Título: Continual learning in chest radiographs of active tuberculosis and pneumonia using images generated by adversarial networks
Autores: Regina Reis da Costa Alves, Frederico Caetano Jandre de Assis Tavares, Jose Manoel de Seixas, Otto Tavares Nascimento, Joao Victor da Fonseca Pinto, Anete Trajman
Resumo: Lower respiratory infections, including tuberculosis (TB) and pneumonia, rank among the top 10 leading causes of death worldwide. Chest radiographs (CXRs) are recommended as a screening and triage tool and computer-aided detection (CAD) softwares are an alternative to analyzing CXR. Continual learning (CL) is an option to obtain models that can identify multiple diseases by continuously learning a diverse range of radiological signs associated with each disease. In this work, we tested a CL model, Learning Without Forgetting, in learning pneumonia and TB detection using synthetic images of TB to enlarge the dataset, produced by two different Generative Adversarial Networks (GANs) and incorporated in the training process using different approaches. After learning TB detection, the models performance in pneumonia detection has improved. Also, a potential improvement in TB detection was observed when synthetic data was used to fine tune the fully-connected layers of the model
Palavras-chave: nan
Páginas: 8
Código DOI: 10.21528/CBIC2023-133
Artigo em pdf: CBIC_2023_paper133.pdf
Arquivo BibTeX: CBIC_2023_133.bib