High-field mr diffusion-weighted image denoising using a joint denoising convolutional neural network.

Area of Science:

• Medical Imaging
• Neuroimaging
• Diffusion-Weighted Imaging

Background:

• Low signal-to-noise ratio (SNR) limits high-resolution diffusion-weighted imaging (DWI).
• Conventional denoising methods have limitations in feature extraction and parameter tuning.
• Existing convolutional neural network (CNN) models do not address multichannel DWI denoising.

Purpose of the Study:

• To introduce a joint denoising CNN (JD-CNN) model.
• To enhance the SNR of multiple b-value DWI.

Main Methods:

• Developed a multichannel CNN model (JD-CNN) for joint denoising of multiple b-value DWI.
• Evaluated performance against Total Variation (TV) and BM3D denoising methods.
• Utilized 11.7T MRI data from healthy rats and rats with focal cortical dysplasia.

Main Results:

• JD-CNN significantly improved image quality, with Peak SNR (PSNR) increasing from 23.15 ± 1.77 to 42.94 ± 2.12.
• JD-CNN outperformed TV and BM3D on high b-value DWIs (PSNR: 46.52 ± 0.98).
• Reduced normalized mean square error (NMSE) of apparent diffusion coefficient (ADC) from 0.72 ± 0.13 to 0.45 ± 0.06 (P < 0.01).

Conclusions:

• The JD-CNN model effectively removes noise across various levels in multiple b-value DWI.
• Improved diffusion parameter estimation demonstrates the method's potential clinical utility.
• This approach shows promise for advancing DWI applications.