Deep learning for image segmentation, registration, and imputation using limited amounts of labeled data

Abstract

Deep learning technologies for image processing have been widely studied to detect objects, restore image quality, etc. However, models with high performance typically require large amounts of labeled data, which can be difficult to obtain in practice. To address this, recently, deep learning methods using limited amounts of labeled data have been developed. Regarding this, I developed deep learning methods for various image processing tasks using unlabeled or small amounts of labeled data. First, I proposed a semi-/un-supervised learning-based image segmentation method using a novel loss function that takes images without any labels. Also, I presented an unsupervised multiscale image registration method based on the cycle consistent model to handle high-resolution images while preserving topology. In addition, I designed a diffusion-based unsupervised image registration method that provides intermediate deformations between source and target images, leading to deforming images with a less folding problem. Moreover, I proposed a missing image imputation method using the diffusion model, which generates missing data without losing the identity of observations. Lastly, I presented a diffusion adversarial representation learning model that synthesizes semantic images and thereby learns image segmentation in a self-supervised manner. These learning-based image processing methods provide desired outputs in real time and can be useful in various imaging fields.