Image inpainting methods for elimination of non-anatomical objects in medical images

Sammanfattning: This project studies the removal of non-anatomical objects from medical images. During tumor ablation procedures, the ablation probes appear in the image, hindering the performance of segmentation, registration, and dose computation algorithms. These algorithms can also be affected by artifacts and noise generated by body implants. Image inpainting methods allow the completion of the missing or distorted regions, generating realistic structures coherent with the rest of the image. During the last decade, the study of image inpainting methods has accelerated due to advances in deep learning and the increase in the consumption of multimedia content. Models applying generative adversarial networks have excelled at the task of image synthesis. However, there has not been much study done on medical image inpainting. In this project, a new inpainting method is proposed for recovering missing information from medical images. This method consists of a two-stage model, where a coarse network is followed by a refinement network, both of which are U-Nets. The refinement network is trained together with a discriminator, providing adversarial learning. The model is trained on a dataset of CT images of the liver and, in order the mimic the areas where information is missing, regular and irregular shaped masks are applied. The trained models are compared both quantitatively and qualitatively. Due to the lack of standards and accurate metrics in image inpainting tasks, results cannot be easily compared to current approaches. However, qualitative analysis of the inpainted images shows promising results. In addition, this project identifies the Frechet Inception Distance as a more valid metric than older metrics commonly used for evaluation of image inpainting models. In conclusion, this project provides an inpainting model for medical images, which could be used during tumor ablation procedures and for noise and artifact elimination. Future research could include implementing a 3D model to provide more coherent results for inpainting patients - a stack of images - instead of single images.