SPICER: Self-supervised learning for MRI with automatic coil sensitivity estimation and reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Magnetic Resonance Imaging (MRI) reconstruction often requires fully sampled data, limiting acceleration.
• Coil sensitivity maps (CSMs) are crucial for MRI reconstruction but can be challenging to estimate accurately, especially with limited data.
• Deep learning models show promise for improving MRI reconstruction but often require fully sampled training data.

Purpose of the Study:

• To introduce SPICER, a novel deep model-based architecture (DMBA) for joint MRI reconstruction and coil sensitivity map (CSM) estimation.
• To enable efficient training of MRI reconstruction models using only noisy, undersampled k-space measurements.
• To achieve state-of-the-art performance in accelerated MRI acquisition settings without relying on fully sampled reference data.

Main Methods:

• SPICER employs a two-module architecture: a CNN-based CSM estimation module and a DMBA-based MRI reconstruction module.
• The reconstruction module integrates a physical measurement model with learned CNN priors.
• A self-supervised learning strategy allows training without fully sampled ground-truth data.

Main Results:

• SPICER achieves state-of-the-art performance in highly accelerated MRI acquisition settings.
• The method demonstrates the critical contribution of its DMBA, CSM estimation, and training loss components.
• SPICER outperforms pre-estimation methods for CSMs, particularly when autocalibration signal (ACS) data is limited.

Conclusions:

• SPICER successfully reconstructs high-quality MRI images and CSMs from noisy, undersampled data.
• The method surpasses other self-supervised learning approaches and matches supervised methods like E2E-VarNet.
• SPICER offers a robust solution for accelerated MRI reconstruction without the need for fully sampled data.