Adversarial Representation Learning for Medical Imaging

Domingues, José David Miranda Barreira

http://hdl.handle.net/10451/56666

Use this identifier to reference this record.

Name:	Description:	Size:	Format:
TM_José_Domingues.pdf		40.99 MB	Adobe PDF	Download

Send Feedback

Authors

Domingues, José David Miranda Barreira

Advisor(s)

Garcia, Nuno Ricardo da Cruz

Tomás, Helena Isabel Aidos Lopes

Abstract(s)

Breast cancer is a significant cause of death worldwide, especially among women, being one of the hottest topics in the medical area. In 2020, according to the World Health Organization, there were 2 million women diagnosed with this disease and 685.000 deaths globally. Thus, demonstrating the enormous impact that this disease has and hence the theme of this work being focused on breast cancer. Nowadays, most medical cases use CAD (Computer-Aided Diagnosis) systems in various ways to prevent and help doctors attenuate the impact of cancer by combining their expertise and the advanced technology we have today to perform various tasks. These learning-based systems use many high-quality datasets to extract and identify core aspects and execute multiple tasks. However, there is significant difficulty accessing these datasets because of data protection rules or even different data sharing policies, allied to the nonexistence of suitable enough public datasets and labelled data. Regarding this problem and the growing use of CADs systems in the breast cancer topic, this work proposes generating mammograms based on a single mammogram allowing health entities to generate their mammograms and, thus, a highquality dataset. With that goal, this project uses the base work of ConSinGAN to generate images based on a single one and an innovative way of gaining more image variability by using single image composition harmonisation. The results underwent a validation process, where the images’ quality, diversity and impact were analysed. In terms of real-life usage, there is still a long way to go since such images need to be validated by real doctors and generated at much higher resolutions. However, for now, it is already a significant step toward this purpose.