RapidParc 1.0.2

Rapid registration-free tractogram parcellation using global context

RapidParc

RapidParc is a fast, accurate, and lesion-robust transformer-based algorithm for registration-free parcellation of streamlines from diffusion MRI tractography. It assigns one label to each streamline in a tractogram and is designed for efficient large-scale inference.

If you use this repository, please cite:

Justus Bisten, Valentin von Bornhaupt, Johannes Grün, Tobias Bauer, Theodor Rüber, Thomas Schultz.
RapidParc: A Global-Context Transformer for Parallel, Accurate, and Lesion-Robust Tractogram Parcellation.
Imaging Neuroscience (2026).
DOI: 10.1162/IMAG.a.1168

Table of Contents

• Problem Description
• Method
• Installation
• Quick Start
• Reproduce Paper Results
• Training
• Known Issues

Problem description

Tractogram parcellation assigns anatomical labels to streamlines obtained from diffusion MRI tractography. RapidParc performs this task without anatomical registration and uses a transformer architecture to incorporate global context across streamlines.

Method

At inference time, streamlines are shuffled and processed in groups of size eval_context_size. This makes it possible to scale to large tractograms while retaining contextual information beyond the individual streamline. The method is registration-free, which is especially useful when anatomical deformation, lesions, or post-surgical alterations reduce the reliability of registration-based pipelines.

Key Features:

• Registration-free inference
• Fast, parallelized tractogram processing
• Strong benchmark performance on the TractCloud test split
• Improved robustness in anatomically altered or lesioned brains
• Simple usage from both Python and command line

RapidParc predicts one of 43 classes per streamline, including an outlier class, Other.

Pretrained models:

Model	Intended use
rapidparc	Default general-purpose model for standard inference.
hemiaug	When your data may contain strong unilateral anatomical changes, lesions, or post-surgical alterations.
rapidparc_v8	When you want to study the effect of using 8 instead of 15 resampled vertices.

Installation

This project was developed with PyTorch 2.4 on Python 3.12.3 and CUDA 12.4.

1. (optional, recommended) Create and activate a virtual environment

python3 -m venv .venv 
source .venv/bin/activate
pip install -U pip

2. Install PyTorch

Install PyTorch using pip by following the official instructions: https://pytorch.org/get-started/locally/

3. Install RapidParc

pip install RapidParc

Quick start

Python Interface

To parcellate a tractogram in Python:

from RapidParc import RapidParc

predictions = RapidParc(...)

The RapidParc function accepts multiple input-formats like list, torch.Tensor and numpy.ndarray. Input streamlines do not need to have the same number of supporting points. RapidParc resamples them automatically.

Example:

from RapidParc import RapidParc
import torch

# Your tractogram goes here
inputTractogram = torch.rand([5000, 15, 3])

predictions = RapidParc(
    model_name_or_path = "rapidparc",
    inputTractogram = inputTractogram,
    eval_batch_size = 128,
    eval_context_size = 2000,
    device = torch.device("cpu"),
    print_time = True,
    print_class_distribution = False
)

print(f"Output shape: {predictions.shape}")
print(f"Output: {predictions}")
classes, counts = torch.unique(predictions, return_counts=True)
print(f"Output summary: {[f'{pred}: {num}' for pred, num in zip(classes, counts)]}")

Time to load model: 0.05s
Time to shuffle data: 0.00s
5000 Streamlines are loaded and preprocessed.
100%|█████████████████████████████████████| 1/1 [00:00<00:00,  5.28it/s]
Prediction is done in 0.20s
Total time: 0.27s

Output shape: torch.Size([5000])
Output: tensor([42, 42, 42,  ..., 42, 42, 42])
Output summary: ['33: 1', '42: 4999']

In this example, randomly generated streamlines are mostly classified as outliers, which is expected.

Command-line usage (.tck tractograms)

If your tractogram is stored as a .tck file, you can run parcellation directly from the terminal. Each assigned tract results in a new .tck output file inside the specified directory.

