to train new models in any reassonable amount of time.

Using

• Any images should be of the form, where xxxx is a number with leading zeros, eg 1==0001

INITIAL_XXXX_WHATEVER.tif

• The required lut.csv for applying models should be of the following form, if we have two subjects, for example, with a um to pixel of 0.76 and 0.85 respectively.

    INITIAL_0001, 0.76
    INITIAL_0002, 0.85

• To run the code open a cmd prompt, or terminal and enter:

cone_detector

Features

After running cone_detector from a terminal a gui will launch asking what you want to do.

Apply existing models

• Required: folder of tifs, lut.csv for each subject in folder
• Applies model to tifs to estimate locations
• Can simply trust the algorithm, or manually correct each image
• Outputs locations and stats for each image

Build training data sets for training new models

• Required: folder of tifs
• Create labeled data in format used by tensorflow to train new models
• Can select a model to aid the annotations, or do completely by hand
• Will save data set as tfrecord, to train new models

Train new models

• Required: training data set built using cone_detector
• Required: a validation data set created using cone_detector
• Will run same training regime described in the paper
• Saves new model, which can be applied in cone_detector