MRI denoising with a non-blind deep complex-valued convolutional neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• High signal-to-noise ratio (SNR) in MR images is crucial for accurate diagnosis.
• Existing MRI denoising methods often overlook phase information, limiting their effectiveness.
• Developing advanced denoising techniques is essential for improving MRI diagnostic capabilities.

Purpose of the Study:

• To design and implement a complex-valued convolutional neural network (CNN) for MRI denoising.
• To leverage both magnitude and phase information for superior denoising performance.
• To evaluate the efficacy of the proposed method on simulated and real-world low-field MRI data.

Main Methods:

• A complex-valued CNN incorporating a noise level map (non-blind DnCNN) was developed.
• The network was trained using pairs of ground truth and simulated noise-corrupted MR images.
• Performance was quantitatively and qualitatively assessed against real-valued CNNs and other algorithms.

Main Results:

• Complex-valued models demonstrated superior performance in key metrics like normalized root-mean-square error and peak SNR.
• The non-blind DnCNN effectively handled spatially varying parallel imaging noise.
• Significant improvements in SNR and visual quality were observed for in vivo low-field MRI data.

Conclusions:

• The proposed non-blind DnCNN offers an efficient and effective solution for MRI denoising.
• This represents the first application of non-blind DnCNN in medical imaging.
• The method has the potential to enhance low-field MRI, improving diagnostics in resource-limited settings.