Diffusion tensor estimation with transformer neural networks.

Area of Science:

• Neuroimaging
• Medical Physics
• Computational Neuroscience

Background:

• Diffusion Tensor Imaging (DTI) is crucial for studying brain white matter development and degeneration.
• Standard DTI requires numerous high-quality measurements, leading to long scan times.
• This poses challenges for vulnerable populations like neonates.

Purpose of the Study:

• To develop a novel DTI method for accurate diffusion tensor estimation using minimal measurements.
• To reduce scan times in DTI, particularly for neonates and infants.

Main Methods:

• Utilized transformer networks to learn relationships between diffusion signals and tensors in neighboring voxels.
• Employed a two-network model: one for initial tensor estimation, a second for refinement.
• Tested the method on three independent datasets.

Main Results:

• The proposed method accurately estimates diffusion tensors from only six diffusion-weighted measurements.
• Achieved estimations comparable to standard methods using 30-88 measurements.
• Demonstrated significant superiority over three competing methods.

Conclusions:

• This novel DTI method enables highly accurate brain white matter assessment with significantly reduced scan times.
• Promises more reliable neuroimaging for non-cooperative patients, including neonates and infants.