DL-based tractography with spatial guidance for public (non-)pathological datasets

This record provides pretrained bundle-specific recurrent neural network tractography models and derived streamline tractograms for the arcuate fasciculus (AF), corticospinal tract (CST), and optic radiation (OR). The release includes diffusion-only and spatial-coordinate-augmented gated recurrent unit (GRU) models. Models were trained on the TractoInferno training split and evaluated on the TractoInferno test set, the BTC preoperative brain-tumor dataset, and a subset of the UCSF-PDGM dataset. Resulting tractograms are provided separately as MRtrix .tck files. Input data and preprocessing All diffusion MRI data were harmonized to a common Repulsion100 acquisition scheme with 100 gradient directions using spherical-harmonic resampling. Diffusion-weighted images were normalized by the corresponding b = 0 s/mm² image. At each streamline point, the model receives interpolated local diffusion information: diffusion-only input: (batch_size, 1, 100) diffusion and spatial-coordinate input: (batch_size, 1, 103) For coordinate-augmented models, three normalized spatial coordinates are appended to the diffusion signal. Coordinates were obtained by deformable atlas registration of T1-weighted MRI to the MNI152 atlas. The resulting atlas-space coordinates were normalized to the range [0, 1]. For BTC, only the b = 1200 s/mm² shell was used to better match the TractoInferno acquisitions, which use b-values between 700 and 1000 s/mm². Model architecture and training The models use GRU-based recurrent neural networks for point-wise streamline direction prediction. Depending on the model configuration, networks contain one to three GRU layers with 512 hidden units, layer normalization, and recurrent dropout of 0.3. A final fully connected layer predicts a continuous fiber-direction vector. Training was performed bundle-wise using TractoInferno reference streamlines (resampled to 1 mm step size) using the original TractoInferno split. Models were optimized for 20 epochs using mean squared error between predicted and reference directions, the Adam optimizer with a learning rate of 0.0001, gradient clipping with norm 5, and batches of 1,000 streamlines. During training, reference coordinates were augmented with zero-mean Gaussian noise with σ = 0.03 mm. Tractography generation Tracking was initialized using TractSeg v2.8 seed masks and default TractSeg seeding, cropping, and filtering procedures. Seed masks were resampled to generate five equally spaced seed points per cubic millimeter. Uniform spatial jitter sampled from U[-0.5, 0.5] was added to each seed coordinate, and one streamline was generated per seed point. Tracking used a step size of 1.0 mm and stopped when either: a streamline reached a maximum length of 250 mm, or the streamline left the corresponding TractSeg bundle mask. Generated streamlines were subsequently filtered using the corresponding TractSeg ending mask. Applying the pretrained models Each pretrained model is specific to a white-matter bundle and input configuration. Select the checkpoint corresponding to the desired bundle and model type: Diffusion-only models: input shape (batch_size, 1, 100) Coordinate-augmented models: input shape (batch_size, 1, 103) The following steps are recommended: - resample diffusion MRI data to the Repulsion100 gradient scheme and normalized by the corresponding b = 0 s/mm² image - at every tracking point, interpolate the 100 diffusion features from the preprocessed dMRI data - for coordinate-augmented models, additionally interpolate or transform the point location into normalized MNI152 atlas space and append the three coordinates, scaled to [0, 1] - tracking requires seed points and a bundle-specific tracking mask in the same physical space as the dMRI data (to reproduce the study protocol, use TractSeg seed masks) - initialize the recurrent model state and iteratively predict a fiber direction - normalize the predicted direction vector and advance the streamline by 1.0 mm per step - stop tracking when the streamline leaves the corresponding bundle mask or reaches a to-be-defined maximum length (250 mm have been used in this study) - optionally, perform further filtering of the reconstructed filtering (using e.g. TractSeg masks) Included data This release contains: pretrained model checkpoints for AF, CST, and OR (diffusion-only and coordinate-augmented model variants) tractography results for the TractoInferno test set tractography results for the BTC dataset tractography results for the selected UCSF-PDGM subset Source datasets Models and tractograms are based on the following publicly available datasets: TractoInferno (https://openneuro.org/datasets/ds003900/versions/1.1.1) preoperative Brain Tumor Connectomics (https://openneuro.org/datasets/ds001226/versions/5.0.1) University of California San Francisco Preoperative Diffuse Glioma MRI (https://www.cancerimagingarchive.net/collection/ucsf-pdgm/) This work is a retrospective secondary analysis of publicly available data. No new human participant data were acquired.