FONDUE: Robust resolution-invariant denoising of MR images using Nested UNets

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• High-resolution magnetic resonance imaging (MRI) enhances neuroanatomical detail but increases noise.
• Increased noise in structural brain MRIs can negatively impact segmentation and morphometric analysis.
• Current denoising methods may struggle with the diverse resolutions and parameters of modern MRI acquisitions.

Purpose of the Study:

• To introduce a novel, fast, robust, and resolution-invariant deep learning method for denoising structural human brain MRIs.
• To evaluate the proposed method's performance across various field strengths, voxel sizes, and scanner vendors.
• To demonstrate the method's effectiveness on both healthy and diseased participants across a wide age range.

Main Methods:

• Development of a deep learning model named Fast-Optimized Network for Denoising through residual Unified Ensembles (FONDUE).
• Exploration of denoising T1-weighted brain images acquired at different field strengths (1.5T-7T) and voxel sizes (1.2 mm-250 µm).
• Testing on data from multiple scanner vendors (Siemens, GE, Phillips) and diverse participant groups (healthy/diseased, various ages).

Main Results:

• FONDUE demonstrated stable denoising performance across multiple resolutions, comparable or superior to state-of-the-art methods.
• The method was significantly faster (orders of magnitude) and more cost-effective on a Graphics Processing Unit (GPU).
• FONDUE achieved top performance on at least one metric across all test datasets, indicating strong generalization and stability.

Conclusions:

• FONDUE offers high-quality, robust, and fast structural MRI denoising suitable for large-cohort studies.
• Its efficiency and low GPU memory requirements make it widely applicable.
• The open-source availability of FONDUE facilitates its broad adoption in neuroimaging research.