FISH_spots Dataset

The manually verified in situ hybridization fluorescence images and point coordinate dataset. This dataset contains images and annotations for the task of single-molecule fluorescence in situ hybridization (FISH) spot detection, supporting 2D, 3D, and simulated noisy data. The structure is designed for deep learning model development, training, and evaluation.

Directory Structure

FISH_spots/
├── 2d/
│ ├── csv/
│ ├── image/
│ ├── image_raw/
│ └── meta.csv
├── 3d/
│ ├── csv/
│ ├── image/
│ ├── meta_test.csv
│ └── meta_valid.csv
└── noise/
├── csv/
├── image/
├── meta.csv
├── meta_test.csv
└── meta_valid.csv

Download

# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/datasets/GangCaoLab/FISH_spots

Data Content and Annotation Format

• image directory
  Contains 512×512 image patches used for training. The images are ready for use by deep learning models, in standard formats such as PNG or TIFF.

• csv directory
  Stores CSV files containing annotated spot positions for each corresponding image patch. Each row records the coordinates and other metadata of detected or simulated spots.

• meta.csv*
  meta.csv, provide global metadata.

2D Dataset

• Includes data used in the U-FISH publication, comprising both real experimental data and simulated data.
• Covers 7 different experimental sources, providing diversity (real and simulated conditions).
• Contains over 4,000 images and 160,000 annotated spot positions.
• Suitable for benchmarking model generalization and spot detection in varied scenarios.

3D Dataset

• Mainly generated via simulation, representing spot distributions under 3D imaging conditions.
• Provides 3D image patches and coordinate-based annotations.
• Includes meta_test.csv and meta_valid.csv for easy split management.

Noise Dataset

• Contains simulated noisy images aimed at testing model robustness to noise.
• Structure mirrors the 3D dataset, with patches, annotations, and metadata files.

Citation

If you use this dataset in your research, please cite the U-FISH paper.