Table 4.

Advantages and limitations for deep learning methods in MRI resolution enhancement

Literature	Main contribution	Advantages	Limitations
Zhao et al. [144]	Reviewed the SMORE algorithm and demonstrated its potential use in both research and clinical scenario	•   SMORE showed improved visualization of brain white matter lesions in FLAIR images acquired from multiple sclerosis patient •   Showed improved visualization of scarring in cardiac left ventricular remodeling after myocardial infarction •   Showed improved performance in parcellation of the brain ventricular system •   Both visual and quantitative metrics of resolution enhancement are demonstrated	•   Limited methodology contribution; based on a previously introduced SR technique called SMORE
Pham et al. [145]	Studied deep 3D CNNs for the super-resolution of brain MRI	•   Investigated monomodal super-resolution in terms several factors: optimization methods, weight initialization, network depth, residual learning, filter size in convolution layers, number of the filters, training patch size, and number of training subjects •   Emphasized that one single network can efficiently manage multiple arbitrary scaling factors based on a multiscale training approach •   Extend super-resolution networks to the multimodal super-resolution using intramodality priors •   Investigate the impact of transfer learning skills onto super-resolution performance in terms of generalization among different datasets •   Learnt models were used to enhance real clinical low-resolution images
Chen et al. [146]	Introduced a novel 3D Densely Connected Super-Resolution Networks (DCSRN) to restore HR features of structural brain MR images	•   Experiments conducted on clinical data demonstrated superior performance over bicubic interpolation as well as other deep learning SOTA in restoring 4 × resolution-reduced images	•   Insufficient experiments and datasets utilized in the study
Sui et al. [42]	Developed a model to construct images with spatial resolution higher than can be practically obtained by direct Fourier encoding	•   Provided an estimate of the spatial gradient prior from the LR inputs for the HR reconstruction •   Incorporated the anisotropic acquisition schemes •   The model was trained over the LR images themselves only •   Developed a closed-form solution to the SRR model was developed to obtain the HR reconstruction •   Assessed performance on the simulated and clinical data •   Reported superior SRR over SOTA •   Obtained better images at lower or the same cost in scan time than direct HR acquisition	•   Experiments limited to brain data
Xue et al. [147]	Proposed the progressive sub-band residual learning SR network (PSR-SRN) which contained focused on missed high-frequency residuals as well as on reconstructing refined MR image	•   Introduced a brain-like mechanisms (in-depth supervision and local feedback mechanism) and progressive sub-band learning strategy to emphasize variant textures of MRI •   Illustrated superior performance over traditional and deep learning MRI SR methods	•   Experiments only on brain data
Li et al. [45]	Introduced DeepVolume, a two-step deep learning architecture to address the challenge of accurate thin-section MR image reconstruction	•   Extensive experiments illustrate that DeepVolume can produce SOTA reconstruction results by embedding more anatomical knowledge •   Practical and clinical value of DeepVolume is validated by applying the brain volume estimation and voxel-based morphometry •   Reliable brain volume estimation in the normalized space based on the thick-section MR images compared with SOTA	•   Experiments may be not dependable when SPM cannot provide accurate segmentation results
Shi et al. [148]	Proposed a novel residual learning-based SR algorithm for MRI, which combines multi-scale GRL and shallow network block-based LRL	•   Simulated and real MRI datasets used for evaluation •   Superior performance over SOTA CNN-based SR algorithms	•   High computational complexity •   Incompatible with 3D data
Shi et al. [149]	Proposed a progressive wide residual network with a fixed skip connection (FSCWRN)	•   Ability to relax the problems of feature degradation and diminishing feature reuse •   Experiments on real and simulated data show competitive performance over SOTA	•   Incompatible with 3D data
Kang et al. [153]	Presented a novel HR-MRI generation approach based on CNNs and multi-resolution analysis	•   Improved resolution T2 data with the prior information provided by HR T1 data •   Took advantage of structural similarity between the modalities •   Experiments on a real MRI dataset show improved over SOTA single- and multi-modal networks •   Effectively restore the edge details even with 4 × magnification	•   Experiments only on brain data
Dong et al. [154]	Presented MAU-Net for spatial resolution enhancement, based on the observation that multi-parametric MRI provide relevant priors for MRSI enhancement	•   Model trained on in vivo brain imaging data from patients with high-grade gliomas •   Combined loss function consisting of pixel, structural, and adversarial loss •   Ability to reconstruct high-quality metabolic maps with a high-resolution of 64 × 64 from a low-resolution of 16 × 16 with better performance over SOTA	•   Limited experiments and datasets utilized
Zhang et al. [156]	Proposed SOUP-GAN: Super-resolution Optimized Using Perceptual-tuned Generative Adversarial Network in order to produce thinner slices	•   Outperformed other conventional resolution-enhancement methods and previous SR work on medical images based on both qualitative and quantitative comparisons •   A novel 3D SR interpolation technique, providing potential applications for both clinical and research applications
Sui et al. [157]	Designed a deep architecture that utilizes an adversarial scheme with a generative neural network against its degradation counterparts	•   Achieved high-quality brain MRI at an isotropic resolution of 0.125 mm3 with 6 min of imaging time •   Experiments on simulated and clinical data •   Demonstrated superior SRR results SOTA deconvolution-based methods	•   Limited experiments and datasets
Jiang et al. [158]	Proposed the Fused Attentive Generative Adversarial Networks(FA-GAN) to generate SR MR image from LR images	•   Designed a global feature fusion module, including the channel attention module, the self-attention module, and the fusion operation in order to enhance the key features MR images •   Introduced the spectral normalization process to make the discriminator network stable during training	•   Relatively small training cohort •   Insufficient experiments and datasets
Chen et al. [160]	Proposed a novel 3D neural network design, namely, a multi-level densely connected super-resolution network (mDCSRN) with GAN-guided training	•   Trains and infers quickly relative to other SOTA •   Promotes realistic output •   Experiments on a real MR data show superior performance recovering 4 × resolution-downgraded images and runs 6 × faster compared with SOTA	•   Limited experiments and datasets