DDTracking: Generative Diffusion MRI Tractography

• DDTracking is a learning-based diffusion MRI tractography framework that formulates streamline progression as a conditional denoising diffusion process.
• It combines local spherical-harmonic diffusion evidence with global history through a dual-pathway encoding network to predict anatomically valid streamlines.
• Empirical evaluations demonstrate improved valid connection and overlap metrics, highlighting its enhanced anatomical continuity across diverse datasets.

to confirm paper existence. Final can cite only (Li et al., 6 Aug 2025). Need use arxiv search tool though. Let's do search for (Li et al., 6 Aug 2025) and maybe title. DDTracking is a learning-based diffusion MRI tractography framework that formulates streamline propagation as a conditional denoising diffusion process rather than as a hand-designed local tracking rule, a discrete direction classifier, or an explicit fiber-orientation-distribution sampler. It predicts the next streamline orientation by conditioning a generative model on both local diffusion evidence around the current point and the global history of the streamline, with the stated aim of improving anatomical plausibility, long-range consistency, and robustness across heterogeneous datasets (Li et al., 6 Aug 2025).

1. Problem formulation and conceptual basis

In diffusion MRI tractography, a streamline is propagated step by step from seed points. If the current point is $p_t$, the next point is computed as

$p_{t+1} = p_t + \alpha y_t,$

where $\alpha$ is the step size and $y_t \in \mathbb{R}^3$ is the propagation orientation vector. DDTracking treats prediction of $y_t$ as the central learning problem. The framework is motivated by the observation that local diffusion signals are noisy and often ambiguous, especially in crossing, fanning, and branching regions, so small local orientation errors can accumulate into anatomically implausible streamlines over long trajectories (Li et al., 6 Aug 2025).

The method is positioned against both model-based and learning-based tractography. Traditional deterministic tractography is described as efficient but fragile in complex fiber configurations, while probabilistic tractography is described as more flexible but still dependent on explicit orientation distributions or handcrafted assumptions. Existing machine-learning methods are characterized as improving over these baselines while still often relying on approximations such as fiber-orientation-distribution peaks or discrete direction sampling, or emphasizing either local context or sequential history without jointly modeling both. DDTracking is therefore defined as a generative alternative that attempts to model the conditional distribution of valid next-step orientations directly.

A central claim of the framework is that tractography is a trajectory-generation problem. On that view, the next orientation should be generated from a conditional distribution that reflects both the immediate local diffusion structure and the accumulated streamline context. This suggests a plausible implication: DDTracking is less a local decision rule than a learned sequential prior over anatomically valid continuation.

2. Conditional denoising diffusion formulation

DDTracking represents a streamline as

$S=\{p_1,p_2,\dots,p_n\},$

with equally spaced points. For each point $p_t$, it extracts a local spherical-harmonic feature tensor

$D(p_t)\in \mathbb{R}^{3\times3\times3\times m},$

where the $3\times 3\times 3$ neighborhood contains the center voxel and its 26 neighbors. In the reported implementation, $l_{\max}=6$, giving $p_{t+1} = p_t + \alpha y_t,$0 spherical-harmonic coefficients per voxel.

The generative component does not use a standard discrete DDPM schedule. Instead, it adopts a continuous-time decoupled process in which the clean orientation $p_{t+1} = p_t + \alpha y_t,$1 is attenuated toward zero while Gaussian noise is injected. With $p_{t+1} = p_t + \alpha y_t,$2, $p_{t+1} = p_t + \alpha y_t,$3, and $p_{t+1} = p_t + \alpha y_t,$4, the forward process is

$p_{t+1} = p_t + \alpha y_t,$5

The reverse process is then learned as a conditional denoising model that reconstructs a valid orientation from $p_{t+1} = p_t + \alpha y_t,$6 given streamline context. The conditional distribution is written as

$p_{t+1} = p_t + \alpha y_t,$7

where $p_{t+1} = p_t + \alpha y_t,$8 denotes global context and $p_{t+1} = p_t + \alpha y_t,$9 denotes local context (Li et al., 6 Aug 2025).

The paper states that training minimizes expected reconstruction error for both the attenuation term and the noise term, with dynamic weights

$\alpha$0

and applies Smooth L1 loss for robustness. Conceptually, this means the model is trained not merely to regress a direction vector, but to denoise a corrupted orientation under local-global conditioning.

3. Dual-pathway local-global spatiotemporal architecture

The architectural core of DDTracking is a dual-pathway encoding network. A spatial encoder processes each local tensor $\alpha$1 and outputs two embeddings,

$\alpha$2

The paper assigns distinct roles to these outputs: $\alpha$3 is passed to the temporal encoder, while $\alpha$4 is used as local conditioning for the diffusion predictor. The spatial encoder is described as using dual 3D convolution branches followed by multilayer perceptrons, though exact kernel sizes and depths are not specified.

The temporal pathway takes the sequence

$\alpha$5

and processes it with a two-stacked GRU, yielding a context embedding

$\alpha$6

This recurrent branch is intended to encode long-range streamline dependencies and maintain continuity over variable-length trajectories. The local and global representations are then coupled inside the diffusion module: local context is provided by $\alpha$7, while global context combines $\alpha$8 with a sinusoidal positional embedding of diffusion step $\alpha$9 (Li et al., 6 Aug 2025).

The conditional diffusion predictor itself is described as a 1D convolutional symmetric encoder-decoder with FiLM-modulated residual blocks. Its role is to predict the quantities needed for reverse denoising and thus produce the next orientation estimate $y_t \in \mathbb{R}^3$0. In architectural terms, DDTracking therefore separates fine-scale local signal encoding, streamline-history modeling, and conditional generative prediction rather than collapsing them into a single regression module.

