Effective deep-learning brain MRI super resolution using simulated training data.

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• High-resolution (HR) magnetic resonance imaging (MRI) is crucial for diagnostics but often unavailable or artifact-prone.
• Low-resolution (LR) MRI is common, and deep learning (DL) super-resolution (SR) can enhance it.
• Current DL-SR methods require paired HR-LR training data.

Purpose of the Study:

• To investigate the efficacy of simulated brain MRI data for training DL-based SR networks.
• To assess if simulated data can augment existing training datasets.

Main Methods:

• Simulated a large, diverse, artifact-free brain MRI dataset with voxel alignment and varying resolutions.
• Trained four DL-SR networks using simulated data and augmented real data.
• Evaluated trained networks on real-world MRI data from multiple sources.

Main Results:

• Generated 0.7mm SR images from 1mm LR T1w brain MRI.
• Trained networks significantly improved LR image sharpness.
• Networks augmented with simulated data showed superior performance across multiple real datasets compared to those trained solely on real data.

Conclusions:

• Simulated paired HR-LR brain MRI data is effective for training and augmenting DL-SR networks.
• Using simulated data for augmentation enhances SR network generalizability across diverse real MRI datasets.