Patch-to-Image Fully Convolutional Networks training for Retinal image segmentation

This repository contains the IPython code for the paper

Taibou Birgui Sekou, Moncef Hidane, Julien Olivier and Hubert Cardot. From Patch to Image Segmentation using Fully Convolutional Networks - Application to Retinal Images.

Given a retinal image database and a fully convolutional network (FCN) f, this tool first pre-trains it on an on-the-fly generated patch database, then fine-tunes it on the original full-sized images.

Setup

Environment: The following software/libraries are needed:

• TensorFlow 1.12.0
• numpy 1.15.1
• h5py 2.7.0
• matplotlib 2.0.2
• scikit-image 0.14.1
• scikit-learn *.19.1

Datasets: The following datasets are used in our experiments:

• DRIVE
• STARE
• CHASE_DB1
• IDRiD

Data preprocessing: All the images are preprocessed using:

• Gray scale conversion
• Gamma correction (with gamma=1.7)
• CLAHE normalization