Bilevel Optimized Implicit Neural Representation for Scan-Specific Accelerated MRI Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Deep learning (DL) for accelerated magnetic resonance imaging (MRI) requires large datasets and struggles with generalization.
• Self-supervised DL methods offer scan-specific reconstruction but need complex hyperparameter tuning and have acceleration limitations.
• Existing methods face challenges in data dependency, generalization, and efficiency for accelerated MRI.

Purpose of the Study:

• To develop a data-free, scan-specific MRI reconstruction method using bilevel-optimized implicit neural representations (INRs).
• To automatically optimize reconstruction hyperparameters for diverse MRI acquisitions without requiring training data.
• To achieve high-quality, accelerated MRI reconstructions efficiently.

Main Methods:

• Formulated MRI reconstruction as a bilevel optimization problem.
• Employed Gaussian process regression for optimizing INR hyperparameters.
• Utilized a trainable positional encoder and a multilayer perceptron within the INR framework.

Main Results:

• Achieved tailored, data-free MRI reconstruction by automatically optimizing hyperparameters.
• Demonstrated computational efficiency, with offline optimization in minutes and reconstruction in seconds.
• Obtained comparable or improved image quality versus existing model-based and self-supervised methods.

Conclusions:

• The bilevel-optimized INR approach enables efficient, scan-specific MRI reconstruction without training data.
• This method offers a robust solution for accelerating MRI acquisition while maintaining high image quality.
• The approach shows promise for improving the adaptability and performance of deep learning in medical imaging.