FDoSR-Net: Frequency-Domain Informed Auto-Encoder Network for Arbitrary-Scale 3D Whole-Heart MRI Super-Resolution

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Three-dimensional (3D) whole-heart (WH) magnetic resonance imaging (MRI) is crucial for cardiac assessment.
• Current super-resolution (SR) methods often struggle to maintain fine details in 3D WH MRI.
• Enhancing the resolution of 3D WH MRI is essential for improved diagnostic accuracy.

Purpose of the Study:

• To develop a 3D super-resolution (SR) network for arbitrary-scale 3D whole-heart (WH) MRI.
• To maintain fine image details during the super-resolution process.
• To outperform existing SR algorithms in terms of image quality and detail preservation.

Main Methods:

• Utilized a 3D autoencoder framework with frequency-domain regularization for training.
• Trained the network for arbitrary factor SR capabilities.
• Evaluated the method using 120 3D WH MR volumes acquired with four different sequences.

Main Results:

• The proposed method demonstrated superior performance across quantitative metrics: PSNR (mean improvement 4.5%), SSIM (mean improvement 2.2%), MSE (mean improvement 48.2%), and RMSE (mean improvement 31.0%).
• Qualitative visual comparisons confirmed the method's effectiveness in preserving fine image details.
• Outperformed state-of-the-art deep learning methods (ACNS, TFC) and nearest neighbor interpolation.

Conclusions:

• The developed 3D SR network effectively performs arbitrary-scale super-resolution for 3D WH MRI.
• The method excels at preserving fine image details, surpassing current state-of-the-art techniques.
• This advancement holds significant potential for improving cardiac MRI diagnostics.