Table 5.
Data-Consistency Layers and Unrolled Networks-based MRI Reconstruction Models.
| Ref. | Year | Network | Contributions | Unsolved Challenges |
|---|---|---|---|---|
| [33] | 2018 | VN | Preserved essential features of MR images, including pathologies not present in the training dataset | Suffer from residual artifacts that are particularly evident in the axial sequences |
| [34] | 2018 | VN | Provided rapid reconstruction speed of approximately 0.2 s per section | Variation in reconstruction times based on hardware models, the use of constant regularizations, and the absence of fully sampled data |
| [76] | 2018 | MoDL | Achieved faster convergence per iteration using numerical optimization blocks for data-consistency and required less training data | Use of many conjugate gradient steps in data-discrepancy layers may lead to increased computational time, possibly reducing reconstruction speed |
| [77] | 2019 | PC-CNN | Improved image accuracy by enforcing data consistency and enhanced convergence | Computational complexity, data dependency, limited interpretability, and sensitivity to noise and artifacts |
| [78] | 2020 | jVN | Image quality was improved, and blurring was reduced through the learning of efficient regularizers | Generalization to unseen data or different acquisition scenarios |
| [79] | 2020 | DeepcomplexMRI | No sensitivity information calculation required for resolving aliasing and channel correlations | High acceleration factors can result in persistent blurriness in the reconstructed MRIs |
| [80] | 2020 | Dense-RNN | Showed potential for capturing long-range dependencies among image units | Does not completely address the slow convergence issue inherent in proximal gradient descent methods |
| [81] | 2020 | TVINet | Ensured data consistency and preserved the fine details in the reconstructed MRI | Time-consuming and computationally expensive hyperparameter tuning, lack of uncertainty quantification in deterministic predictions |
| [82] | 2020 | FlowVN | Achieved accurate reconstructions of pathological flow in a stenotic aorta within a short timeframe of 21 s | Large training data requirement, interpretability |
| [83] | 2022 | CNN & UNet | Enhanced unfolding structures without complexity increase, using an adaptively calculated noise parameter for improved reconstruction performance | Suffer from training instability, slow convergence, and limited explainability, which can hinder its practical applicability and interpretability |
| [84] | 2022 | DEMO | Efficiently removed CS-MRI artifacts, such as motion, zebra, and herringbone artifacts | High computational requirements, including GPUs, for training and inference |
| [85] | 2023 | DIRCN | Used long-range skip connections to improve gradient and information flow | Model trained on retrospective public domain data, needs to be tested on clinically valid prospective data |