Table 4.
Comparison of encoder and decoder-based models.
| Ref. | Year | Network | Contributions | Unsolved Challenges |
|---|---|---|---|---|
| [63] | 2019 | DISN | Improved MRI reconstruction quality and robustness to misregistration errors | Limited generalization to unseen data and imaging conditions, lack of interpretability and explainability as black box models |
| [64] | 2019 | VDDCN | Made the network easy to train using dense connections and alleviated gradient-vanishing problem | Limited generalization to unseen data and imaging conditions, data scarcity and the need for large amounts of labeled training data |
| [65] | 2019 | IFR-Net | Improved network capacity with better feature refinement and fully learned parameters | Limited generalization to unseen data and imaging conditions, prone to overfitting |
| [66] | 2020 | DECN | Reduced structural reconstruction errors and improved MRI quality | Lack of interpretability and explainability as black-box models, may lead to artifacts and noise |
| [67] | 2020 | NISTAD | Reduced reconstruction time, simplified hyperparameter tuning, and a simpler network architecture with fewer parameters | Not efficient for highly undersampled image sequence reconstruction and might not be realistic enough for real clinical scans |
| [68] | 2021 | X-net & Y-net | Reduced number of trainable parameters, leading to a more efficient and streamlined model architecture | Lower computational efficiency due to the incorporation of additional network branches and the increased complexity of the model |
| [70] | 2022 | DFCN | Reconstruction quality improved by eliminating aliasing effects utilizing correlation information between adjacent slices | Time-consuming and computationally expensive hyperparameter tuning, may lead to artifacts and noise |
| [71] | 2022 | HIWDNet | Achieved accurate cross-domain MRI reconstruction by leveraging image and wavelet domains. Efficiently reconstructed the structure while removing aliasing artifacts. | The complex architecture and intricate interactions of HIWDNet may hinder interpretability |
| [72] | 2023 | DSMENet | Enhanced detail and structure information, adapted to diverse MRI scenarios, and offered improved visual effects and generalization. Proved to be a competitive candidate for real-time MRI applications | Complex architecture and intricate interactions of DSMENet limit its interpretability |
| [73] | 2023 | SCU-Net | Achieved superior deghosting performance even at high acceleration factors, leading to high-quality complex MRIs | Relied on sparsified complex data and required further investigation into its effectiveness in handling complex anatomical structures and capturing fine details in highly undersampled MRI data |
| [61] | 2023 | RNLFNet | Effectively captured long-range spatial dependencies in the frequency domain, leading to enhanced MRI reconstruction | May have limitations when applied to parallel MRI and dynamic MRI |
| [74] | 2023 | GFN | Maintain more detailed MR images by capturing edge structures in gradient images | Lack of interpretability and explainability as black-box models, may lead to artifacts and noise |