Tractography-Driven Synthetic Data Generation for Fiber Bundle Segmentation in Tracer Histology (MICCAI)

About

This is a Pytorch implementation for the paper "Tractography-Driven Synthetic Data Generation for Fiber Bundle Segmentation in Tracer Histology" (MICCAI 2026) by Kyriaki-Margarita Bintsi, Sparsh Makharia, Yaël Balbastre, Joselyn Romero Avila, Julia F. Lehman, Suzanne N. Haber, and Anastasia Yendiki

Training mixes real and synthetic patches in each batch according to a configurable ratio (--synthetic_ratio); validation always uses held-out real data only.

This repo does not ship any data or trained checkpoints — the commands below only run once you point them at your own real data, registered .trk bundle files, and blockface NIfTI volume (see "Data" below).

Repository layout

• train.py — main training entry point (real + synthetic mixed training).
• model.py — FlexibleUNet, an nnUNet-style 2D Lightning UNet for the binary segmentation task.
• losses.py — DiceBCELoss (the loss used for all reported training runs).
• datamodules/ — datamodule.py / dataset.py: Lightning datamodule/datasets. Real data is loaded from nnUNet-style paired PNG directories; synthetic data is generated on-the-fly per batch via Syntract (see below) with zero on-disk caching.
• predict.py — sliding-window inference with Gaussian-weighted patch blending and nnUNet-style mirroring TTA. Loads a single checkpoint by default; pass --kfold only if you have a 5-fold cross-validation ensemble to load.
• evaluate.py — computes the paper's reported numbers (Dice, IoU, clDice, cluster-wise sensitivity/FDR). Reads raw OME-Zarr histology + outline/fiber-class masks per subject. Not a CLI — edit the "Adapt this part" block at the top of the file for your subject/paths, then run it.
• visualize_predictions.py — qualitative visualization: overlays predicted vs. ground-truth fiber-bundle boundaries on the histology slide for a subject and saves a figure. Also not a CLI — same "Adapt this part" pattern.

Setup

pip install -r requirements.txt

Data synthesis (Syntract)

Synthetic patch generation is provided by Syntract (Sparsh Makharia, LINC Team), which turns tractography (.trk) + a registered NIfTI volume into dark-field-microscopy-style patches with matching fiber masks.

Clone it directly into this repo's root:

git clone https://github.com/Sparsh57/Syntract.git
pip install -r Syntract/requirements.txt

This gives you a Syntract/ folder alongside train.py.

Data

None of the paths below ship with this repo — supply your own:

• Real training/validation data: nnUNet-style directories of paired PNG images + binary masks (--real_image_dir, --real_label_dir, --val_image_dir, --val_label_dir).
• Registered tractography: a directory of .trk files already registered into the same space as your blockface volume (--trk_dir).
• Blockface volume: the registered NIfTI volume synthetic patches are extracted from (--input_nifti).
• Optionally, a white matter mask NIfTI (--white_mask).
• For evaluate.py and visualize_predictions.py (edited directly in each script, see below): a root directory containing sub-<SUBJECT>/micr/*.ome.zarr histology + masks/*.ome.zarr (Outline, Fiber_dense_bundle, Fiber_moderate_bundle, Fiber_light_bundle) per evaluated subject.

Training

python train.py \
    --real_image_dir /path/to/imagesTr/ \
    --real_label_dir /path/to/labelsTr/ \
    --val_image_dir /path/to/imagesTs/ \
    --val_label_dir /path/to/labelsTs/ \
    --input_nifti /path/to/registered_blockface_volume.nii.gz \
    --synthetic_ratio 0.7 --epochs 200 --batch_size 8 --batches_per_epoch 80 \
    --checkpoint_dir results_mixed_training/my_run/ --wandb_name my_run \
    --lr 1e-4 --warmup_epochs 2

Inference / evaluation

python predict.py \
    --input_folder /path/to/imagesTs_SUBJECT/ \
    --output_folder results_mixed_training/evaluate/SUBJECT/my_run/ \
    --model_folder results_mixed_training/my_run/ \
    --saved_model best

evaluate.py and visualize_predictions.py are not CLIs: edit the subject, predicted_folder (or predicted_dir), dirname, and slides variables in the "Adapt this part" block near the top of each file for your data, then run e.g. python evaluate.py.

Citation

If you use this code, please cite our paper:

@article{bintsi2026tractography,
  title={Tractography-Driven Synthetic Data Generation for Fiber Bundle Segmentation in Tracer Histology},
  author={Bintsi, Kyriaki-Margarita and Makharia, Sparsh and Balbastre, Ya{\"e}l and Avila, Joselyn Romero and Lehman, Julia F and Haber, Suzanne N and Yendiki, Anastasia},
  journal={arXiv preprint arXiv:2606.26898},
  year={2026}
}

and Syntract for the data synthesis pipeline.