Table 8.

Contributions and Limitations of GAN-Based MRI Reconstruction Models.

Ref.	Year	Contributions	Unsolved Challenges
[35]	2018	Under 100 ms, a 256 × 256 MRI can be reconstructed with high quality (over 42 dB in average at 40% sampling rate)	Limited generalization to unseen data and imaging conditions, data scarcity and the need for large amounts of labeled training data
[36]	2019	Reduced motion artifacts and motion blurring consistently by retrospectively correcting MR images with simulated motion	reconstructed MRIs still have a certain amount of smoothness
[37]	2020	Achieved high acceleration factors, successfully recovered pathologies, and could jointly reconstruct and synthesize the target contrast	Large paired datasets are required for training, and further optimization and generalization are necessary to handle diverse multi-contrast imaging scenarios
[128]	2020	Reduced training time and improved network training stability and network generalization	May not fully capture the clinical significance of the phase information
[121]	2021	Balanced edge features against global high-level features for improved reconstruction accuracy	Lack of interpretability and explainability as black box models
[130]	2021	Reconstructed finer MRI texture details and effectively removed artifacts, all while utilizing fewer model parameters	Need to evaluate the generalizability and robustness of the approach across various imaging conditions
[131]	2022	For real-valued activations, a learnable complex-valued activation was developed to solve the transferability issues	Prone to overfitting, lack of interpretability and explainability as black box models
[132]	2022	Captured global context, recovered fine-structural details, and had low model complexity with improved learning behavior	Reliance on fully-supervised training with high-quality datasets, which could be challenging to compile, and the potential challenges in generalizing the model to nonrectilinear orientations
[133]	2023	Efficiently captured both long-distance dependencies and local information	High hardware requirements are associated with the increasing network parameters
[134]	2023	Improved spatiotemporal information was achieved between adjacent views, with a specific focus on reconstructing the local cardiac regions	Not suitable for multi-coil data and requires a large number of network parameters
[135]	2023	Used overall and regional perspectives to remove noise and restore the fine details	Limited generalization to unseen data and imaging conditions, high computational requirements