RD-DIP: Rician denoising deep image prior

Abstract

Rician denoising of magnetic resonance images (MRI) is a fundamental problem in medical image processing. Although variational methods can address Rician noise, the mathematical complexity of the underlying models makes them challenging to implement. Neural networks offer a compelling alternative but require extensive labelled datasets and are prone to introducing artifacts. This study proposes advanced tailored models for MRI Rician denoising within the unsupervised Deep Image Prior (DIP) framework, eliminating the need for additional data and reducing training-related artifacts. To our knowledge, this is the first application of the Rician maximum likelihood in the DIP framework. Our experiments, conducted on the Brainweb dataset and a set of real MRI scans, show the superiority of the proposed models over traditional and recent state-of-the-art unsupervised approaches. Additionally, we provide a detailed pipeline to ensure the reproducibility of our experiments. The code is available at https://github.com/heqro/rd-dip.