This repository is the code for paper 'Surveillance of neonatal jaundice using sternal skin images via deep learning'. The paper is currently under review and has not been formally accepted.

Overview

This code uses convolutional neural network to estimate neonatal jaundice of sternal skin images.

Contents

imgs: 12 skin images of 4 newborns with labels ('imgs/labels.csv'), and the prediction results of these images ('imgs/results.csv').

regression.py: The train code.

test.py: The test code.

utils.py: Some functions.

System Requirements

Hardware Requirements

The proposed network requires a computer with GPU to run the convolutional neural network. For optimal performance, we recommend a computer with the following specs:

RAM: 16+ GB
GPU: NVIDIA 1080Ti

OS Requirements

The code has been tested on Linux operating systems.

Package dependencies

Users should install the following python packages prior to run the code:

python3.5

tensorflow-gpu

keras

pandas

numpy

opencv

Instructions

We have provided the trained weight of the proposed network in [model], you can download the weight and run 'test.py' directly to estimate the bilirubin value of images in folder 'imgs'. We also provide the prediction results of 'test.py' for these images in 'imgs/results.csv'.

The 'regression.py' is used to train the proposed network, but the training data is too large to provide, so 'regression.py' may not run in your computer.