Synthetic Diffusion Tensor Imaging Maps Generated by 2D and 3D Probabilistic Diffusion Models: Evaluation and...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• Diffusion tensor imaging (DTI) is crucial for brain microstructure analysis but faces challenges like high cost, long acquisition times, and artifacts.
• Data scarcity and privacy concerns limit the training of deep learning models for DTI analysis.
• Synthetic DTI generation is gaining traction to overcome these limitations and enhance data availability.

Purpose of the Study:

• To evaluate the quality, fidelity, and downstream application value of synthetic DTI mean diffusivity (MD) maps generated by 2D and 3D denoising diffusion probabilistic models (DDPMs).
• To assess the computational efficiency and data augmentation utility of these synthetic DTI methods in classification tasks.
• To provide a benchmark analysis of synthetic diffusion-weighted imaging approaches.

Main Methods:

• Generation of synthetic DTI MD maps using 2D slice-wise and 3D volume-wise DDPMs.
• Evaluation of image quality, fidelity, and diversity of generated synthetic DTI maps.
• Assessment of downstream task performance (sex and dementia classification) using 2D and 3D convolutional neural networks (CNNs) with augmented data.
• Benchmarking computational efficiency and performance trade-offs.

Main Results:

• 3D volume-wise DDPM synthesis demonstrated superior performance in downstream classification tasks compared to 2D slice-wise synthesis.
• Synthetic DTI data effectively augmented training datasets, improving the performance of CNNs for sex and dementia classification.
• DDPMs showed advantages in fidelity, diversity, controllability, and stability over traditional generative models like GANs and VAEs.

Conclusions:

• 3D DDPMs are a highly effective method for generating high-quality synthetic DTI data, outperforming 2D approaches for downstream applications.
• Synthetic DTI generation using DDPMs offers a viable solution for data augmentation, addressing data scarcity and privacy issues in neuroimaging research.
• This study provides valuable insights into the trade-offs of different synthetic diffusion-weighted imaging techniques, guiding future research and clinical applications.