Advancing 1.5T MR imaging: toward achieving 3T quality through deep learning super-resolution techniques.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• 3T MRI scanners offer superior image quality and signal-to-noise ratio (SNR) for diagnosing complex neurological conditions compared to 1.5T MRI.
• However, the high cost, accessibility limitations, and increased susceptibility to image distortions of 3T scanners lead to data heterogeneity in neuroimaging studies.
• This heterogeneity poses challenges for data comparison and harmonization across different healthcare institutions.

Purpose of the Study:

• To investigate the efficacy of deep learning-based super-resolution techniques in enhancing 1.5T MRI images.
• To achieve image quality comparable to 3T MRI scans using lower-field strength scanners.
• To provide a cost-effective and accessible method for obtaining high-resolution neuroimaging data.

Main Methods:

• Three deep learning-based super-resolution techniques were evaluated.
• 1.5T MRI images were enhanced to simulate 3T quality.
• Image quality was assessed using metrics including Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS), and Intensity Differences in Pixels (IDP).

Main Results:

• The Transformer Enhanced Generative Adversarial Network (TCGAN) significantly outperformed other evaluated deep learning methods.
• TCGAN effectively reduced pixel differences, enhanced image sharpness, and preserved crucial anatomical details.
• The enhanced 1.5T images showed high similarity and visual quality comparable to 3T MRI counterparts.

Conclusions:

• TCGAN presents a viable and cost-effective alternative for generating high-quality neuroimaging data.
• This deep learning approach mitigates the need for expensive 3T MRI scans.
• The method aids in overcoming data inconsistencies and harmonization challenges in multi-center neuroimaging studies.