Add dataset card

README.md

@@ -1,36 +1,19 @@
 ---
-dataset_info:
-  features:
-  - name: image_id
-    dtype: string
-  - name: image
-    dtype: image
-  - name: mask
-    dtype: image
-  - name: split
-    dtype: string
-  - name: num_nuclei
-    dtype: int32
-  - name: height
-    dtype: int32
-  - name: width
-    dtype: int32
-  - name: usage
-    dtype: string
-  - name: is_grayscale
-    dtype: bool
-  splits:
-  - name: train
-    num_bytes: 64503513
-    num_examples: 670
-  - name: stage1_test
-    num_bytes: 9680318
-    num_examples: 65
-  - name: stage2_test
-    num_bytes: 13367920
-    num_examples: 106
-  download_size: 87548608
-  dataset_size: 87551751
+license: cc0-1.0
+pretty_name: 2018 Data Science Bowl (BBBC038) - Nuclei Segmentation
+task_categories:
+- image-segmentation
+tags:
+- medical
+- biomedical
+- microscopy
+- histopathology
+- fluorescence
+- nuclei
+- cell-segmentation
+- bbbc038
+size_categories:
+- n<1K
 configs:
 - config_name: default
   data_files:
@@ -41,3 +24,73 @@ configs:
   - split: stage2_test
     path: data/stage2_test-*
 ---
+
+# 2018 Data Science Bowl (BBBC038) - Nuclei Segmentation
+
+2D light-microscopy **cell-nucleus segmentation** assembled across many imaging
+experiments (humans, mice, flies; 22 cell types, 15 resolutions, 30+ experiments).
+The collection deliberately spans **multiple modalities**: fluorescence
+(DAPI / Hoechst), brightfield **H&E histopathology**, and other brightfield -
+making it a standard cross-modality nuclei-segmentation benchmark.
+
+This is the official **BBBC038v1** release (Broad Bioimage Benchmark Collection),
+the same data used in the Kaggle 2018 Data Science Bowl. **License: CC0 / public
+domain.**
+
+## Contents & splits
+
+| Split | Images | Nuclei | Ground-truth source |
+|-------|-------:|-------:|---------------------|
+| `train`        (stage1_train) | 670 | 29,461 | native per-nucleus PNG instance masks |
+| `stage1_test`  (stage1_test)  |  65 |  4,152 | RLE in `stage1_solution.csv` (post-competition) |
+| `stage2_test`  (stage2_test)  | 106 |  3,716 | RLE in `stage2_solution_final.csv` (post-competition) |
+| **Total** | **841** | **37,329** | |
+
+**Faithful-naming notes**
+- Most papers cite "DSB2018" = `stage1_train` (670) only, since that is the only
+  split distributing *native* instance masks. This repo ships the **full 3-stage**
+  set; the test-stage GT was decoded from the official solution-CSV RLE.
+- The raw `stage2_test_final` archive contains ~3,019 images, but only **106 are
+  scored** - the rest are intentional **decoys** flagged `Usage=Ignored`. **Only
+  the 106 scored images are included here.**
+
+## Ground truth
+
+`mask` is a **binary semantic** nucleus mask (mode `L`, values `{0, 255}`): the
+**union of all per-nucleus instances**. For `train` it is the union of the native
+per-nucleus PNG masks; for the test splits it is the union of the RLE-decoded
+nuclei. The RLE decoder was validated against the native train masks
+(pixel agreement = 1.000000). The original per-nucleus **instance** masks remain
+available at [BBBC038](https://bbbc.broadinstitute.org/BBBC038) for
+instance-segmentation use.
+
+## Columns
+
+| Column | Type | Notes |
+|--------|------|-------|
+| `image_id`     | string | source hash id |
+| `image`        | Image  | RGB (RGBA fluorescence normalized to RGB) |
+| `mask`         | Image  | binary semantic, `{0,255}` |
+| `split`        | string | `stage1_train` / `stage1_test` / `stage2_test` |
+| `num_nuclei`   | int32  | nuclei in this image |
+| `height`,`width` | int32 | image dimensions |
+| `usage`        | string | `null` (train) / `Public` (s1) / `Private` (s2) |
+| `is_grayscale` | bool   | derived (R==G==B): fluorescence/brightfield vs H&E color |
+
+`metadata.xlsx` (repo root) is the official 43-row per-experiment provenance table
+(cell type, stain, SNR, resolution).
+
+## Provenance, overlap & integrity
+
+- **Provenance:** official BBBC038v1 (Broad Institute), CC0. Counts reconcile with
+  the paper (670 / 65 / 106).
+- **Overlap (leakage hazards):** a small fraction of images overlap **BBBC039**.
+  The H&E subset shares source-level (TCGA-derived) lineage with H&E nuclei sets
+  such as MoNuSeg / PanNuke, though no individually-confirmed shared images.
+- **Curated collection:** assembled from 30+ independent experiments / donor labs.
+
+## Citation
+
+Caicedo, J.C., Goodman, A., Karhohs, K.W., et al. *Nucleus segmentation across
+imaging experiments: the 2018 Data Science Bowl.* **Nature Methods** 16(12),
+1247-1253 (2019). doi:10.1038/s41592-019-0612-7