Taming diffusion transformers for high-fidelity MRI super-resolution.

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• Diffusion models enhance MRI super-resolution but U-Net architectures struggle with global context and scalability.
• Transformer models offer better global reasoning but are underexplored for MRI super-resolution.
• Existing methods lack efficient latent space decoding for high-fidelity MRI reconstruction.

Purpose of the Study:

• To propose DiTMSR, a novel framework for MRI super-resolution leveraging diffusion transformers.
• To address limitations in global context modeling and latent space decoding in current MRI super-resolution techniques.

Main Methods:

• Developed a conditional Diffusion Transformer for progressive denoising of latent MRI inputs.
• Designed a hybrid Mamba decoder with a content preservation module and MambaVision Mixer for efficient decoding.
• Integrated Transformer-based diffusion models for improved global reasoning and multi-scale dependency modeling.

Main Results:

• DiTMSR demonstrated superior performance in MRI super-resolution compared to state-of-the-art methods.
• The framework successfully recovered structurally faithful MR latent features.
• Experiments on public and clinical datasets validated the effectiveness of DiTMSR.

Conclusions:

• DiTMSR offers a promising advancement in MRI super-resolution by effectively combining diffusion transformers and Mamba decoders.
• The proposed framework addresses key challenges in global context modeling and latent space decoding for high-fidelity MRI reconstruction.
• DiTMSR shows potential for improving diagnostic accuracy through enhanced MRI image quality.