A unified hybrid transformer for joint MRI sequences super-resolution and missing data imputation

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• High-resolution multi-modal magnetic resonance imaging (MRI) is vital for diagnosis.
• Acquiring multiple MRI sequences is often limited by cost and image quality issues.

Purpose of the Study:

• To develop a unified framework for simultaneous super-resolution (SR) and imputation of missing MRI sequences.
• To address limitations in MRI acquisition for clinical and research applications.

Main Methods:

• Proposed SIFormer, a hybrid framework using a generator and discriminator.
• Generator features a dual branch attention block (Transformer + CNN) and a learnable gating adaptation MLP.
• Framework reconstructs under-sampled images and imputes missing sequences in one process.

Main Results:

• SIFormer outperformed six state-of-the-art methods in quantitative performance.
• Achieved visually superior results for image SR and synthesis across multiple datasets.
• Demonstrated effectiveness on multi-center, multi-contrast MRI data.

Conclusions:

• SIFormer offers a promising solution for enhancing MRI data acquisition.
• Potential to supplement clinical and research MRI sequence acquisition.
• Applicable to both healthy individuals and brain tumor patients.