Generative Adversarial Network for Generation of Artificial Microwave Data for Stroke Detection

Abstract

This study aims to explore the possibilities of generating microwave data with a Generative Adversarial Network (GAN), in order to expand the existing data set and increase the performance of a stroke detection algorithm. Key challenges of the project relate to the small data set size and samples with many features. The generation of data was done with a Conditional Wasserstein Generative Adversarial Network. Due to the low data regime, the effects of adding DeLiGAN was also investigated. In addition to generating data with a GAN, this study also covers methods for the evaluation of generated data. To evaluate the quality of the generated data, a separate classifier network is utilised. Evaluation of the generated data in classification problems, as well as visualisation of distribution coverage, indicate that the data is of good quality and represent the distribution of original data well. However, results also show that the generated data cannot completely substitute the real data, and is deemed to be lacking in some quality measure. Still, the results are promising and the project concludes that it certainly is possible to generate microwave data which is to be used for stroke detection, with great potential for further improvements.