DDTracking: A Deep Generative Framework for Diffusion MRI Tractography with Streamline Local-Global Spatiotemporal Modeling

This paper presents DDTracking, a novel deep generative framework for diffusion MRI tractography that formulates streamline propagation as a conditional denoising diffusion process. In DDTracking, we introduce a dual-pathway encoding network that jointly models local spatial encoding (capturing fine-scale structural details at each streamline point) and global temporal dependencies (ensuring long-range consistency across the entire streamline). Furthermore, we design a conditional diffusion model module, which leverages the learned local and global embeddings to predict streamline propagation orientations for tractography in an end-to-end trainable manner. We conduct a comprehensive evaluation across diverse, independently acquired dMRI datasets, including both synthetic and clinical data. Experiments on two well-established benchmarks with ground truth (ISMRM Challenge and TractoInferno) demonstrate that DDTracking largely outperforms current state-of-the-art tractography methods. Furthermore, our results highlight DDTracking’s strong generalizability across heterogeneous datasets, spanning varying health conditions, age groups, imaging protocols, and scanner types. Collectively, DDTracking offers anatomically plausible and robust tractography, presenting a scalable, adaptable, and end-to-end learnable solution for broad dMRI applications. Code is available at: https://github.com/yishengpoxiao/DDtracking.git

💡 Research Summary

DDTracking introduces a novel deep generative framework for diffusion MRI tractography that treats streamline propagation as a conditional denoising diffusion process. The method begins by extracting spherical harmonic (SH) coefficients from a 3 × 3 × 3 voxel neighborhood around each point on a streamline, producing a 4‑D feature tensor. These features are processed through a dual‑pathway spatial encoder consisting of two separate 3D convolutional branches. One branch yields a spatial embedding (zₜ) that feeds into a recurrent neural network (GRU) to capture long‑range temporal dependencies along the streamline, producing a global context vector (cₜ). The other branch generates a local embedding (vₜ) that serves as a fine‑grained conditioning signal.
Both the global context (cₜ combined with a sinusoidal positional embedding of the diffusion step) and the local context (vₜ) are fused via Feature‑wise Linear Modulation (FiLM) inside a 1‑D convolutional diffusion model. In the forward diffusion stage, the clean orientation vector is linearly attenuated to zero while Gaussian noise is gradually added; the reverse diffusion stage uses the conditioned 1‑D CNN to iteratively denoise and recover the next propagation orientation. This design enables the model to learn complex multimodal trajectory distributions, handling crossing, fanning, and branching fibers without relying on discrete orientation sampling or explicit fiber orientation distributions.
The authors evaluate DDTracking on two benchmark datasets with ground truth (ISMRM 2015 Tractography Challenge and TractoInferno) as well as a variety of synthetic and clinical scans covering different scanners, b‑values, age groups, and pathologies. Across all metrics—precision, recall, bundle overlap, and connectivity—DDTracking outperforms state‑of‑the‑art deterministic, probabilistic, and recent deep learning approaches (e.g., RNN‑based, reinforcement‑learning, and uncertainty‑aware methods). Notably, the model shows robust generalization to unseen acquisition protocols and maintains high anatomical plausibility in regions with tight bottlenecks or high curvature. Computationally, the 1‑D diffusion backbone runs efficiently on modern GPUs, achieving near‑real‑time tractography speeds.
In summary, DDTracking unifies local spatial detail and global temporal coherence within a conditional diffusion framework, delivering anatomically accurate, robust, and scalable tractography. The open‑source code and pretrained models facilitate adoption in research and clinical pipelines, representing a significant step forward for data‑driven white‑matter mapping.