HD-MCVN: hybrid-domain multi-contrast variational network for MRI super-resolution

Area of Science:

• Medical Imaging
• Deep Learning
• Image Reconstruction

Background:

• Magnetic Resonance Imaging (MRI) resolution is crucial for diagnostics but limited by scan time, signal-to-noise ratio, and hardware.
• Deep learning-based multi-contrast super-resolution is a growing research area.
• Existing methods often ignore valuable k-space data and lack physically interpretable fusion strategies.

Purpose of the Study:

• To develop an interpretable MRI super-resolution framework that leverages both image and k-space data.
• To improve the quality and anatomical accuracy of reconstructed MRI images.
• To address limitations of current deep learning approaches in MRI super-resolution.

Main Methods:

• Proposed the Hybrid-Domain Multi-Contrast Variational Network (HD-MCVN), integrating variational optimization and deep learning.
• Incorporated frequency-domain k-space information and structural priors from high-resolution (HR) reference images.
• Jointly optimized reconstruction in image and k-space domains using a data fidelity layer (DFL) and structural texture refinement layer (STRL).
• Employed a hybrid texture loss function for supervising image content and edge detail reconstruction.

Main Results:

• HD-MCVN demonstrated superior performance on multiple MRI datasets.
• Achieved PSNR improvements of 0.3-0.6 dB and SSIM gains of 0.001-0.003 under ×4 undersampling.
• Showed reduced HFEN, indicating improved structural fidelity and preservation of fine anatomical details.

Conclusions:

• HD-MCVN offers a novel, interpretable framework for MRI super-resolution by fusing hybrid-domain information and multi-contrast priors.
• The method enhances clinical reliability and diagnostic potential through superior performance and interpretability.
• This approach holds significant promise for advancing medical image analysis and diagnostic practices.