Datasets:

HTW-KI-Werkstatt
/

SynthMT

pretty_name: SynthMT
annotations_creators:
  - no-annotation
language:
  - en
license: cc-by-4.0
tags:
  - microscopy
  - synthetic-data
  - microtubules
  - irm
  - tirf
  - biology
  - instance-segmentation
  - computer-vision
  - foundation-models
  - segmentation
task_categories:
  - image-segmentation
task_ids:
  - instance-segmentation
size_categories:
  - 1K<n<10K
dataset_info:
  features:
    - name: id
      dtype: string
    - name: image
      dtype: image
    - name: mask
      list: image
  splits:
    - name: train
      num_bytes: 3984597522
      num_examples: 6600
  download_size: 3963881685
  dataset_size: 3984597522
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

SynthMT: A Synthetic Benchmark for Automated Microtubule Segmentation

Authors: Mario Koddenbrock*, Justus Westerhoff*, Dominik Fachet, Simone Reber, Felix Gers, Erik Rodner
- *Equal contribution
Affiliations: HTW Berlin, BHT Berlin, MPI for Infection Biology

Project Page: https://datexis.github.io/SynthMT-project-page/ Code & Pipeline: https://github.com/ml-lab-htw/SynthMT
Paper: Synthetic Data Enables Human-Grade Microtubule Analysis with Foundation Models for Segmentation
Dataset: https://huggingface.co/datasets/HTW-KI-Werkstatt/SynthMT

*Equal contribution

🧬 Overview

SynthMT is a synthetic, expert-validated dataset designed to benchmark segmentation models on in vitro microtubule (MT) images visualized in interference reflection microscopy (IRM)–like conditions.
It provides:

Realistic synthetic MT images
Pixel-perfect instance segmentation labels
A generation pipeline that adapts to any real microscope domain without the need for ground-truth annotations
A comprehensive benchmark of classical (FIESTA), microscopy-specialized (StarDist, TARDIS, µSAM, CellSAM, Cellpose-SAM), and general-purpose foundation models (SAM, SAM2, SAM3, SAM3Text)

A core result of the associated paper:
→ SAM3Text, prompted with “thin line” and tuned on only 10 synthetic images, achieves human-grade performance on unseen real data.

🖼 Example from the Dataset

(a) IRM-like synthetic image	(b) Ground-truth instance mask
(c) FIESTA anchor point predictions	(d) SAM3Text segmentation

Overview of SynthMT and example predictions.

📦 Dataset Structure

Each sample in SynthMT contains:

Field	Type	Description
`id`	`string`	Unique image identifier
`image`	`Image`	Synthetic IRM-like image, decoded from PNG. Can be converted to a numpy array `(H, W, 3)` for in-memory processing.
`mask`	`Array3D`	Stack of instance masks with shape `(C, 512, 512)` and `uint16` dtype, where `C` = number of instances in the image. Background pixels = 0. Stored in the dataset as a `Sequence(Image())` but can be stacked in memory for in-memory pipelines.

🧫 Biological Motivation

Microtubules (MTs) are cytoskeletal filaments essential for intracellular transport, cell motility, and mitotic spindle formation. Measuring MT count, length, and curvature is critical for in vitro reconstitution experiments, drug discovery, and mechanistic cell biology.

However:

Manual MT annotation is time-consuming and unscalable
IRM/TIRF imaging varies significantly across labs (domain shift)
No large, labeled benchmarks existed for MT segmentation

SynthMT directly addresses this gap.

📥 Installation & Loading

Install the Hugging Face datasets library:

pip install datasets

Load the dataset entirely in memory with masks stacked:

from datasets import load_dataset
import numpy as np

ds = load_dataset("HTW-KI-Werkstatt/SynthMT", split="train")

sample = ds[0]

# Image as numpy array (H, W, 3) = (512, 512, 3)
img_array = np.array(sample["image"].convert("RGB"))

# Masks as stacked numpy array (C, H, W)
mask_stack = np.stack([np.array(mask.convert("L")) for mask in sample["mask"]], axis=0)

No disk I/O is required — everything can be used in-memory.

🔗 Links

Project Page: https://datexis.github.io/SynthMT-project-page/
This Dataset: https://huggingface.co/datasets/HTW-KI-Werkstatt/SynthMT
Real IRM in-vitro MT dataset (also used in our paper): https://huggingface.co/datasets/HTW-KI-Werkstatt/IRM-in-vitro-microtubules
Code & Generation Pipeline: https://github.com/ml-lab-htw/SynthMT
Paper: https://doi.org/10.1371/journal.pcbi.1013901

📄 Citation

If you use this dataset, please cite:

@article{koddenbrock2026synthetic,
    doi = {10.1371/journal.pcbi.1013901},
    author = {Koddenbrock, Mario AND Westerhoff, Justus AND Fachet, Dominik AND Reber, Simone AND Gers, Felix A. AND Rodner, Erik},
    journal = {PLOS Computational Biology},
    publisher = {Public Library of Science},
    title = {Synthetic data enables human-grade microtubule analysis with foundation models for segmentation},
    year = {2026},
    month = {05},
    volume = {22},
    url = {https://doi.org/10.1371/journal.pcbi.1013901},
    pages = {1-25},
    abstract = {Studying microtubules (MTs) and their mechanical properties is central to understanding intracellular transport, cell division, and drug action. While important, experts still need to spend many hours manually segmenting these filamentous structures. The suitability of state-of-the-art methods for this task cannot be systematically assessed, as large-scale labeled datasets are missing. We address this gap by introducing the synthetic dataset SynthMT, produced by tuning a novel image generation pipeline on real-world interference reflection microscopy (IRM) frames of in vitro reconstituted MTs without requiring human annotations. In our benchmark, we evaluate nine fully automated methods for MT analysis in both zero- and Hyperparameter Optimization (HPO)-based few-shot settings. Across both settings, classical algorithms and current foundation models still struggle to achieve the accuracy required for biological downstream analysis on in vitro MT IRM images that humans perceive as visually simple. However, a notable exception is the recently introduced SAM3 model. After HPO on only ten random SynthMT images, its text-prompted version SAM3Text achieves near-perfect and in some cases super-human performance on unseen, real data. This indicates that fully automated MT segmentation has become feasible when method configuration is effectively guided by synthetic data. To enable progress, we publicly release the dataset, the generation pipeline, and the evaluation framework.},
    number = {5}
}

🏷 License

CC-BY-4.0 - See LICENSE for details.

🙏 Acknowledgements

Our work is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) Project-ID 528483508 - FIP 12. We would like to thank Dominik Fachet and Gil Henkin from the Reber lab for providing data, and also thank the further study participants Moritz Becker, Nathaniel Boateng, and Miguel Aguilar. The Reber lab thanks staff at the Advanced Medical Bioimaging Core Facility (Charité, Berlin) for imaging support and the Max Planck Society for funding. Furthermore, we thank Kristian Hildebrand and Chaitanya A. Athale (IISER Pune, India) and his lab for helpful discussions