An untrained deep learning method for reconstructing dynamic MR images from accelerated model-based data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Accelerated MRI acquisition reduces scan times but often compromises image quality.
• Deep neural networks (DNNs) show potential for reconstructing MR images from undersampled data.
• Determining optimal training duration for DNNs in MRI reconstruction is challenging without ground truth.

Purpose of the Study:

• To implement physics-based regularization as an automated stopping condition for tuning an untrained DNN.
• To reconstruct MR images from accelerated data using this novel approach.

Main Methods:

• A ConvDecoder (CD) neural network was trained with a physics-based regularization term derived from the spoiled gradient echo equation.
• Retrospective acceleration of k-space data (R = 8, 12, 18, 36) was performed.
• Reconstructions were compared using CD, CD with regularization (CD+r), locally low-rank (LR), and L1-wavelet regularization (L1), evaluated by normalized RMS error, concordance correlation coefficient, and structural similarity index.

Main Results:

• CD+r reconstructions, guided by the regularization stopping condition, showed comparable structural similarity to CD and LR methods.
• CD+r significantly outperformed L1 regularization in structural similarity.
• For acceleration factors R≥12, L1 and LR methods lost spatially refined details compared to CD+r, while CD+r preserved them.

Conclusions:

• An untrained DNN combined with physics-based regularization is a viable method for determining optimal training cessation.
• This approach eliminates the need for fully-sampled ground truth data in accelerated MRI reconstruction.
• The proposed method shows promise for improving the efficiency and quality of accelerated MRI.