Table 2.

Comparative Analysis of Input domain-based MRI Reconstruction Models.

Ref.	Year	Input Domain	Contributions	Unsolved Challenges
[46]	2018	Dual	Implemented KIKI-net, a cross-domain CNN that operates sequentially on k-space, image, k-space, and image to achieve better image reconstruction and minimize aliasing artifacts	Increasing noise levels in data potentially lead to blurred output images and lower PSNR, affecting performance
[9]	2019	k-space	Developed RAKI, a k-space method for non-linear reconstruction of undersampled data from autocalibration signal data, using subject-specific neural networks without extensive training databases	CNN architecture heuristically selected, performance may vary with different network parameters; fixed learning rates in gradient descent algorithm may not be optimal for all applications
[4]	2020	Dual	A dual domain recurrent network was developed to restore both the image and k-space domains, with an embedded T1 prior for enhanced restoration quality	Limited generalization to unseen data and imaging conditions, data scarcity, and the need for large amounts of labeled training data
[47]	2020	Dual	Implemented MRI dual-domain reconstruction network (MD-Recon-Net) to explore the latent relationship between k-space and spatial data	Limited generalization to unseen data and imaging conditions
[48]	2021	k-Space	Employed a residual encoder-decoder network with self-attention layers to achieve adaptive focus and enhance interpolation performance	Potential sensitivity to variations in acquisition parameters and noise levels
[49]	2022	Dual	Utilized complex-valued operations on a cross-domain neural network called the Primal-Dual net (PD-net) for reconstruction and provided an optimal representation of magnitude and phase information in the data	Limited generalization to unseen data and imaging conditions
[50]	2022	Dual	Preserved structure details and removed aliasing artifacts using double-domain GAN	Limited validation on clinical usability, further experiments needed to introduce additional analysis measurements
[51]	2022	k-space	Achieved high-fidelity multi-coil MRI reconstruction using recurrent variational network	Required more memory during training to accumulate gradients for back-propagation during loss function computation
[52]	2022	Dual	Utilized spatial and Fourier domain convolutional layers in an interleaved hybrid domain CNN model, incorporating local residual connections to enhance the reconstruction performance	Still have some residual blur or ringing artifacts that could affect the accuracy of fine details in the reconstructed images