Bidirectional Mapping Perception-enhanced Cycle-consistent Generative Adversarial Network for Super-resolution of...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• High-resolution (HR) structural magnetic resonance imaging (MRI) is crucial for visualizing neurodegeneration and quantitative analysis.
• Super-resolution (SR) reconstruction enhances image resolution without hardware upgrades, but existing methods often require paired data and lack quantitative validation.
• Current SR techniques struggle with quantitative analysis and often depend on paired image datasets.

Purpose of the Study:

• To propose a semi-supervised generative adversarial network (GAN) model for brain MRI super-resolution using unpaired data.
• To evaluate the quantitative accuracy and clinical relevance of the generated synthetic HR images.
• To improve the diagnostic capability of low-resolution (LR) MRI scans through SR reconstruction.

Main Methods:

• Developed a semi-supervised GAN with cycle-consistency for training with unpaired data and perceptual loss for preserving high-frequency details.
• Utilized 363 subjects with both HR and LR scans and 217 subjects with HR scans only for model training and validation.
• Extracted voxel-based and surface-based morphological features for quantitative comparison between synthetic and real HR images.

Main Results:

• The model achieved excellent performance with low error rates and high similarity metrics (MAE: 0.049±0.021, MSE: 0.0059±0.0043, PSNR: 29.41±3.71, SSIM: 0.914±0.048).
• Eight morphological metrics demonstrated significant agreement (P<0.0001) between synthetic and real HR images.
• Disease diagnostic accuracy using synthetic HR images was within 5% of real HR images, significantly outperforming LR images.

Conclusions:

• The proposed semi-supervised GAN model effectively reconstructs high-resolution brain MRIs from low-resolution inputs.
• The synthetic HR images are quantitatively accurate and suitable for clinical applications, including disease diagnosis.
• This technology enables accurate image quantification and enhances the utility of MRI for neurodegenerative disease research and clinical practice.