Novel Deep Image Prior Network with Structured Sparsity Enhances Highly Accelerated MRI

A new deep learning network (SSCA-DIPNet) consistently outperformed baseline methods for MRI reconstruction at up to 20× acceleration, using limited training data from two clinical datasets and the fastMRI dataset. It preserves anatomical detail and reduces artifacts.

• SSCA-DIPNet integrates channel-adaptive structured sparsity, channel-attention deep image prior, and a globally learnable fusion mechanism built upon ISTA-Net+.
• Validated on two clinical datasets (179 training samples each) and fastMRI raw k-space data at 5x, 10x, and 20x acceleration factors.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.