Calibrationless reconstruction of uniformly-undersampled multi-channel MR data with deep learning estimated ESPIRiT...

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Parallel imaging reconstruction in MRI relies on accurate coil sensitivity maps.
• Traditional methods for estimating these maps, such as Eigenvalue Approach to Autocalibrating Parallel MRI (ESPIRiT) maps, often require dedicated calibration scans or sufficient autocalibration data.
• This dependence on calibration data can limit acceleration factors and increase scan times.

Purpose of the Study:

• To develop a truly calibrationless reconstruction method for parallel MRI.
• To derive Eigenvalue Approach to Autocalibrating Parallel MRI (ESPIRiT) maps directly from uniformly-undersampled multi-channel MR data using deep learning.
• To enable faster and more efficient MRI acquisition without compromising image quality.

Main Methods:

• A U-Net based deep learning model was designed to estimate multi-channel ESPIRiT maps directly from uniformly-undersampled multi-channel multi-slice MR data.
• The model was trained on fully-sampled multi-slice axial brain datasets.
• A hybrid loss function was implemented to leverage subject-coil geometric parameters, imposing constraints on ESPIRiT maps at both original and standard reference locations.

Main Results:

• The deep learning model robustly predicted multi-channel ESPIRiT maps from undersampled k-space data.
• The predicted maps were highly comparable to reference maps derived from central k-space lines.
• The derived maps facilitated excellent parallel MRI reconstruction performance, even at high acceleration factors, with minimal increase in artifacts compared to using reference maps.

Conclusions:

• A novel deep learning approach enables direct estimation of ESPIRiT maps from undersampled MRI data, eliminating the need for calibration scans.
• This method offers a general strategy for calibrationless parallel imaging reconstruction.
• The approach learns coil and protocol-specific information, paving the way for accelerated and more efficient MRI.