MCDM-DTI: Accelerating Diffusion Tensor Imaging based on Multi-Conditional Denoising Diffusion Probability Model

Diffusion tensor imaging (DTI) is a significant non-invasive neuroimaging technique with valuable applications in clinical medicine. To address the lengthy acquisition times and high costs of traditional DTI technology, we propose a multi-conditional diffusion model for accelerated DTI (MCDM-DTI). This model generates high-quality DTI maps using only six diffusion-weighted images (DWIs) and one b0 image. The model leverages the diffusion process of Denoising Diffusion Probabilistic Models (DDPMs) to learn the underlying data distribution of DTI maps. By incorporating the DWIs and the b0 image as conditional inputs to a U-Net denoising network, MCDM-DTI produces more accurate DTI parameter maps. The introduction of Class Label Embedding technology extracts category-specific features, enabling the concurrent generation of all category parameter maps. Additionally, a binary image erosion algorithm can eliminate edge noise. Validated on the Human Connectome Project dataset (HCP), MCDM-DTI demonstrates superior performance in generating FA, MD, and AD, along with good robustness against noise, compared to previously studied methods. Our approach achieves over a 5-fold boost in imaging speed while maintaining high-resolution DTI maps, significantly reducing acquisition time.