scidata-benchmark/Care2D

This dataset contains 24 pairs of 2D fluorescence microscopy images of DCIS.COM Lifeact-RFP cells (F-actin), each acquired at low and high signal-to-noise ratio (SNR). It was released as the example dataset for training and testing CARE (content-aware image restoration) denoising networks in ZeroCostDL4Mic (see its Wiki for details). The task is to restore the high-SNR image from the low-SNR one.

Usage

One HDF5 file per split: train.h5, test.h5. Download only the split you need and index it like a numpy array; only the samples you access are read from disk.

from huggingface_hub import hf_hub_download
import h5py

f = h5py.File(hf_hub_download("scidata-benchmark/Care2D", "test.h5", repo_type="dataset"))

len(f["low_snr"])        # 2
f["low_snr"][0]          # (1024, 1024) float32, network input
f["high_snr"][0]         # (1024, 1024) float32, target
f["name"].asstr()[:]     # ['23', '24']

With 🤗 datasets: load_dataset("scidata-benchmark/Care2D", data_files={"test": "test.h5"}, split="test", streaming=True).with_format("numpy").

Splits

The original train/test split is kept as is.

split	pairs	images
train	22	01–22
test	2	23–24

Fields

Each split file contains, for its N pairs:

field	description	dtype / shape
low_snr	low signal-to-noise image (input)	float32, N × 1024 × 1024
high_snr	high signal-to-noise image of the same field of view (target)	float32, N × 1024 × 1024
name	image identifier (original file {name}.tif)	bytes, N

Pixel size is 40 nm × 40 nm. Values are the original 32-bit intensities, not normalized: the two channels have different ranges (low SNR roughly 0–5,000, high SNR up to ~64,000, with a few small negative values), so normalize each image before training or comparing. Samples are sorted by name, and low_snr[i] and high_snr[i] always form a pair.

Acquisition

Modality	Structured Illumination Microscopy (SIM), 60× 1.42 NA objective
Sample	DCIS.COM Lifeact-RFP cells
Label	Lifeact-RFP (fluorescence)
Image size	1024 × 1024 pixels, 40 nm pixel size

Attribution

Source: Zenodo

Original dataset: https://doi.org/10.5281/zenodo.3713330 ("ZeroCostDL4Mic - CARE (2D) example training and test dataset")

Original author: Guillaume Jacquemet

If you use this dataset, please cite:

@dataset{jacquemet_2020_3713330,
  author    = {Jacquemet, Guillaume},
  title     = {ZeroCostDL4Mic - CARE (2D) example training and test dataset},
  publisher = {Zenodo},
  year      = {2020},
  doi       = {10.5281/zenodo.3713330},
}