SNRAware: Improved Deep Learning MRI Denoising with SNR Unit Training and G-factor Map Augmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• MRI denoising is crucial for image quality.
• Current deep learning methods can be limited in performance and generalization.

Purpose of the Study:

• To develop and evaluate a novel deep learning MRI denoising method.
• Leverage quantitative noise distribution from MRI reconstruction for improved performance and generalization.

Main Methods:

• Trained 14 transformer and convolutional models using the SNRAware scheme on 2,885,236 cardiac cine images.
• SNRAware simulates synthetic datasets and provides quantitative noise distribution to models.
• Tested models on in-distribution and out-of-distribution datasets (real-time cine, perfusion, neuro, spine MRI).

Main Results:

• SNRAware training outperformed standard training across all 14 models.
• Transformer models showed superior performance over convolutional models.
• The best model generalized to diverse out-of-distribution scans, improving CNR by up to 6.5x.

Conclusions:

• The SNRAware training scheme effectively improves deep learning MRI denoising.
• This method enhances denoising performance and generalization capabilities across different MRI applications.