A dual-task dual-domain model for blind MRI reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• Magnetic Resonance Imaging (MRI) acquisition acceleration is crucial for reducing scan times.
• Current cascade models for accelerated MRI reconstruction often require a known sampling mask (mask prior).

Purpose of the Study:

• To develop a method for MRI reconstruction that does not rely on a known mask prior.
• To improve the performance of accelerated MRI reconstruction without prior information.

Main Methods:

• An auxiliary network was designed to estimate the sampling mask from k-space data.
• A dual-domain cascade network utilizing partial convolutional layers was proposed for k-space inpainting.
• The model was evaluated as a blind reconstruction task without mask prior.

Main Results:

• The proposed auxiliary network successfully estimated the sample mask from k-space data.
• The dual-domain cascade network demonstrated effective k-space inpainting for MRI reconstruction.
• The blind reconstruction model achieved superior performance at 4x and 8x acceleration compared to existing methods.

Conclusions:

• MRI reconstruction can be effectively performed without a mask prior by estimating it using an auxiliary network.
• Rethinking MRI reconstruction as a k-space inpainting task with a dual-domain cascade network offers significant advantages.
• The developed blind reconstruction model represents a breakthrough in accelerated MRI acquisition.