An Arbitrary-Scale Super-Resolution Network for Multi-Contrast MRI With Permuted Cross-Attention

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Low-resolution (LR) magnetic resonance imaging (MRI) hinders clinical diagnosis and research.
• Multi-contrast MRI super-resolution (SR) utilizes complementary information but faces challenges like misalignment and limited dependency capture.
• Existing transformer networks struggle with long-range dependencies and fixed integer scaling.

Purpose of the Study:

• To develop a novel arbitrary-scale SR network for multi-contrast MRI.
• To overcome limitations of existing SR techniques in precision, dependency capture, and scaling flexibility.

Main Methods:

• Implemented a deformable registration module for precise spatial alignment of multi-contrast MRI.
• Employed a permuted cross-attention transformer to capture long-range dependencies effectively.
• Introduced a ref-scale ensemble implicit attention module for arbitrary-scale upsampling and high-frequency information integration.

Main Results:

• Validated the proposed method's superiority on two public MRI datasets.
• Demonstrated significant improvements in multi-contrast MRI SR reconstruction quality.
• The approach enables flexible, arbitrary-scale upsampling, outperforming existing methods.

Conclusions:

• The novel arbitrary-scale SR network significantly enhances multi-contrast MRI quality.
• The method shows substantial potential for improving clinical diagnosis and research applications.
• The developed technique offers a flexible and effective solution for MRI super-resolution.