Augmenting Image Datasets with GANs

Introduction

While in space, astronauts are exposed to harmful cosmic radiation. By accurately categorizing radiation, scientists and engineers can design better shielding materials and develop appropriate countermeasures to mitigate the harmful effects of radiation on astronauts' health.

However, creating a comprehensive dataset for radiation classification is challenging due to the limited amount of available data. To overcome these limitations and improve the accuracy of the radiation classification model, the augmentation of the dataset with generated images can be a valuable approach.

Getting Started

Setting Up the Environment

The environment can be setup by following the instructions in setup/environments/README.md.

Downloading the Dataset

The data comes from the NASA BPS Microscopy Dataset which contains Fluorescence microscopy images of individual nuclei from mouse fibroblast cells, irradiated with Fe particles or X-rays. It is stored on an AWS S3 bucket and can be downloaded by running src/data_utils.py. The dose_Gy_specifier and hr_post_exposure_val variables can be changed to match the desired subset of the data.

Models

This repository contains a vanilla GAN model at src/gan.py and a ResNet-101 model at src/model/resnet101.py.

Credit

This was made by shenaniGANs (Brandon Huynh, Olivia Ih, Dennis Lustre, Sharon Ma).