Exploring Generative Adversarial Networks to Improve Prostate Segmentation on MRI

Abstract

Prostate cancer is the second most occurring cancer and the sixth leading cause of cancer death among men worldwide. The number of cases is expected to increase dramatically due to population growth and increased expected lifetime. The magnetic resonance imaging (MRI) examination is an essential and a comfortable tool towards a precise diagnosis at an early stage. The examination method is already used at several hospitals, but its effective use depends on the expertise of clinical personnel.

This thesis will explore how generative adversarial networks can improve prostate segmentation on MRI. Different architecture within the topic of deep learning have proven to be accurate in biomedical image segmentation. However, it depends on a large volume of training data that is hard to obtain due to privacy policy. This thesis investigates the possibilities for generating new anonymized training data to improve biomedical image segmentation.

The final results improve the segmentation score compared to just using the original data. An underperforming segmentation network limits the segmentation results compared to other networks using the same data, but present the potential for expanding the dataset using generated data and improve the segmentation results.