Code for [CDMRI 2025] "Streamline Signature Net(SSN): Efficient White Matter Pathway Recognition for Bundles Parcellation Using Path Signature", Computational Diffusion MRI https://link.springer.com/chapter/10.1007/978-3-032-12837-9_4

Uploaded all files. But comments and instructions are still under developing.

License

This repository contains code under two licenses:

• File in src/new/lars.py is licensed under Apache-2.0, see LICENSE.Apache-2.0
• Files in src/ except src/new/lars.py are licensed under BSD-3-Clause, see LICENSE.BSD-3-Clause

run

conda create --name SSN python=3.12 conda activate SSN

pip install git+https://github.com/SlicerDMRI/whitematteranalysis.git

pip install torch torchvision torchaudio pip install argparse signatory h5py matplotlib scikit-learn==1.5.2 conda install -c conda-forge libstdcxx-ng=13

git clone https://github.com/RenchZhao/Streamline_Signature_Net.git cd Streamline_Signature_Net

if need to prepare datasets

conda create -n pnlpipe3 python=3.6 conda install -c mrtrix3 mrtrix3

⚠️ Important Bug Fix: LPS/RAS Coordinate Transformation

Problem Description

When using MRtrix3's tckconvert to generate VTK files from TCK tractography, the output coordinates are in LPS (Left-Posterior-Superior) coordinate system. However, the whitematteranalysis (wma) library reads these coordinates and labels them as RAS (Right-Anterior-Superior) without actual transformation.

The original inference.py contains a "bug-that-works-by-accident":

def RAS2LPS_transform(points):
    points[:,:,0] = -points[:,:,0]  # In-place modification!
    points[:,:,1] = -points[:,:,1]  # In-place modification!
    return points

This in-place modification accidentally converts LPS→RAS when data.numpy() is saved, making the output VTK files loadable by dipy. However, this is incorrect design because:

1. It relies on unintended side effects (numpy view sharing)
2. It modifies data during inference instead of preprocessing
3. It breaks when using global index extraction (like in StreamlineBased migration)

Recommended Fix

Modify InferVtkDataset to apply coordinate transformation during initialization:

class InferVtkDataset(data.Dataset):
    def __init__(self, vtk_dataset, script_name='<inference>', normalization=False,
                 data_augmentation=False, transform=None):
        pd_tract = wma.io.read_polydata(vtk_dataset)
        
        # Apply transform during feature generation (LPS→RAS conversion)
        if transform:
            self.features = np.asarray(transform(gen_features(pd_tract)))
        else:
            self.features = np.asarray(gen_features(pd_tract))
        
        # ... rest of initialization

Then in inference calls:

• Pass the coordinate transform to InferVtkDataset
• Set data_transform=None in model_inference() to avoid in-place modification during inference

# Define LPS→RAS transform
def LPS_to_RAS_transform(points):
    feat = points.copy()
    feat[:,:,0] = -points[:,:,0]  # L → R
    feat[:,:,1] = -points[:,:,1]  # P → A
    return feat

# Create dataset with transform
test_dataset = InferVtkDataset(input_vtk, transform=LPS_to_RAS_transform)

# Run inference without transform
test_predicted_lst, pred_time, streamline_di = model_inference(
    model, test_data_loader, streamline_di, label_names,
    script_name, logger, device, thresh,
    data_transform=None,  # No transform during inference!
    result_transform=result_transform
)

This approach:

1. ✅ Properly separates preprocessing from inference
2. ✅ Avoids in-place modification side effects
3. ✅ Makes the coordinate system explicit and traceable
4. ✅ Works correctly with global index extraction

---

Possible Runtime Errors

ImportError: /home/user/anaconda3/envs/SSN/bin/../lib/libstdc++.so.6: version `GLIBCXX_3.4.32' not found (required by /home/user/anaconda3/envs/SSN/lib/python3.12/site-packages/signatory/_impl.cpython-312-x86_64-linux-gnu.so) run: conda install -c conda-forge libstdcxx-ng=13

"cross_val_txt" variable in new/gen_train_h5.py contains subjects of 5-fold cross validation of HCP dataset: https://zenodo.org/records/1285152 You can copy contents below without quotation mark to your own "5_fold.txt": ''' fold1 = ['992774', '991267', '987983', '984472', '983773', '979984', '978578', '965771', '965367', '959574', '958976', '957974', '951457', '932554', '930449', '922854', '917255', '912447', '910241', '907656', '904044'] fold2 = ['901442', '901139', '901038', '899885', '898176', '896879', '896778', '894673', '889579', '887373', '877269', '877168', '872764', '872158', '871964', '871762', '865363', '861456', '859671', '857263', '856766'] fold3 = ['849971', '845458', '837964', '837560', '833249', '833148', '826454', '826353', '816653', '814649', '802844', '792766', '792564', '789373', '786569', '784565', '782561', '779370', '771354', '770352', '765056'] fold4 = ['761957', '759869', '756055', '753251', '751348', '749361', '748662', '748258', '742549', '734045', '732243', '729557', '729254', '715647', '715041', '709551', '705341', '704238', '702133', '695768', '690152'] fold5 = ['687163', '685058', '683256', '680957', '679568', '677968', '673455', '672756', '665254', '654754', '645551', '644044', '638049', '627549', '623844', '622236', '620434', '613538', '601127', '599671', '599469'] '''

new/get_label_dict.py is used to generate "label_di" variable in new/gen_train_h5.py

cite

@inproceedings{zhao_streamline_2026, address = {Cham}, title = {Streamline {Signature} {Net} ({SSN}): {Efficient} {White} {Matter} {Pathway} {Recognition} for {Bundles} {Parcellation} {Using} {Path} {Signature}}, isbn = {978-3-032-12837-9}, shorttitle = {Streamline {Signature} {Net} ({SSN})}, doi = {10.1007/978-3-032-12837-9_4}, language = {en}, booktitle = {Computational {Diffusion} {MRI}}, publisher = {Springer Nature Switzerland}, author = {Zhao, Renzhi and Zhang, Xin and Tan, Zihao and Xu, Jiakun and Yang, Zhenyu and Wu, Ye and Xu, Xiangmin}, editor = {Chamberland, Maxime and Chen, Yuqian and Filipiak, Patryk and Hendriks, Tom and Lv, Jinglei and Shailja, S. and Thompson, Elinor}, year = {2026}, keywords = {Diffusion MRI, Path Signature, Bundles Parcellation, Tractogram}, pages = {30--43}, }