DIRECTION: Deep cascaded reconstruction residual-based feature modulation network for fast MRI reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep cascaded networks are used to accelerate Magnetic Resonance Imaging (MRI).
• Increasing network depth in cascaded architectures can lead to performance plateaus or degradation without proper guidance.
• Existing methods often lack effective mechanisms to guide individual subnetworks within a cascaded structure.

Purpose of the Study:

• To propose a novel parallel network architecture, DIRECTION, for enhanced MRI reconstruction.
• To introduce a Reconstruction Residual-Based Feature Modulation Mechanism (RRFMM) to guide subnetworks using reconstruction residuals.
• To improve the performance and overcome limitations of traditional deep cascaded networks in MRI acceleration.

Main Methods:

• Developed a parallel architecture named DIRECTION.
• Introduced the Reconstruction Residual-Based Feature Modulation Mechanism (RRFMM) using a Residual Attention Modulation Block (RAMB).
• Implemented Cross-Stage Feature Reuse Connection (CSFRC) and Reconstruction Dense Connection (RDC) to reduce information loss and enhance feature representation.

Main Results:

• The proposed RRFMM effectively guides each subnetwork with unique optimization objectives.
• CSFRC and RDC contribute to reduced information loss and improved feature representation.
• Experiments on the fastMRI knee dataset demonstrated significant performance boosts over state-of-the-art methods.

Conclusions:

• The DIRECTION architecture, with its RRFMM, CSFRC, and RDC, effectively enhances MRI reconstruction performance.
• This parallel approach offers a superior alternative to simply increasing the cascading number in deep networks.
• The method achieves state-of-the-art quantitative and qualitative results in accelerated MRI.