Example:

rapidparc --tck_path tractogram.tck \
 --out_path parcellated_tractogram \
 --model_name_or_path rapidparc \
 --eval_batch_size 64 \
 --eval_context_size 2000 \
 --device cpu \
 --print_time \
 --print_class_distribution

Time to load tck file with 926641 streamlines: 1.50s
Time to shorten streamlines: 5.54s
Time to load model: 0.07s
Time to shuffle data: 0.08s
926641 Streamlines are loaded and preprocessed.
100%|█████████████████████████████████████| 8/8 [00:12<00:00,  1.60s/it]
Prediction is done in 12.83s
Class distribution
      Class       | Count
 0 AF                (6254) ▇▇▇▇▇▇
 1 CB               (14252) ▇▇▇▇▇▇▇▇▇▇▇▇▇
 2 EC                (3795) ▇▇▇▇
 3 EmC               (5990) ▇▇▇▇▇▇
 4 ILF              (17420) ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
 5 IOFF             (19687) ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
 6 MdLF             (10216) ▇▇▇▇▇▇▇▇▇▇
 7 SLF-I            (10937) ▇▇▇▇▇▇▇▇▇▇
 8 SLF-II           (28028) ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
 9 SLF-III           (5348) ▇▇▇▇▇
10 UF                (4185) ▇▇▇▇
11 CST              (24662) ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
12 CR-F             (14493) ▇▇▇▇▇▇▇▇▇▇▇▇▇
13 CR-P              (2089) ▇▇
14 SF               (23246) ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
15 SO                (5810) ▇▇▇▇▇▇
16 SP                (5294) ▇▇▇▇▇
17 TF               (40690) ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
18 TO                (4753) ▇▇▇▇▇
19 TT                (8700) ▇▇▇▇▇▇▇▇
20 TP                (9958) ▇▇▇▇▇▇▇▇▇
21 PLIC              (4345) ▇▇▇▇
22 CC1               (1477) ▇▇
23 CC2              (18125) ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
24 CC3               (7703) ▇▇▇▇▇▇▇
25 CC4               (9848) ▇▇▇▇▇▇▇▇▇
26 CC5               (7267) ▇▇▇▇▇▇▇
27 CC6              (18489) ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
28 CC7               (8638) ▇▇▇▇▇▇▇▇
29 CPC                (875) ▇
30 ICP                (866) ▇
31 Intra-CBLM-I-P    (6059) ▇▇▇▇▇▇
32 Intra-CBLM-PaT   (10004) ▇▇▇▇▇▇▇▇▇
33 MCP                (929) ▇
34 Sup-F            (56703) ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
35 Sup-FP            (5456) ▇▇▇▇▇
36 Sup-O             (7084) ▇▇▇▇▇▇▇
37 Sup-OT            (6879) ▇▇▇▇▇▇▇
38 Sup-P            (23330) ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
39 Sup-PO            (6748) ▇▇▇▇▇▇
40 Sup-PT           (19768) ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇
41 Sup-T            (11187) ▇▇▇▇▇▇▇▇▇▇
42 Outlier 'Other' (429054) ###################################################
Total time: 13.28s
Files saved in parcellated_tractogram/

The same function is also available from Python:

from RapidParc import RapidParcTckEval

RapidParcTckEval(...)

Reproducing results

To benchmark RapidParc on the TractCloud test split, use RapidParc.test. The RapidParc.test function automatically fetches the TractCloud dataset during evaluation.

Example:

import RapidParc
import torch

acc_43, f1_43 = RapidParc.test(
    model_name_or_path = "rapidparc_v8",
    applyTestSetAugmentations = False,
    eval_batch_size = 128,
    eval_context_size = 2000,
    device = torch.device("cpu"),
    print_classification_report = True
)
print(f"Accuracy: {100 * acc_43:.2f} %, Macro F1 Score: {100 * f1_43:.2f} %")

This prints out a classification report and creates a confusion matrix plot in rapidParc_v8/plots. You can also execute this via the CLI using rapidparc-test.

Retrain RapidParc on the TractCloud train split

To retrain RapidParc on the TractCloud training split, a minimal run can be started using:

rapidparc-train

A list of all configurable attributes can be obtained from cli_train.py. The parameters we used for each model can be found in the .yaml files attached to the GitHub releases.

Known issues

• Orientation Errors: If evaluation results are unexpectedly poor, your tractogram may have an axis flip or coordinate-system mismatch.
• Outlier Class Ratio: The Other class may contain a large fraction of streamlines, especially in whole-brain tractograms. This is often expected and wanted.