RAVEN: Robust, generalizable, multi-resolution structural MRI upsampling using autoencoders

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• Magnetic Resonance Images (MRIs) offer high inter-tissue contrast, reflecting neuroanatomical changes in aging and disease.
• Standard MRI resolutions limit the detection of subtle, early-stage pathological changes.
• Increasing MRI acquisition resolution faces challenges like noise, longer scan times, and patient discomfort.

Purpose of the Study:

• To introduce a robust and generalizable single-image super-resolution network for brain MRIs.
• To address the limitations of standard MRI resolution for detecting early neurodegenerative changes.
• To provide an open-access tool for advanced brain MRI analysis.

Main Methods:

• Developed Resolution Augmentation with Variational auto-Encoder Networks (RAVEN), a deep learning model utilizing generative adversarial networks (GANs).
• Trained and evaluated RAVEN on diverse in-vivo and ex-vivo brain MRIs across various modalities (T1w, T2w, T2*) and field strengths (3T-7T).
• Assessed RAVEN's capability to achieve target voxel sizes as small as 0.5 mm isotropic with arbitrary upsampling factors.

Main Results:

• RAVEN successfully upsampled brain MRIs to high resolutions (0.5 mm isotropic), preserving anatomical details.
• The network demonstrated state-of-the-art performance compared to existing deep learning and non-deep learning super-resolution methods.
• RAVEN effectively preserved true anatomical information in upsampled images.

Conclusions:

• RAVEN offers a powerful solution for enhancing brain MRI resolution, enabling earlier detection of neuroanatomical changes.
• The method is generalizable across different MRI modalities and field strengths.
• RAVEN's open-access availability facilitates its adoption in research and clinical settings.