Deep learning for medical image restoration

Image restoration refers to the process of inspecting a degraded image and recovering the underlying artifact-free counterpart through discarding the artifacts. Medical image restoration is a crucial part of automating medical image analysis and is the crux of many high-risk applications such as computer-aided diagnosis, and computer-aided diagnosis, therapy planning and delivery, and computer-aided intervention. In this thesis, we study and develop approaches to restore artifact-free medical images from their corrupted counterparts using deep learning models. This thesis begins with a thorough and comprehensive examination of the recent advances in image restoration based on deep learning both in natural and medical imaging. The thesis then proposes five novel approaches to image restoration by leveraging deep neural networks. The first two contributions target image super-resolution of confocal microscopy images using supervised training schemes. In particular, we firstly make the first practical effort to frame the image restoration problem in miniaturized confocal laser endomicroscopy devices as a classic deep learning based image super-resolution task. Secondly, we focus on improving the learning capacity of neural networks in a limited computational budget. In the next two works, we turn our attention to the problem of image denoising in fluorescence microscopy imaging. First, we propose a neural network model for disentangling the signal and noise components of an input noisy image, without the need for any ground truth training data. The second contribution centres around proposing a non-local patch-wise Bayesian mean filtering in the context of neural networks. Finally, the thesis tackles the problem of joint attenuation and scatter correction in positron emission tomography reconstruction to remove visual artifacts and quantitative errors. The Thesis concludes with a discussion of the limitations of the proposed models and important directions for future research.