tomzhaosky/Cryo-Sim

CryoSim-SNR-0.05 Dataset

This repository contains a synthetic low-SNR Cryo-EM filament tracing dataset used by FPicker: Topology-Guided Evolution for Filament Tracing in Low-SNR Microscopy [arXiv] | [Springer]

The dataset is generated with the FPicker synthetic Cryo-EM simulator (data_generate.py). It follows a COCO-style layout and provides both grayscale microscopy images and topology-aware centerline annotations for helical filaments.

Overview

• Task: filament detection and tracing in low-SNR Cryo-EM images.
• Signal-to-noise ratio: SNR = 0.05 for this release folder, corresponding to approximately -13.0 dB under 10 * log10(SNR).
• Image format: 8-bit single-channel grayscale PNG.
• Image size: 1024 x 1024.
• Annotation format: COCO-style JSON with additional centerline topology fields.
• Category: one foreground class, helix (category_id = 1).
• Total size on disk: approximately 19.5 GiB.

Directory Structure

cryosim_snr_0.05/
|-- annotations/
|   |-- instances_train2025.json
|   |-- instances_val2025.json
|   `-- instances_test2025.json
|-- images/
|   |-- train2025/
|   |-- val2025/
|   `-- test2025/
`-- masks/
    |-- train2025/
    |-- val2025/
    `-- test2025/

Each image has a matching mask with the same file name under the corresponding split:

images/train2025/0_000001.png
masks/train2025/0_000001.png

File names follow the pattern:

<generation-rank>_<image-id>.png

The rank prefix records the distributed generation worker that produced the image.

Dataset Splits

Split	Images	Helix instances	Image files	Mask files	Defocus range
train2025	16,000	95,837	16,000	16,000	1.5000-2.9999 um
val2025	2,000	11,881	2,000	2,000	1.5001-2.9992 um
test2025	2,000	12,115	2,000	2,000	1.5003-3.0000 um
Total	20,000	119,833	20,000	20,000	-

The centerline annotations contain about 25 points per filament on average. Points are uniformly resampled along each projected filament with an approximate step size of 15 pixels.

Annotation Format

The annotation files are COCO-style JSON files with the following top-level fields:

{
  "info": {
    "year": 2025
  },
  "images": [],
  "annotations": [],
  "categories": [
    {
      "id": 1,
      "name": "helix"
    }
  ]
}

Each image record contains:

{
  "id": 1,
  "file_name": "0_000001.png",
  "width": 1024,
  "height": 1024,
  "defocus": 2.000996319120829
}

Each annotation record contains:

{
  "id": 1,
  "image_id": 1,
  "category_id": 1,
  "bbox": [519, 208, 306, 285],
  "area": 87210,
  "line": [[825.89, 208.36], [813.81, 217.21]],
  "resampled_line": [[825.89, 208.36], [813.81, 217.21]],
  "edges": [[0, 1]],
  "seed_point": [825.89, 208.36],
  "iscrowd": 0
}

Field meanings:

• bbox: bounding box around the centerline points in [x, y, width, height] format.
• line: ordered filament centerline points in image pixel coordinates.
• resampled_line: the uniformly resampled version of line; in this release it matches line.
• edges: sequential topology connections between centerline point indices.
• seed_point: the first point of the resampled centerline.
• defocus: simulated defocus value in micrometers.

Coordinates use the standard image convention: origin at the top-left corner, x increasing to the right, and y increasing downward.

Masks

The files under masks/ are 8-bit single-channel PNG masks generated from the projected centerlines. They are useful for visualization, debugging, and segmentation-style baselines. The JSON centerline annotations should be treated as the authoritative tracing target for topology-aware methods.

Generation Pipeline

The dataset was generated by the FPicker Cryo-EM simulator. At a high level, the generator:

1. Parses atomic coordinates from mmCIF structures.
2. Builds helical filament stacks using configured rise and twist parameters.
3. Applies straight, worm-like-chain, or sine-wave bending models.
4. Samples random 3D rotations and translations for each filament.
5. Projects atoms into a 2D image grid.
6. Adds structured distractors such as ice contamination blobs, linear artifacts, and carbon-edge-like high-contrast regions.
7. Applies Gaussian blur, contrast transfer function simulation, defocus variation, B-factor attenuation, and white noise according to the configured SNR.
8. Normalizes images with Z-score clamping and exports 8-bit grayscale PNGs.
9. Produces binary masks and uniformly resampled centerline annotations.

The generation script depends on the simulator configuration (SIM_CONFIG) used by FPicker. Important configurable factors include image resolution, SNR, defocus range, pixel size range, helical fiber definitions, bending mode, and split ratios.

Loading Example

import json
from pathlib import Path
from PIL import Image

root = Path("cryosim_snr_0.05")
split = "train2025"

with open(root / "annotations" / f"instances_{split}.json", "r") as f:
    coco = json.load(f)

image_info = coco["images"][0]
image = Image.open(root / "images" / split / image_info["file_name"])
mask = Image.open(root / "masks" / split / image_info["file_name"])

anns = [ann for ann in coco["annotations"] if ann["image_id"] == image_info["id"]]
centerlines = [ann["resampled_line"] for ann in anns]

Recommended Use

This dataset is intended for research on:

• low-SNR Cryo-EM filament detection;
• open-curve filament tracing;
• topology-preserving centerline prediction;
• segmentation, detection, and tracking baselines for filamentous structures;
• sim-to-real robustness studies for signal-starved scientific imaging.

Because the data are synthetic, models trained on this dataset should be validated carefully before deployment on real Cryo-EM micrographs.

Citation

If you use this dataset or the FPicker codebase, please cite:

@inproceedings{zhao2026fpicker,
  title={FPicker: Topology-Guided Evolution for Filament Tracing in Low-SNR Microscopy},
  author={Zhao, Tingyin and Huang, Mingtao and Shen, Yuan},
  booktitle={European Conference on Computer Vision},
  pages={116--133},
  year={2026},
  organization={Springer}
}

License

This dataset, including images, masks, annotations, and accompanying dataset documentation, is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

Commercial use is not permitted without prior written permission from the authors. See LICENSE for details.