Go to MIDL 2026 Conference homepageAuto-N2N: Self-supervised MRI Denoising using Synthetic Pair Generation and Mixture of Experts
Keywords: MRI Denoising, Unsupervised Learning, Noise2Noise, Deep Learning, Mixture of Experts, Non-Local Means
TL;DR: A new self-supervised deep learning method for MRI denoising by using a single noisy image to generate training pairs via Non-Local Means (NLM) and employing a Mixture of Experts (MoE) to handle diverse noise levels.
Abstract: Magnetic Resonance Imaging (MRI) is the gold standard for neuroimaging, yet its acquisi-
tion process inherently introduces noise that degrades diagnostic quality and quantitative
analysis. Deep learning (DL) methods have achieved state-of-the-art performance, but suf-
fer from a critical limitation: they predominantly rely on supervised learning with clean
ground-truth data, which is clinically usually unavailable or unsupervised approaches, such
as Noise2Noise (N2N), requiring spatially aligned noisy image pairs. This paper presents a
new self-supervised method that eliminates the need for double scanning. We introduce a
synthetic data generation mechanism based on Non-Local Means (NLM) principles to create
training pairs from single volumes. Furthermore, we propose a Mixture of Experts (MoE)
framework in which each expert employs a 3D Residual CNN architecture, enabling the
system to handle the heterogeneous noise levels typically encountered in clinical settings.
Experiments on the Brainweb phantom and OASIS dataset demonstrate that our approach
significantly outperforms state-of-the-art methods, particularly in high-noise regimes, while
reducing inference time from minutes to seconds.
Primary Subject Area: Image Acquisition and Reconstruction
Secondary Subject Area: Unsupervised Learning and Representation Learning
Registration Requirement: Yes
Visa & Travel: Yes
Read CFP & Author Instructions: Yes
Originality Policy: Yes
Single-blind & Not Under Review Elsewhere: Yes
LLM Policy: Yes
Submission Number: 203
Loading