Mouse brain MR super-resolution using a deep learning network trained with optical imaging data.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence

Background:

• Magnetic resonance imaging (MRI) resolution is limited by signal-to-noise ratio, hindering diagnostic value.
• Deep learning-based super-resolution (SR) offers potential for MRI enhancement but requires large datasets, often unavailable for preclinical research.
• Existing SR methods face challenges in adapting to the specific characteristics of MRI data.

Purpose of the Study:

• To evaluate deep learning-based SR performance for mouse brain images using high-resolution auto-fluorescence (AF) data.
• To investigate the transferability of SR networks trained on AF data to enhance preclinical MRI data.
• To explore the use of transfer learning to overcome data limitations in applying SR to MRI.

Main Methods:

• Utilized high-resolution mouse brain auto-fluorescence (AF) data from serial two-photon tomography (STPT) for initial deep learning SR model training.
• Assessed SR network performance by matching resolutions of training and target data.
• Applied transfer learning to fine-tune the AF-trained SR network using a limited dataset of high-resolution mouse brain MRI data.

Main Results:

• Optimal SR performance was achieved when training and target data resolutions were consistent.
• The effectiveness of the SR network on MRI data was influenced by the tissue contrast within the MRI images.
• Transfer learning successfully adapted the SR network, trained on AF data, to enhance the resolution of mouse brain MRI data.

Conclusions:

• Deep learning SR networks trained on high-resolution data from one modality (e.g., AF) can be effectively adapted for MRI data enhancement.
• Transfer learning is a viable strategy to overcome data scarcity issues in applying deep learning SR to preclinical MRI.
• This approach holds promise for improving the resolution and diagnostic utility of low-SNR MRI scans.