DM-Net: a physics-model-independent direct mapping approach for calibration-free multi-coil MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning (DL) significantly advances multi-coil MRI reconstruction.
• Current DL models often rely on physics-based methods requiring coil sensitivity, limiting performance or complexity.
• Optimizing reconstruction without explicit coil sensitivity is a key challenge.

Purpose of the Study:

• To introduce DM-Net, a novel physics-model-independent DL approach for MRI reconstruction.
• To demonstrate optimal reconstruction performance without explicit coil sensitivity input.
• To evaluate DM-Net against established reconstruction methods like SENSE and GRAPPA.

Main Methods:

• A densely connected convolutional network (DM-Net) was designed for direct mapping, intrinsically exploiting coil sensitivity and channel correlation.
• Comparative models included DMwS-Net (direct mapping with coil sensitivity) and CCR-Nets (coil-by-coil reconstruction).
• Models were trained and tested on 5440 3D Fast Fourier Transform (3DFT) MRI images from 17 subjects.

Main Results:

• DM-Net demonstrated superior reconstruction performance compared to comparative methods.
• The feasibility of excluding precalculated coil sensitivity maps from DL model input was confirmed.
• DM-Net's applicability to k-space data lacking fully sampled calibration regions was shown, potentially improving image quality.

Conclusions:

• DM-Net offers a reliable, simplified, and faster approach to DL-based MRI reconstruction.
• Excluding explicit coil sensitivity from DL models is a viable strategy for improved performance.
• This work broadens the scope of DL applications in medical image reconstruction.