Enhancing quality and speed in database-free neural network reconstructions of undersampled MRI with SCAMPI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning models for MRI reconstruction typically require extensive training on large datasets.
• This training process can lead to overfitting, where models perform poorly on data not seen during training.
• Untrained neural networks offer an alternative by learning directly from the reconstruction task itself.

Purpose of the Study:

• Introduce SCAMPI (Sparsity Constrained Application of deep Magnetic resonance Priors for Image reconstruction), an untrained deep neural network for MRI reconstruction.
• Expand the Deep Image Prior approach with a multidomain, sparsity-enforcing loss function.
• Achieve higher image quality and faster convergence compared to existing methods.

Main Methods:

• Utilized two-dimensional MRI data from the FastMRI dataset with Cartesian undersampling.
• Reconstructed data at various acceleration rates for both single-coil and multi-coil acquisitions.
• Compared SCAMPI against state-of-the-art Compressed Sensing methods and the ConvDecoder network.

Main Results:

• SCAMPI demonstrated superior performance in reducing undersampling artifacts and error metrics, particularly in multi-coil imaging.
• The U-Net architecture combined with an advanced loss function resulted in faster convergence and higher image quality than ConvDecoder.
• SCAMPI successfully reconstructed multi-coil data without explicit coil sensitivity profile knowledge and is effective for single-coil k-space data.

Conclusions:

• SCAMPI's untrained nature prevents overfitting to dataset-specific features, as network parameters are tuned solely on reconstruction data.
• This method achieves superior results and faster reconstruction times compared to baseline untrained neural network approaches.
• SCAMPI offers a novel and effective tool for reconstructing undersampled MRI data.