benchmark-data / README.md
5xuekun's picture
Card: the NeurIPS22-CellSeg folder no longer serves a baselines benchmark
15be1eb verified
|
Raw History Blame Contribute Delete
3.85 kB
metadata
license: other
license_name: per-folder
pretty_name: SHAPE benchmark data
tags:
  - biology
  - microscopy
  - cell-segmentation
  - cell-painting
  - benchmark
size_categories:
  - 10K<n<100K

SHAPE benchmark data

Inputs of the public benchmarks in the SHAPE repository (benchmarks/). Each folder is laid out the way that benchmark's driver reads it, so a download into the repository's data/ folder is ready to use. From the repository root:

hf download SHAPE-agent/benchmark-data --repo-type dataset --include "NeurIPS22-CellSeg/*" --local-dir data
hf download SHAPE-agent/benchmark-data --repo-type dataset --include "CPJUMP1/*" --local-dir data
hf download SHAPE-agent/benchmark-data --repo-type dataset --include "JUMP/*" --local-dir data

A folder holds thousands of files, and the Hub limits how many requests a client may make in five minutes (fewer without a login). If hf download stops with 429 Too Many Requests, wait five minutes and run the same command again: it keeps the files already downloaded and fetches the rest. hf auth login raises the limit.

Folder Files Size Benchmark Content
NeurIPS22-CellSeg/ 2,001 3.4 GB segmentation_NeurIPS22-CellSeg, cropper_NeurIPS22-CellSeg images/ (1,000 images) and labels/ (<name>_label.tiff instance masks) from the NeurIPS 2022 Cell Segmentation Challenge; NeurIPS22-CellSeg-cell_image_metrics.csv, the per-image Cellpose-SAM AP50 that splits the images into hard (< 0.30), medium and easy (> 0.60) bands (213 / 171 / 616 images)
CPJUMP1/ 5,001 12.8 GB full_CPJUMP1 1,000 five-channel views (5,000 TIFFs) of the CPJUMP1 CRISPR plates BR00117000 and BR00117004 (A549 Cas9) under raw/<plate>/<well>/, and manifest.json, the selection the harness pins by sha256 (7cf7a28b...)
JUMP/ 1,158 15.4 GB batch_correction_JUMP well-level profiles of JUMP sources 2, 3, 6, 8 and 10 in the inputs/<source>/workspace/profiles/ layout of Arevalo et al., the compound, plate and well metadata (inputs/metadata/), and inputs/microscope_config.csv (jump-cellpainting/datasets at commit 181fa0dc)
ModalityProbe3/ 3,000 7.2 GB full_CPJUMP1 (ModalityProbe3) raw TIFFs of the six cell type x perturbation groups only, without the manifest; build this dataset with the repository's benchmarks/full_CPJUMP1/runners/download_modality_probe3.py

RxRx1, used by representation_RxRx1, is not included; obtain it from rxrx.ai.

Sources and licences

The folders keep the licence of the data they come from.

  • NeurIPS22-CellSeg/images, NeurIPS22-CellSeg/labels: the NeurIPS 2022 Cell Segmentation Challenge dataset (Zenodo 10719375), CC BY-NC-ND 4.0. Cite Ma, J. et al. The multimodality cell segmentation challenge: toward universal solutions. Nat. Methods 21, 1103-1113 (2024).
  • CPJUMP1/raw, ModalityProbe3/raw: the Cell Painting Gallery, cpg0000-jump-pilot, CC0 1.0. Cite Chandrasekaran, S. N. et al. Three million images and morphological profiles of cells treated with matched chemical and genetic perturbations. Nat. Methods 21, 1114-1121 (2024).
  • JUMP/inputs/source_*: the Cell Painting Gallery, cpg0016-jump, CC0 1.0. JUMP/inputs/metadata and microscope_config.csv: jump-cellpainting/datasets. The benchmark follows Arevalo, J. et al. Evaluating batch correction methods for image-based cell profiling. Nat. Commun. 15, 6516 (2024).
  • NeurIPS22-CellSeg/NeurIPS22-CellSeg-cell_image_metrics.csv and CPJUMP1/manifest.json: SHAPE, CC BY 4.0.