Deep learning-based diffusion tensor image generation model: a proof-of-concept study

Area of Science:

• Medical Imaging
• Neuroimaging
• Artificial Intelligence in Medicine

Background:

• Diffusion Tensor Imaging (DTI) provides crucial microstructural information about white matter tracts.
• Conventional Diffusion Weighted Imaging (DWI) is more widely available but offers less detailed directional information.
• Synthesizing DTI from DWI could enhance the utility of existing imaging datasets.

Purpose of the Study:

• To develop and validate an image-to-image translation model for synthesizing DTI from DWI.
• To assess the quantitative similarity between original and synthetic DTI data.
• To explore the potential of AI in generating advanced neuroimaging metrics from standard sequences.

Main Methods:

• An image-to-image translation model was trained to synthesize DTI from DWI using six motion probing gradient (MPG) directions.
• Data from 32 healthy volunteers were used, with DTI and DWI acquired prospectively.
• Regions of Interest (ROIs) in key brain structures were analyzed for mean values and signal-to-noise ratio (SNR), with Bland-Altman analysis performed.

Main Results:

• The model successfully synthesized DTI from conventional DWI.
• While synthetic data showed slightly lower SNR and higher variability, Bland-Altman plots indicated similar distributions between original and synthetic DTI.
• Quantitative analysis in ROIs demonstrated comparable results between the two DTI types.

Conclusions:

• Image-to-image translation models can effectively generate synthetic DTI from conventional DWI.
• This approach holds potential for expanding DTI analysis capabilities using existing DWI data.
• Further validation is warranted, but the method shows promise for neuroimaging research.