---
language:
- en
license: other
license_name: mixed-source-terms
license_link: LICENSE.md
task_categories:
- image-segmentation
tags:
- microscopy
- instance-segmentation
- weakly-supervised
- autoresearch
size_categories:
- 1K<n<10K
---
# LuminBench: Weakly Supervised Cell Segmentation

A checksum-bound training and validation package for agent-driven research with two labelled images per group. The scientific protocol and runnable code are maintained in [LuminBench-Weakly-Supervised-CellSeg](https://github.com/Lumin-Science/LuminBench-Weakly-Supervised-CellSeg). The project uses the [LB-Template](https://github.com/Lumin-Science/LB-Template) organization for ordinary experiments and agent research.

| Split | Images | Masks | Role |
|---|---:|---:|---|
| training_labelled | 24 | 24 | Two human-labelled examples in each of 12 groups |
| training_unlabelled | 967 | 0 | Public image-only pseudo-label pool |
| validation | 111 | 111 | Fixed hill-climbing evaluation covering 11 public groups |
| test | 0 included | 0 included | The 65 test images and masks are deliberately not distributed here |

There are 1,102 images and 135 visible masks. Public images retain their original bytes. Exact image/mask paths and SHA-256 values are in `manifests/*.csv`; `dataset.json` binds those manifests to the frozen project configuration. Masks are integer instance IDs, with zero background; this is not a semantic-class mask dataset. Different official splits reuse filenames, so always resolve a record through its split manifest.

The 11 public groups are appearance clusters, not verified biological cell types. Group 11 is Jianglab microscopy. The two Jianglab training images are `20240413/103.tif` and `20240422_2/304.tif`; their masks are lossless instance-mask TIFFs extracted from the original Cellpose annotations. The other 13 Jianglab images are reserved for the owner's test. No Jianglab validation images are supplied, by the owner's explicit split instruction. The public collection includes multiple imaging modalities; do not describe every image as fluorescence microscopy.

The validation set consists of 101 original tuning images and five original training image/mask pairs from each previously absent group, 2 and 10. Promoted source IDs: `cell_00675`, `cell_00676`, `cell_00680`, `cell_00682`, `cell_00683`, `cell_00763`, `cell_00769`, `cell_00771`, `cell_00773`, `cell_00774`. All ten are removed from training and pseudo-label pools. The training duplicate `cell_00550` is excluded because it matches a tuning image. No hidden masks for the 967 unlabelled images are included.

## Intended protocol

Initialize from the official MEDIAR **public-only phase-1** checkpoint, SHA-256 `f68fb3b3a193a10018c2108003caf0fca1c66028d59a3db175ca7fb718ea8a9b`. MEDIAR phase-2 and final target-finetuned weights are forbidden in this task. Train with only the 24 visible human masks plus pseudo masks generated from the 967 unlabelled images. Validation masks are for scoring, not training. Optimize the mean of per-group image-mean object F1 at IoU 0.5 over the 11 represented validation groups, using the project's fixed boundary and matching rules. Freeze model selection before owner test evaluation. Prior research already exposed test results, so future reuse is not a fresh blind evaluation.

## Download and verification

Use the immutable commit published in the project's `docs/dataset_release.json`, rather than a moving `main`, and run `bash scripts/setup.sh --data-only`. The project verifier checks the exact frozen identities, every payload hash and the absence of unexpected payload files. This dataset uses ordinary files and CSV manifests to preserve original mixed image formats and integer instance masks; it is not an automatic imagefolder classification dataset.

## Sources and terms

Public images and masks: Jun Ma, Ronald Xie, Shamini Ayyadhury, Cheng Ge, and collaborators, [NeurIPS 2022 Cell Segmentation Competition Dataset](https://doi.org/10.5281/zenodo.10719375), under **CC BY-NC-ND 4.0**. Public source image and mask bytes are unchanged; this package supplies a selected collection and split metadata. Jianglab images were supplied by the project owner, who explicitly authorized public upload of these two training images and their annotations. No broader reuse license for Jianglab material was specified. See [LICENSE.md](LICENSE.md) for component-specific terms; the code repository's MIT license does not license these images.