FlexDTI: flexible diffusion gradient encoding scheme-based highly efficient diffusion tensor imaging using deep...

Area of Science:

• Medical Imaging
• Neuroimaging
• Computational Neuroscience

Background:

• Deep neural network methods for diffusion tensor imaging (DTI) typically require matching diffusion gradient numbers and directions between data and training sets.
• This limitation hinders the application of DTI in scenarios with varying or incomplete diffusion encoding schemes.

Purpose of the Study:

• To develop and evaluate FlexDTI, a novel dynamic-convolution-based method for efficient DTI reconstruction.
• To achieve high-quality DTI parametric mapping using a flexible diffusion encoding gradient scheme.

Main Methods:

• FlexDTI utilizes dynamic convolution kernels to embed diffusion gradient direction information directly into signal feature maps.
• Network generalization for a flexible number of diffusion gradient directions is achieved by setting the maximum input channels.
• The method was trained and validated using datasets from the Human Connectome Project and local hospitals.

Main Results:

• FlexDTI successfully reconstructs high-quality diffusion tensor-derived parameters irrespective of changes in the number and directions of diffusion encoding gradients.
• It demonstrates a significant reduction in normalized root mean squared error: approximately 50% for fractional anisotropy and 15% for mean diffusivity compared to a state-of-the-art deep learning method.
• The method shows superior performance over other advanced tensor parameter estimation techniques.

Conclusions:

• FlexDTI effectively learns diffusion gradient direction information, enabling generalized DTI reconstruction with flexible gradient schemes.
• The developed network balances flexibility in diffusion encoding with high-quality reconstruction of DTI parameters.
• This approach offers a significant advancement for DTI analysis in diverse clinical and research settings.