Related Experiment Videos
Multi-Frequency Feature Guided Progressive Divide-and-Conquer Network for Accelerated MRI Reconstruction
Hanshuo Zhu1,2, Xiaozhen Ren3,4, Xiaqiong Fan1
1School of Artificial Intelligence and Big Data, Henan University of Technology, Zhengzhou, 450001, China.
Journal of Imaging Informatics in Medicine
|August 4, 2026
Summary
Accelerated magnetic resonance imaging (MRI) uses a novel network to reduce artifacts from undersampling. The Multi-Feature Guided Progressive Divide-and-Conquer network improves image quality for faster, more accurate diagnoses.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Signal Processing
Background:
- Magnetic resonance imaging (MRI) is a crucial diagnostic tool but suffers from long scan times, leading to motion artifacts and reduced accuracy.
- K-space undersampling accelerates MRI acquisition but degrades image quality, posing a challenge for clinical applications.
- Existing reconstruction methods struggle to balance acceleration speed with diagnostic image fidelity.
Purpose of the Study:
- To develop an advanced deep learning network for accelerated MRI reconstruction that overcomes the limitations of current methods.
- To improve the quality of undersampled MRI images, particularly in critical diagnostic regions.
- To enable faster and more accurate MRI scans without compromising diagnostic information.
Main Methods:
- Proposing the Multi-Feature Guided Progressive Divide-and-Conquer (MFG-PDAC) network, incorporating three novel modules: Multi-Frequency Gated Attention (MFGA), Edge Enhanced Feature Modulation (EEFM), and Frequency-Aware Data Consistency (FREDC).
- Implementing a closed-loop mechanism involving feature selection, spatial optimization, and frequency-domain correction for synergistic optimization.
- Utilizing U-Net skip connections for dynamic fusion of multi-frequency features and gradient modulation to enhance edge response and high-frequency reconstruction.
- Employing a feedback mechanism where FREDC refines MFGA based on frequency band errors.
Main Results:
- MFG-PDAC achieved a peak signal-to-noise ratio (PSNR) of 37.28 dB and a structural similarity index measurement (SSIM) of 0.909 at 8x acceleration on the fast MRI knee dataset.
- The proposed network significantly outperformed current mainstream reconstruction methods in quantitative metrics.
- Superior reconstruction quality was observed in critical anatomical areas, including bone-soft tissue interfaces and ligament textures.
Conclusions:
- The MFG-PDAC network offers an accurate and efficient solution for accelerated MRI reconstruction, addressing key challenges in image quality degradation.
- This approach demonstrates significant potential for clinical translation, enabling faster MRI scans and improving diagnostic accuracy.
- The synergistic integration of multi-feature guidance, spatial optimization, and frequency-domain correction provides a robust framework for advanced MRI image reconstruction.