4. Data representation, training workflow, and tractography procedure

Input diffusion-weighted images are normalized by the $y_t \in \mathbb{R}^3$1 image and projected onto a spherical-harmonic basis using DIPY with the Descoteaux07 basis. For several external datasets, preprocessing through pnlpipe includes brain extraction, eddy current correction, EPI distortion correction, rigid registration to MNI space, and resampling to $y_t \in \mathbb{R}^3$2. HCP-YA, ISMRM, and TractoInferno are used in their provided or benchmark-standard preprocessed forms.

Training is reported on HCP-YA and TractoInferno. HCP-YA contains 105 subjects, of which 10 are used for training in the reported setup. TractoInferno is used with 198 training, 58 validation, and 28 test subjects. The implementation uses PyTorch v2.2.2, AdamW, an initial learning rate of $y_t \in \mathbb{R}^3$3, learning-rate reduction by a factor of 10 if validation loss does not improve for 50 epochs, a minimum learning rate of $y_t \in \mathbb{R}^3$4, and early stopping with patience 120 epochs. Hardware is a Linux workstation with 16 $y_t \in \mathbb{R}^3$5 32 GB RAM and 6 $y_t \in \mathbb{R}^3$6 NVIDIA GeForce RTX 4090 (24 GB) (Li et al., 6 Aug 2025).

At tractography time, the framework places five seeds per voxel in the white matter mask. For each seed, it repeatedly extracts the local $y_t \in \mathbb{R}^3$7 spherical-harmonic neighborhood, computes $y_t \in \mathbb{R}^3$8, $y_t \in \mathbb{R}^3$9, and $y_t$0, predicts the next orientation, and advances the streamline by one voxel. Stopping occurs when the streamline leaves the mask or when the curvature-based angular threshold exceeds $y_t$1. This makes the system operationally deterministic at inference, even though the orientation predictor is trained as a conditional generative model.

5. Empirical evaluation and reported performance

The paper reports an ablation on 5 unseen HCP-YA subjects comparing temporal-only, spatial-plus-temporal, spatial-plus-temporal-plus-generative, and full DDTracking variants. Cluster detection rate rises from 93.73% for the temporal-only model to 95.62% for the full model; tract detection rate is 100% for all ablations; and $y_t$2 increases from $y_t$3 to $y_t$4. This progression is used to support the claim that both the generative formulation and the explicit local conditioning improve tract reconstruction quality (Li et al., 6 Aug 2025).

On the ISMRM 2015 Tractography Challenge, DDTracking reports $y_t$5 valid connections, $y_t$6 invalid connections, $y_t$7 no-connections, $y_t$8 overlap, $y_t$9 overreach, and $S=\{p_1,p_2,\dots,p_n\},$0 F1. The paper interprets this as best performance in valid connections, invalid connections, no-connections, and overlap, while not being best in overreach. On TractoInferno, DDTracking reports Dice $S=\{p_1,p_2,\dots,p_n\},$1, Overlap $S=\{p_1,p_2,\dots,p_n\},$2, and Overreach $S=\{p_1,p_2,\dots,p_n\},$3. The paper describes Dice as tied for the highest reported value and Overlap as the highest, again with increased overreach.

Generalization is assessed on HCP-YA, PPMI, ABIDE, CNP, BrainTumor, and Stroke. Reported tract detection rates are 100%, 99%, 99%, 94%, 95%, and 81%, respectively. Reported $S=\{p_1,p_2,\dots,p_n\},$4 overlap with UKF, iFOD2, and TractSeg remains above 0.72 across most datasets, with values ranging roughly from 0.72 to 0.92 depending on comparator and cohort. The paper uses these results to argue that DDTracking generalizes across health conditions, age groups, imaging protocols, and scanner types.

6. Significance, position in the literature, and limitations

DDTracking is presented as a tractography system that combines three elements often treated separately in prior work: local spatial modeling, global temporal modeling, and continuous generative orientation prediction. In that sense, it occupies a distinctive position relative to deterministic tracking rules, probabilistic model-based tracking, RNN-based trackers such as Learn to Track, reinforcement-learning approaches such as Track-to-Learn, and transformer-based streamline predictors. Its defining move is to make the next-step orientation a denoising target conditioned on both local diffusion evidence and streamline history (Li et al., 6 Aug 2025).

The framework’s reported strengths are sensitivity and continuity: it improves valid connections and overlap metrics, attains strong cluster and tract detection rates, and maintains broad cross-dataset generalization. A plausible implication is that the diffusion formulation acts as a trajectory prior that regularizes local ambiguity without discarding local evidence. At the same time, the paper also identifies several limitations. It uses no anatomical priors such as ACT; the temporal encoder is a GRU rather than a Transformer or Mamba-style alternative; the spherical-harmonic representation is mainly suited to single-shell data; and the current implementation predicts a single next orientation rather than a multi-sample probabilistic set of continuations.

The benchmark results also indicate a trade-off between coverage and specificity. DDTracking often improves overlap and valid-connection measures while showing comparatively high overreach, particularly on TractoInferno. That pattern suggests a method oriented toward recovering more anatomically plausible connectivity may still require stronger termination control or anatomical constraints to improve specificity. Within the tractography literature, DDTracking is therefore most accurately understood as a generative local-global streamline predictor: anatomically ambitious, empirically strong, and explicitly designed to replace handcrafted propagation rules with end-to-end learned conditional denoising.