Beyond Nyquist: A Comparative Analysis of 3D Deep Learning Models Enhancing MRI Resolution.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• High-spatial resolution MRI provides detailed structural information crucial for diagnosis and treatment.
• Acquiring high-resolution MRI data is limited by reduced spatial coverage, lower signal-to-noise ratio (SNR), and extended scan times.
• Deep learning-based super-resolution techniques offer a promising solution to overcome these limitations in MRI.

Purpose of the Study:

• To compare the performance and robustness of various state-of-the-art 3D convolutional neural network models for MRI super-resolution.
• To identify the optimal deep learning model for enhancing the resolution of structural MRI data.

Main Methods:

• Evaluation of five 3D convolutional neural network models: RRDB, SPSR, UNet, UNet-MSS, and ShuffleUNet.
• Utilized the public IXI dataset, artificially downsampling structural MRI images with factors ranging from 8 to 64.
• Assessed model performance using the Structural Similarity Index Measure (SSIM) metric on a test set.

Main Results:

• All evaluated models demonstrated good performance in the super-resolution task.
• The UNet model consistently achieved the highest performance across all downsampling factors.
• The SPSR model exhibited the poorest performance, while UNet and UNet-MSS showed top overall results. RRDB performed poorly relative to other models.

Conclusions:

• UNet and UNet-MSS are highly effective deep learning models for MRI super-resolution, offering robust performance.
• Deep learning techniques, particularly UNet, can significantly improve the quality of low-resolution MRI data, addressing acquisition limitations.
• The study provides valuable insights for selecting appropriate deep learning architectures for enhancing MRI resolution in clinical and research settings.