Denoising of multi b-value diffusion-weighted MR images using deep image prior

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Computational Imaging

Background:

• Diffusion-weighted magnetic resonance imaging (DW-MRI) is clinically valuable but suffers from low signal-to-noise ratio (SNR), especially at high b-values.
• Noise in DW-MRI data can compromise image quality and the accuracy of derived parameters, limiting its diagnostic utility.

Purpose of the Study:

• To introduce and evaluate a novel image denoising method for multiple b-value DW-MRI data using deep image prior (DIP).
• To address the SNR limitations of DW-MRI and improve the quality of intravoxel incoherent motion (IVIM) parametric images.

Main Methods:

• A deep image prior (DIP) based denoising approach was developed, utilizing high-quality prior images as network input and noisy DW images as output.
• The DIP model was trained using early stopping to learn image content and suppress noise, enabling simultaneous denoising across multiple b-values.
• Performance was assessed using simulated data (digital phantom with Rician noise) and real DW-MRI datasets, comparing DIP against local principal component analysis (LPCA).

Main Results:

• Simulations demonstrated that the DIP method significantly outperformed LPCA in reducing mean-squared error and improving parameter estimation.
• Evaluation on real DW-MRI data confirmed that the DIP method enhances the quality of IVIM parametric maps.
• The DIP approach effectively denoises multiple b-value DW-MRI data, preserving essential image content.

Conclusions:

• Deep image prior (DIP) is a feasible and effective method for denoising multiple b-value diffusion-weighted MRI data.
• The proposed DIP method offers improved image quality and parameter accuracy compared to traditional methods like LPCA.
• This technique holds promise for enhancing the clinical utility of DW-MRI by mitigating noise-related artifacts.