Real-world unified denoising for multi-organ fast MRI: a large-scale prospective validation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accelerated Magnetic Resonance Imaging (MRI) reduces scan times but often increases image noise, compromising diagnostic quality.
• Current denoising methods struggle to maintain image fidelity and diagnostic reliability in accelerated MRI across diverse clinical data.

Purpose of the Study:

• To develop and validate a unified deep learning-based denoising model for multi-organ accelerated MRI.
• To improve image quality and diagnostic performance of accelerated MRI scans using a novel AI approach.

Main Methods:

• A deep learning model was trained on a large-scale dataset (148,930 pairs) of prospectively collected, real-world noisy and clean MRI images from multiple centers, vendors, organs, and protocols.
• The model was evaluated on independent test sets (20,143 pairs) and external cohorts (46,870 images), assessing performance against state-of-the-art methods and through downstream tissue segmentation tasks.
• Clinical utility was assessed via blinded evaluations by radiologists on image quality, diagnostic confidence, and disease diagnosis.

Main Results:

• The unified deep learning model significantly outperformed existing denoising methods on real-world accelerated MRI data.
• Tissue segmentation accuracy improved by 7.05% (Dice score) across multiple organs when using denoised images compared to noisy images.
• The model demonstrated robust generalization across diverse scanners, protocols, and independent clinical cohorts, maintaining high visual fidelity and diagnostic performance equivalent to clean images, even at 3x acceleration.

Conclusions:

• A unified deep learning denoising model can effectively enhance the quality and diagnostic reliability of accelerated MRI across various clinical scenarios.
• This AI-driven approach facilitates faster MRI acquisition times (under one minute) without sacrificing diagnostic performance, holding significant potential for widespread clinical adoption.
• The model's robustness and generalization capabilities suggest it can be broadly applied to improve clinical MRI workflows.