Lesion filling in magnetic resonance images based on generative adversarial networksBERTONCELJ, MOJCA (Avtor) Špiclin, Žiga (Mentor) lesion fillingmagnetic resonanceimage analysisdeep neural networksgenerative adversarial networksPathological brain lesions may adversely influence the process of magnetic resonance (MR) brain image analysis. For instance, they lead to unreliable and poor brain segmentation and corrupt normal brain tissue volume measurements. In this thesis, we aim to develop and test a novel approach for automated lesion filling using recent deep learning techniques. Based on input T1-weighted MR brain image with lesions we demonstrate the ability of a modified context conditional generative adversarial network (CCGAN) to reproduce normal appearing brain MR images. The original CCGAN was modified by incorporating Wasserstein loss with gradient penalty. The network was able to seamlessly fill a masked-out area of the image without any prior knowledge of the size or location of the mask. We also tested partial convolution networks, which requires explicit knowledge about the location and size of mask, and found they yield comparable results for lesion filling or inpainting. The obtained promising results show that deep learning techniques can successfully address the problem of lesion filling. We hope that this work will facilitate further research and adoption of deep neural networks in health care applications and thereby encourage cooperation between deep learning researchers and clinicians in the healthcare institutes.20192019-02-26 14:05:02Magistrsko delo/naloga106473VisID: 43281sl