Table 1.

Summary information of various MRI post-processing tasks

Task	Motion correction	Instrumental artefacts correction	Sequence artefacts correction	Multi-site normalization	Noise reduction	Noise reduction	Super-resolution
Scopes	Rigid motion, non-rigid motion	Gibbs ringing, B0 inhomogeneity	EPI ghosting and distortion	Multi-site normalization	Anatomical	fMRI, DWI, perfusion, ASL	Anatomical, DWI, spectroscopic
Encoder types	2D Conv	2D Conv, 3D Conv	2D Conv, 3D Conv	2D Conv, 3D Conv	2D Conv, 3D Conv	1D Conv, 2D Conv, 3D Conv, LSTM, MLP	2D Conv, 3D Conv, MLP
Model types	CNN, Inception, ResNet, FCN, VAE, U-Net	CNN, U-Net, ResNet50	AE, U-Net	CNN, U-Net, CNN with attention, VAE, ResNet	CNN, AE	CNN, U-Net, FCNN, transformer	CNN, U-Net, AE
Training types	Fully supervised, adversarial	Fully supervised, adversarial	Fully supervised, adversarial	Fully supervised, adversarial	Fully supervised, adversarial	Fully supervised, adversarial	Fully supervised, adversarial
Loss	MAE, MSE, adversarial loss	MAE, MSE, adversarial loss	MAE, MSE, SSIM loss, gradient loss	MAE, perceptual loss	MAE, MSE, perceptual loss	MAE, MSE, SSIM loss	MAE, MSE, adversarial loss, SSIM loss. triplet loss
Input types	2D image, 2D patch, 3D patch	2D image	2D image, 2D patch	2D image	2D image, 2D patch	2D image, 3D image, fMRI time series, diffusion-weighted data	2D image, 2D patch, 3D patch
Inputs	Images/patches with simulated motion artefacts	Images/patches with simulated Gibbs artefacts or simulated phase errors	Images/patches with ghosting and distortion	Images acquired on different scanners or with different parameters	Images/patches with simulated Gaussian/Poisson/Rician noise	Noisy images and fMRI, undersampled DWI data	Low-resolution images
Outputs	Motion-free images/patches	Artefact-free images/patches	Artefact-free images/patches	Labels for the task of interest	Noise-free image/patch	Noise-free images and fMRI, fully sampled DWI data	High-resolution images
Training data source	Motion artefacts are simulated by adding phase error in k-space	Gibbs ringing artefacts are simulated with cropped kspace, created bias fields for simulating B0 inhomogeneity	Artefact-free images are generated by post-processing steps	Multi-site data acquisitions	Simulation by adding noises to the clean image to generate noisy training samples	Noise-free data is generated by applying post-processing steps	Simulation by undersampling k-space to create low-resolution training samples
Anatomical regions	Brain, liver, abdomen, pelvis	Brain, knee, respiration	Brain	Brain	Brain, knee	Brain, prostate, whole body	Brain, prostate, knee, fetal, cardiac, torso
Metrics	SSIM, RMSE, MMI	SSIM, PSNR, RMSE, HFEN, GSR	GSR, MSE, Mutual Information	MAE, DICE,	SSIM, PSNR, IFC	SSIM, MAE, MSE, RMSE, PSNR	SSIM, PSNR, RMSE, DICE
Examples	MocoNet [19], MedGan [17], NAMER [49]	GRA-CNN [50], InHomoNet [51], DeepResp [34]	Synb0-DisCo [52], S-Net [53]	AD2A [54], DeepHarmony [47]	dDLR [55], RED-WGAN [48], DABN [56]	DeNN [31], 3DConv-LSTM [57], DNIF [24], STFNet [58]	DeepVolume [45], SRGAN [59], SR-q-DL [46]