---
license: cc-by-4.0
pretty_name: CellImageNet
task_categories:
  - image-classification
tags:
  - biology
  - single-cell
  - cell-type-classification
  - DAPI
  - nuclear-morphology
  - spatial-transcriptomics
  - xenium
size_categories:
  - 1M<n<10M
configs:
  - config_name: human
    data_files:
      - split: full
        path: data/human/*.tar
  - config_name: mouse
    data_files:
      - split: full
        path: data/mouse/*.tar
---

# CellImageNet

**CellImageNet** is a large-scale single-cell image database of **paired DAPI
nuclear images with cell-type annotations**, built from publicly available
10x Genomics Xenium data. It contains **~10 million cells** from **42 Xenium
samples — 28 human and 14 mouse tissues** — spanning diverse species, biological
conditions, and tissue types, annotated with **31 harmonized cell-type classes**
(unified from the source datasets' own annotations into a common label set).

Each cell has paired DAPI crops centered on the same cell at complementary context scales:

- **2.5×** — a tight view capturing fine nuclear morphology, and
- **10×** — a wider view capturing the local tissue context / neighbourhood.

Crops are provided at their **native resolution** (variable per sample; they are
*not* pre-resized — resize to a fixed input size, e.g. 224×224, is left to the
downstream model).

## Configurations & splits

| config | content |
|---|---|
| `human` | 28 human Xenium samples (~6.5M cells) |
| `mouse` | 14 mouse Xenium samples (~3.4M cells) |

(Counts are pre-filtering segmentation totals; the released set is marginally
smaller after removing cells with tiny nuclear masks or missing crops.)

This is an unsplit corpus: each config exposes a single `full` split (we do not
ship an official train/test partition). The exact subset used to train MorphPT
is specified in the [MorphPT weights repo](https://huggingface.co/jilab/MorphPT)
under `splits/`.

```python
from datasets import load_dataset
ds = load_dataset("jilab/CellImageNet", "human", split="full", streaming=True)
ex = next(iter(ds))
ex["2p5x.png"], ex["10x.png"], ex["json"]["cell_type"]
```

## Sample schema (WebDataset)

Each sample (one cell) is keyed by `cell_id` with three members:

| member | type | description |
|---|---|---|
| `2p5x.png` | image | 2.5× DAPI crop (grayscale, native resolution) |
| `10x.png` | image | 10× DAPI crop (grayscale, native resolution) |
| `json` | dict | metadata (below) |

`json` fields: `cell_id`, `dataset` (source Xenium sample), `species`
(Human/Mouse), `tissue`, `condition`, `cell_type` (one of the 31 classes below,
plus a small `Unknown` bucket in some mouse samples), `x_centroid`, `y_centroid`
(nuclear centroid, **microns**).

> Note: the field is named `cell_type` (the fine-grained cell label). It is
> *not* the coarse morphology "group" used by the MorphPT router — that grouping
> lives in the model repo, not in this dataset.

## Cell-type classes

The 31 harmonized cell-type labels in `cell_type`:

<details>
<summary>All 31 classes</summary>

Astrocytes · B cells · Brain cancer cells · Cardiac muscle cells · Chondrocytes ·
Colon cancer cells · Endothelial cells · Ependymal cells · Epithelial cells ·
Erythrocytes · Fibroblasts · Kidney cancer cells · Liver cancer cells ·
Lung cancer cells · Mesangial cells · Microglia · Myeloid cells · NK cells ·
Neurons · OPCs · Oligodendrocytes · Ovary cancer cells · Pancreas cancer cells ·
Pericytes · Schwann cells · Skeletal muscle cells · Skin cancer cells ·
Smooth muscle cells · Stem and progenitor cells · Stromal cells · T cells

</details>

`Unknown` is **mouse-only** (~134k cells, ≈3.8% of the mouse split; no human cell
carries it) and marks cells left unannotated in the source. Filter it out if you
need a clean 31-class label space.

## How it was built

Source: 42 Xenium samples (28 human, 14 mouse) from the
[10x Genomics datasets portal](https://www.10xgenomics.com/datasets). From each
tissue-wide DAPI image we used the `morphology_mip` maximum-intensity-projection
channel (or `morphology_focus` when unavailable). Nuclear segmentation masks
(10x Xenium Onboard Analysis) were converted to pixels at 0.2125 µm/px; cells
with rasterized nuclear area < 5 px or a bounding box < 10 px in either dimension
were removed. For each cell, two square crops centred on the nuclear centroid
were extracted at context scales r = 2.5 and r = 10 (side length S_r = r·d, with
d the per-sample mean nuclear bounding-box size) and zero-padded at image
borders. Crops are stored at native resolution.

## License & attribution

CellImageNet is a **derivative work** of publicly available 10x Genomics Xenium
datasets. The underlying imaging data is distributed by 10x Genomics under the
**Creative Commons Attribution 4.0 International ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/))**
license. Because CellImageNet is derived from CC BY 4.0 material, the image crops
are released under **CC BY 4.0**; the cell-type annotations and derived metadata
contributed by the CellImageNet authors are likewise released under CC BY 4.0.
See [`LICENSE`](LICENSE) for the full statement.

Under CC BY 4.0 you may share and adapt this dataset, including commercially,
provided you (1) credit 10x Genomics and the CellImageNet authors, (2) link the
license, and (3) **indicate that changes were made** — the images here have been
cropped/re-framed and re-annotated and are **not** the original 10x Genomics
files.

### Source datasets

All 42 source samples are 10x Genomics Xenium In Situ datasets from the
[10x Genomics datasets portal](https://www.10xgenomics.com/datasets). Each is
individually licensed CC BY 4.0 on its dataset page. The complete list of source
samples (dataset name, species, tissue, condition, and its 10x dataset URL) is
provided in **[`attribution_manifest.csv`](attribution_manifest.csv)** in this
repository.

Please cite both 10x Genomics and the individual source datasets in addition to
the CellImageNet/MorphPT paper below.

## Limitations

- **DAPI only** — nuclear morphology, no gene expression or protein channels
  (despite deriving from Xenium spatial-transcriptomics runs).
- **Native-resolution crops** vary in pixel size across samples; downstream
  models must resize to a fixed input.
- **Unsplit and imbalanced** — no official train/test split, and class frequency
  is highly skewed (tissue/condition sampling reflects the source datasets, not a
  balanced design). Subsample or reweight for classifier training.
- Labels are the source annotations harmonized into 31 classes; ≈3.8% of mouse
  cells (none in human) are `Unknown`.

## Relation to MorphPT

CellImageNet is the training corpus for **MorphPT**, a visual foundation model
for cell classification. MorphPT was trained on a human-only, per-class
subsampled subset of CellImageNet.

- Code: <https://github.com/AnitaCao/MorphPT>
- Model weights: <https://huggingface.co/jilab/MorphPT>


## Citation

```bibtex
@article{cao2026visual,
  title   = {A visual foundation model for cell classification},
  author  = {Cao, Ting and Zhuang, Haotian and Zhang, Boxuan and
             Pang, Zhiping P. and Tang, Ruixiang and Liu, Dongfang and
             Ji, Zhicheng},
  year    = {2026}
}
```