SHFormer: Dynamic spectral filtering convolutional neural network and high-pass kernel generation transformer for...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Attention Mechanisms (AM) enhance imaging by focusing on vital information and inter-region relationships.
• Accelerated Magnetic Resonance Image (MRI) reconstruction can benefit from AM due to non-local influences in Fourier domain measurements.
• Existing AM models struggle with high-frequency details and require mode-specific retraining for multimodal MRI data.

Purpose of the Study:

• To develop a scalable MRI reconstruction method addressing limitations of current AM-based models.
• To enhance high-frequency detail propagation for superior image quality.
• To enable feature reusability across diverse, unseen multimodal MRI domains.

Main Methods:

• Proposed a neuromodulation-based discriminative multi-spectral AM for MRI reconstruction.
• Integrated a spectral filtering convolutional neural network for transferable feature extraction.
• Utilized a dynamic high-pass kernel generation transformer to focus on high-frequency details.

Main Results:

• Achieved scalable and high-quality MRI reconstruction.
• Demonstrated significant improvements in Peak Signal-to-Noise Ratio (PSNR) by ~1 dB and Structural Similarity Index Measure (SSIM) by ~0.01 under unseen scenarios.
• Showcased effective generalization to deviated MRI data domains without mode-specific retraining.

Conclusions:

• The proposed method offers a significant advancement in accelerated MRI reconstruction.
• It provides practical value for healthcare by enabling high-quality, generalizable MRI analysis.
• The approach facilitates improved diagnostic capabilities through enhanced image fidelity and data versatility.