Table 3.

Advantages and limitations for deep learning methods in MRI denoising

Literature	Main contribution	Advantages	Limitations
Kidoh et al. [55]	Proposed a novel denoising method (dDLR)	•   Neuroradiologists’ assessments and experiments on a clinical dataset shows that dDLR outperforms other methods with respect to SSIM, PSNR	•   Relatively small testing cohort •   Experiments limited to T1, T2, FLAIR, and MPRAGE data
Zhu et al. [103]	Proposed a novel MR image denoising method called DESN which has a novel network architecture and a well-designed loss function	•   A novel loss function which data fidelity loss, image quality penalty and three loss terms: MSE, SSIM, entropy term •   Superior performance over SOTA in generating high-quality MR images with sufficient edge and texture information	•   Requirement of a large amount of training image data pairs to learn the clear image prior information
Christopher et al. [105]	Proposed a variant of ADMM-DIP method for enhancing single coil MR images	•   Achieved Rician noise removal from single MR image by utilizing the combined effect of MSE, KL divergence, and perceptual loss functions	•   Experiments are done only on simulated data •   Hyperparameter tuning is not optimized
Ran et al. [48]	Introduced a novel MRI denoising method based on the residual encoder–decoder Wasserstein GAN (RED-WGAN)	•   Novel loss function combining perpetual loss from VGG-19 network with MSE and adversarial losses •   Achieved superior performance over SOTA in simulated and clinical data •   Comparatively better computational cost	•   Experiments limited to brain data
Tian et al. [120]	Proposed a novel MRI image denoising method using the conditional GANs	•   Experiments conducted on both synthetic and real clinical MRI datasets •   Achieved high SSIM compared to SOTA in high noise levels	•   Experiments are incomprehensive and performed on limited data
Chauhan et al. [14]	Proposed a combined approach of fuzzy logic and a convolutional autoencoder on a brain MR images	•   The combined approach performs better than SOTA	•   Experiments conducted on limited data
Jiang et al. [107]	Proposed the Multi-channel DnCNN (MCDnCNN) method with two main training strategies to denoise images with and without a specific noise level	•   Comprehensive experiments conducted on public and clinical datasets •   Reported high PSNR and SSIM over SOTA •   Showed good generalizable applicability	•   Model incompatibility with 3D volumetric data •   Experiments confined to brain data
Tripathi et al. [108]	Propose a novel CNN-based denoiser called CNN-DMRI	•   End-to-end training scheme utilizing residual learning scheme •   Performance assessed qualitatively and quantitatively on simulated and real data •   Capability to denoise without losing crucial image details	•   Suboptimal computational time
Gregory et al. [109]	Proposed the HydraNet, a multi-branch deep neural network architecture that learns to denoise MR images at a multitude of noise level	•   Compatible with numerous factors such as pulse sequences, reconstruction methods, coil configurations, and physiological activities	•   Incompatible with volumetric denoising
Naseem et al. [111]	Proposed the Cross-Modality Guided Denoising Network (CMGDNet) for removing Rician noise in T1 data	•   Compatibility with cross-modal medical imaging •   Exploited complementary information existing in cross-modal images and improved the learning capability	•   Experiments limited to public datasets •   Experiments limited to brain data
Wu et al. [112]	Proposed a denoising method named 3D-Parallel-Rician Net, which combines global and local information to remove noise in MR images	•   Introduced a powerful dilated convolution residual (DCR) module to expand the receptive field of the network •   Introduced a depth wise separable convolution residual (DSCR) module to learn the channel and position information	•   Evaluated only on simulated T1 MR image data •   Requirement for high-quality noise-free ground-truth images
Singh et al. [113]	Proposed a noise filtering network which learns the image details from the image patches pixel-by-pixel from noise residuals to restore the detailed image features in an end-to-end feed-back approach	•   Showed comparable performance with SOTA without losing important image information	•   Insufficient number of experimentations
Tripathi et al. [115]	Proposed a dual path deep convolution network based on discriminative learning for denoising MR images	•   Incorporated depth wise separable convolution to denoise the images of different noise levels •   Yielded better performance as compared with various other networks •   Attained favorable assessments from radiologists	•   Experiments limited to public data and brain data
Yang et al. [118]	Proposed a hybrid regularization model from deep prior and low-rank prior. The local deep prior was explored by a fast flexible denoising convolutional neural network (FFDNet)	•   Compared with the popular CS-MRI approaches, the experimental results demonstrated better performance	•   Limited experiments
Moreno et al. [119]	Evaluated two unsupervised approaches for MRI denoising in the complex image space using k-space data: SURE and blind spot network	•   Methods are evaluated on real knee MRI and synthetic brain MRI data •   Both networks outperformed NLM and prove to be dependable denoising methods	•   Experiments limited to public datasets •   Incomprehensive experiments
Panda et al. [28]	Utilized perceptual loss and MSE for training a network for brain MRI denoising	•   Restored images were visually desirable and contained more anatomically refined features •   The proposed CNN network surpassed SOTA for Rician MRI denoising and obtained high quality brain MR images	•   Comparatively large computational cost for training •   Experiments on limited datasets