[Multimodality-based super-resolution reconstruction for routine brain magnetic resonance images].

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Routine brain MRI often suffers from low resolution and high slice thickness, limiting detailed anatomical visualization.
• Accurate brain structure analysis requires high-resolution imaging, which can be time-consuming or require specialized equipment.

Purpose of the Study:

• To develop a multi-modality-based super-resolution synthesis model for reconstructing high-resolution brain MRI from low-resolution inputs.
• To improve the quality and diagnostic utility of standard brain MRI scans.

Main Methods:

• A structure-constrained image mapping network was employed, utilizing paired low- and high-resolution 2D T1, 2D T2 FLAIR, and 3D T1 MRI data.
• The model extracted features from multiple modalities, including T1 and T2 FLAIR subcortical regions, to reconstruct higher-resolution T1 images.
• Anatomical information from segmentation maps was used as a constraint for adaptive learning of brain tissue structures.

Main Results:

• The proposed method achieved superior performance compared to existing methods, with an average PSNR of 33.11 and SSIM of 0.996.
• Clear reconstruction of brain anatomical structures, including sulci, gyri, and subcortical regions, was achieved.
• The precision of brain volume measurement was significantly enhanced by the super-resolution reconstruction.

Conclusions:

• The developed multi-modality super-resolution (MSCSR) model demonstrates excellent performance in reconstructing high-resolution brain MRI.
• The model effectively leverages multi-modal information and anatomical constraints for superior image reconstruction.
• This approach holds promise for improving diagnostic accuracy and quantitative analysis in neuroimaging.