MEDL-Net: A model-based neural network for MRI reconstruction with enhanced deep learned regularizers.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Model-based networks are crucial for MRI reconstruction.
• These networks often demand significant GPU memory and computational resources.
• Improving efficiency and performance is essential for clinical applications.

Purpose of the Study:

• To enhance the performance of model-based MRI reconstruction networks.
• To reduce the substantial GPU memory requirements of these networks.
• To develop a more efficient deep learning approach for MR image reconstruction.

Main Methods:

• Proposed a model-based neural network with enhanced deep learned regularizers (MEDL-Net).
• MEDL-Net utilizes cascaded submodules mimicking optimization steps, with dense connections and revising blocks (RB).
• A composition loss function was designed for explicit supervision of RBs.

Main Results:

• MEDL-Net quantitatively outperformed state-of-the-art methods across various acceleration rates (4x, 6x).
• Reconstructed MR images demonstrated superior preservation of detailed textures.
• Achieved comparable reconstruction results with fewer cascades compared to existing model-based networks.

Conclusions:

• A more efficient model-based deep network (MEDL-Net) was developed for MR image reconstruction.
• Experimental results confirm improved reconstruction performance and reduced cascade requirements.
• The proposed method effectively alleviates the large demand for GPU memory in MRI reconstruction.