DeepSTI: Towards tensor reconstruction using fewer orientations in susceptibility tensor imaging.

Area of Science:

• Neuroimaging
• Biophysics
• Medical Physics

Background:

• Susceptibility tensor imaging (STI) offers high-resolution insights into brain structure and myelin.
• Current STI methods require extensive data acquisition across multiple orientations, limiting in vivo application.
• Physical constraints in MRI scanners further complicate multi-orientation data collection.

Purpose of the Study:

• To develop an efficient image reconstruction algorithm for susceptibility tensor imaging (STI).
• To overcome the limitations of cumbersome data acquisition in current STI techniques.
• To enable robust in vivo STI with reduced measurement requirements.

Main Methods:

• Proposed DeepSTI, a novel image reconstruction algorithm leveraging data-driven priors.
• Utilized a deep neural network to approximate the proximal operator for STI regularization.
• Employed iterative dipole inversion using the learned proximal network.

Main Results:

• Demonstrated significant improvements in reconstructed tensor images and tractography compared to state-of-the-art methods.
• Achieved accurate tensor reconstruction using significantly fewer than six orientations.
• Showcased promising results from a single orientation in human in vivo data.
• Successfully applied the technique for estimating lesion susceptibility anisotropy in multiple sclerosis patients.

Conclusions:

• DeepSTI significantly enhances the feasibility and efficiency of in vivo susceptibility tensor imaging.
• The method allows for high-quality STI reconstruction with substantially reduced acquisition time and complexity.
• DeepSTI holds promise for improved diagnosis and understanding of neurological disorders like multiple sclerosis.