Título: Application of Generative Adversarial Networks for Synthetic COVID-19 Ultrasound Data Generation
Autores: Pedro S. T. F. Silva, Antonio M. F. L. M. de Sá, Leonardo B. Felix, Wagner C. A. Pereira, Jose M. Seixas
Resumo: Lung ultrasound emerges as a powerful tool for the diagnosis of COVID-19, being a very cost-effective option to other modalities of exams, such as computerized tomography and Xray imaging. There are efforts in trying to employ deep learning to develop systems that can make an automatic diagnosis based on ultrasound exams to assist the medical decision, but they are limited by the amount of data available. The present work tackles this problem by proposing a method using generative adversarial models to create synthetic data and increase the volume of data to train more complex models. To evaluate whether the synthetic data presents a variety close to that of the original data without replicating training samples, it was devised applications of the Kullback-Leiber divergence and L1 norm. Results indicate that the generated data sampled the main features of the ultrasound data, presenting a variety close to the original data. This points to the possibility of using the proposed method as a means to overcome the problem of low data volume for lung ultrasound. which is shown using expert knowledge
Palavras-chave: deep learning, generative adversarial networks, lung ultrasound, COVID-19
Páginas: 7
Código DOI: 10.21528/CBIC2023-033
Artigo em pdf: CBIC_2023_paper033.pdf
Arquivo BibTeX: CBIC_2023_033.bib