Fig. 1

The developmental history of DWI and its derivative models

Table 1

Classification of noise mechanisms and artifacts in dMRI

Fig. 2

Summary of traditional dMRI denoising methods

Fig. 3

Dual-channel input module in the DnCNN framework. Abbreviations: Conv, Convolution; BN, Batch Normalization; ReLU, Rectified Linear Unit; MSE, Mean Squared Error

Fig. 4

Dual-branch network architecture integrating DWI and T1-weighted images[35]. Abbreviations: CSF, cerebrospinal fluid; GM, gray matter; WM, white matter; SD-LSTM, spatial-directional long short-term memory; FC, fully connected; AD, axial diffusivity; Vic, ICVF; Viso, isotropic volume fraction; OD, Orientation Dispersion Index

Fig. 5

Schematic diagram of the 1D-CNN denoising network architecture. The network takes a one-dimensional sequence composed of N highly noisy voxels at the same spatial location as input. It sequentially undergoes two layers of one-dimensional convolution (Conv) and Pooling operations for feature extraction. The first convolutional layer has 16 channels and a kernel size of 1; the second convolutional layer has 32 channels and a kernel size of 8. After feature Flatten, signal reconstruction is performed via a dense connected layer (Dense), with root mean squared error (RMSE) loss serving as the loss function for supervised training

Fig. 6

Architecture diagram of the generative network. Abbreviations: Conv3D, 3D Convolution; BatchNorm3D, 3D Batch Normalization; LeakyReLU, Leaky Rectified Linear Unit

Fig. 7

Architecture diagram of the discriminator network. Abbreviations: Conv3D, 3D Convolution; LeakyReLU, Leaky Rectified Linear Unit; FC, Fully Connected

Table 2

Summary of deep learning-based denoising methods for dMRI

Table 3

Comparison of SSIM metrics for different denoising methods at various noise levels

Table 4

Comparison of PSNR metrics for different denoising methods at various noise levels

Table 5

Comparison of RMSE metrics for different denoising methods at various noise levels

Fig. 8

Visual comparison of different denoising methods at a 6% noise level. From top to bottom, images under conditions of b=0, b=1 000 s/mm², and b=2 000 s/mm² are shown to reflect the performance differences of each method at varying diffusion sensitivities

Fig. 9

Comparison of FA maps (Row 1), color-coded FA maps (Row 2), and local details (Row 3) across different denoising methods at a 4% noise level (b=1 000 s/mm²)

Table 6

Quantitative evaluation of denoising performance on DTI microstructural parameters. Values represent the voxel-wise MAPE of microstructural metrics (FA, MD, AD, RD) relative to the ground truth. Lower values indicate better preservation of microstructural details. Bold values indicate the best performance for each noise level