anon-neuripsdb26/SenoMorph

---
license: cc-by-4.0
task_categories:
- image-classification
- image-to-image
tags:
- cell-morphology
- senescence
- DAPI
- flow-matching
- benchmark
size_categories:
- 100K<n<1M
configs:
- config_name: real
default: true
data_files:
- split: train
path: real-*-train.parquet
- split: validation
path: real-*-val.parquet
- split: test
path: real-*-test.parquet
- config_name: real_lung
data_files:
- split: train
path: real-lung-train.parquet
- split: validation
path: real-lung-val.parquet
- split: test
path: real-lung-test.parquet
- config_name: real_prostate
data_files:
- split: train
path: real-prostate-train.parquet
- split: validation
path: real-prostate-val.parquet
- split: test
path: real-prostate-test.parquet
- config_name: real_skin
data_files:
- split: train
path: real-skin-train.parquet
- split: validation
path: real-skin-val.parquet
- split: test
path: real-skin-test.parquet
- config_name: companion
data_files:
- split: train
path: companion-*-train.parquet
- split: validation
path: companion-*-val.parquet
- split: test
path: companion-*-test.parquet
- config_name: companion_lung
data_files:
- split: train
path: companion-lung-train.parquet
- split: validation
path: companion-lung-val.parquet
- split: test
path: companion-lung-test.parquet
- config_name: companion_prostate
data_files:
- split: train
path: companion-prostate-train.parquet
- split: validation
path: companion-prostate-val.parquet
- split: test
path: companion-prostate-test.parquet
- config_name: companion_skin
data_files:
- split: train
path: companion-skin-train.parquet
- split: validation
path: companion-skin-val.parquet
- split: test
path: companion-skin-test.parquet
---

# SenoMorph

**SenoMorph** is a senescence-conditioned cell morphology benchmark of
single-cell DAPI crops paired with continuous, transcriptome-derived
senescence scores. The dataset accompanies an anonymous NeurIPS 2026
Datasets & Benchmarks submission.

## Composition

| Subset | Cells | Score-conditioned crops |
|:------------------|--------:|------------------------:|
| Real (lung) | 41,702 | — |
| Real (prostate) | 28,845 | — |
| Real (skin) | 16,844 | — |
| **Real total** | **87,391** | — |
| Companion (lung) | 41,702 | 417,020 (×10) |
| Companion (prost.)| 28,845 | 288,450 (×10) |
| Companion (skin) | 16,844 | 168,440 (×10) |
| **Companion total** | **87,391** | **873,910** |

Each instance carries five attributes (paper §2 attribute card):
`tissue`, `cell_type_name`, `senescence_score`, `nuclear_area` (mask
pixel count, threshold > 10), and `dapi_intensity` (mean intensity
within the mask). The image is a `64×64` grayscale uint8 DAPI crop.

## Splits

70 / 10 / 20 train / val / test, stratified by cell type within each
tissue. The companion set inherits its split from the source real cell.

## Real subset

Each real cell has a senescence score produced by `DeepScence` from its
matched single-cell transcriptome and min-max normalized to `[0,1]`
within its tissue. Cell-type labels are produced by `CellTypist`.

## Companion subset

For each real source cell, the model generates ten score-conditioned
images at target scores evenly spaced over `[0,1]`. The companion subset
is intended for downstream augmentation tasks (e.g., senescent-cell
classification) and is **not** used as ground truth for any evaluation
metric in the accompanying paper.

## Loading

Available configs:

| `name=` | Description |
|---|---|
| `real` | All three tissues, real DAPI cells |
| `companion` | All three tissues, generated companion cells |
| `real_lung` / `real_prostate` / `real_skin` | Real cells, single tissue |
| `companion_lung` / `companion_prostate` / `companion_skin` | Companion cells, single tissue |

Each config has `train`, `validation`, and `test` splits.

```python
from datasets import load_dataset

# Full real benchmark, training split (3 tissues combined)
real_train = load_dataset('anon-neuripsdb26/SenoMorph', name='real', split='train')

# Real lung test split only
lung_test = load_dataset('anon-neuripsdb26/SenoMorph', name='real_lung', split='test')

# Companion data (873,910 generated cells), train split
comp_train = load_dataset('anon-neuripsdb26/SenoMorph', name='companion', split='train')

# Companion skin only, validation split
skin_val_comp = load_dataset('anon-neuripsdb26/SenoMorph', name='companion_skin', split='validation')
```

## Schema

```
image PIL.Image (grayscale 64×64, PNG-encoded)
tissue str ('lung' | 'prostate' | 'skin')
cell_type_id int32 (raw CellTypist integer)
cell_type_name str ('Epithelial cells' | 'T cells' | 'Fibroblasts' | ...)
senescence_score float32 in [0, 1]
real: DeepScence min-max norm within tissue
companion: target conditioning score from linspace(0,1,10)
nuclear_area int32 (mask pixel count, threshold > 10)
dapi_intensity float32 (mean intensity within mask)
split str ('train' | 'val' | 'test')
cell_id int64
real: row index in source set
companion: source real cell id (10 companion rows share a cell_id)
```

## License

Released under CC-BY-4.0. Raw Xenium outputs are publicly available from
the 10x Genomics dataset portal. The dataset contains no personally
identifiable information; `64×64` DAPI nuclear crops cannot be traced
back to individual donors.

## Anonymous

This release accompanies an anonymous NeurIPS 2026 Datasets & Benchmarks
submission. Author identification is intentionally suppressed during
review.