abstract: The Covid-19 pandemic is causing worldwide repercussions, with early and rapid detection of the disease being crucial in reducing its spread and mortality rate. Radiological findings can be used to detect the disease, and multiple machine learning models so far have learnt to recognize Covid-19 from chest x-rays with good results. One major obstacle in improving the accuracy and generality of these models is the small size of the available datasets. In this paper we employ StyleGan2 to generate synthetic images for the Covid-19 label of the Covidx5 dataset and then fine-tune an AlexNet model to classify images into three labels: Covid-19, Pneumonia and Normal. Three models are built for comparison: one on the base dataset, one on the augmented dataset and one the dataset with GAN-generated images. Results show that with a Frechet Inception Distance score of 73.471, the images generated by the GAN perform worse than the base and augmented images in multilabel classification. When comparing performance for the Covid-19 label we report ROC-AUC scores of 89.95, 87.06 and 84.26 for the Base, Augmented and GAN-based model respectively, indicating the generated images were not representative of the Covid-19 label. We hope future research into different GAN models can improve on our results.
Giacomo Fiorentini, g.fiorentini@students.uu.nl Maria Galanty, m.galanty@students.uu.nl Luca Lin, l.lin1@students.uu.nl Jakub Myśliwiec, j.mysliwiec@students.uu.nl
link to Covidx5 repository: https://github.com/lindawangg/COVID-Net
link for preprocessed dataset in .rar format: https://drive.google.com/file/d/1SA9jrKHHbxbtwwvruRluhq71sKNUIsjH/view?usp=sharing
with GAN images included in training set: https://drive.google.com/file/d/12Go7M6eYJcfsDFsaujpResF0ZOpgrVaw/view?usp=sharing
network snapshot 3360 for generating StyleGan2 images: https://drive.google.com/file/d/170VuXn3L1akfqAO5FYBZhJLSU3c3YM2k/view?usp=sharing