Xiang-zx-zx/xenium-senescence-demo

Xenium Senescence Benchmark (Demo Preview)

Note: This is a demo preview with 234 sample images. The full dataset (62K labeled cells across 2 tissue samples and 3 magnification scales) will be released upon paper publication.

Overview

A benchmark dataset for predicting cellular senescence from spatial transcriptomics (Xenium) cell images. Cell images are DAPI fluorescence microscopy captures at multiple magnification scales, labeled with senescence scores derived from gene expression (DeepScence).

Key challenge: Can cell morphology (nuclear shape, size, texture) predict senescence state — a biological process typically only measurable through gene expression?

Data Description

Tissue Samples

Sample	Tissue	Condition	Total Cells	Labeled Cells
Colon	Human colon	Non-diseased	270,984	21,510
Ovarian	Human ovary	Cancer (FFPE)	414,693	41,469

Multi-Scale Cell Images

Each cell is captured at 3 magnification scales from the DAPI fluorescence channel:

• 2.5x (56×56 pixels → 224×224): Nucleus only
• 5.0x (112×112 pixels → 224×224): Nucleus + immediate neighborhood
• 10.0x (224×224 pixels): Nucleus + extended tissue context

All images are grayscale (single-channel DAPI fluorescence), stored as RGB for compatibility.

Senescence Labels

Labels are derived from DeepScence gene expression signatures:

• Senescent (top 1% DeepScence score): Cells with highest senescence gene expression
• Normal (bottom 10% DeepScence score): Cells with lowest senescence gene expression
• ds_score: Continuous DeepScence score for regression tasks

Cell Types

Colon (5 types): B cells, Epithelial cells, Myeloid cells, Stem and progenitor cells, T cells

Ovarian (8 types): Endothelial cells, Epithelial cells, Fibroblasts, Myeloid cells, Ovary cancer cells, Pericytes, Smooth muscle cells, T cells

Benchmark Tasks

Task 1: Senescence Score Regression

• Input: Cell image (224×224, grayscale)
• Output: Continuous senescence score (ds_score)
• Metrics: Pearson R, Spearman ρ, MAE
• Split: 90% train / 10% test (fixed seed)

Task 2: Binary Senescence Classification

• Input: Cell image (224×224, grayscale)
• Output: Senescent vs Normal
• Metrics: AUC (primary), Balanced Accuracy, F1, AUPRC
• Split: Train = remaining cells; Test = 1:1 balanced (senescent:normal)

Task 3: Multi-Scale Fusion

• Input: Multiple scale images of the same cell (e.g., 2.5x + 5.0x)
• Output: Senescence score or label
• Challenge: How to best combine nucleus-level and tissue-context information?

Baseline Results

Colon — Binary Classification (Mean AUC across cell types)

Method	Scale	Init	Mean AUC
ResNet-18	5.0x	ImageNet	0.852
ResNet-50	5.0x	ImageNet	0.847
ViT-Small	5.0x	ImageNet	0.892
DINOv2-Small	5.0x	ImageNet	0.738
ViT-Small + MAE	5.0x	Domain MAE	0.892
ViT-Large + MAE	5.0x	Domain MAE	0.905
Multi-scale 2.5x+5.0x + MAE	2.5+5.0x	Domain MAE	0.879
Multi-scale Sep 2.5x+5.0x + MAE	2.5+5.0x	Domain MAE	0.875
Multi-scale 5.0x+10.0x + MAE (Ours)	5.0+10.0x	Domain MAE	0.906

Colon — Regression (Mean across cell types)

Method	Avg Pearson R	Avg Spearman ρ	Avg MAE
ResNet-18 (ImageNet)	0.259	0.215	0.098
ResNet-50 (ImageNet)	0.288	0.226	0.097
ViT-Small (ImageNet)	0.319	0.255	0.097
DINOv2-Small	0.333	0.227	0.093
MAE-ViT (domain pretrained)	0.340	0.258	0.095
Multi-scale 2.5x+5.0x + MAE	0.487	0.352	0.096
Multi-scale 5.0x+10.0x + MAE (Ours)	0.524	0.381	0.097

Ovarian — Binary Classification (Mean AUC across cell types)

Method	Scale	Init	Mean AUC
ViT-Small	10.0x	ImageNet	0.729
ViT-Small	5.0x	ImageNet	0.705
ResNet-18	5.0x	ImageNet	0.684
ViT-Large + MAE	5.0x	Domain MAE	0.679
Multi-scale 5.0x+10.0x + MAE	5.0+10.0x	Domain MAE	0.625

Note: Ovarian cancer tissue shows consistently weak morphology-senescence correlation across all methods (AUC < 0.73), suggesting a genuine biological limitation in this tissue type.

Key Findings

1. Multi-scale + domain-specific MAE pretraining achieves the highest AUC (0.906) on colon tissue
2. ViT-Large + domain MAE (0.905 AUC) outperforms ViT-Large + ImageNet (0.848 AUC), confirming the value of domain-specific pretraining
3. Colon tissue shows strong morphology-senescence signal (AUC up to 1.0 for Stem cells)
4. Ovarian cancer shows weak signal (AUC < 0.73) across all methods — genuine biological limitation
5. Larger model capacity (ViT-Large) benefits most when combined with domain-specific MAE pretraining

Demo Dataset Contents

This preview contains 234 sample images:

• Colon: 90 images | Ovarian: 144 images
• Senescent: 117 | Normal: 117
• Scales: 2.5x, 5.0x, 10.0x
• All cell types represented

images/          # PNG cell images (224×224)
metadata.csv     # cell_id, celltype, ds_score, label, sample, scale

Usage

from datasets import load_dataset

ds = load_dataset("Xiang-zx-zx/xenium-senescence-demo")

# Access an example
example = ds["train"][0]
print(example["image"])       # PIL Image (224x224)
print(example["celltype"])    # e.g. "Epithelial cells"
print(example["label"])       # 0=normal, 1=senescent
print(example["ds_score"])    # continuous DeepScence score
print(example["scale"])       # e.g. "5.0x"

Citation

Paper in preparation for NeurIPS 2026 Datasets & Benchmarks Track.

License

CC-BY-4.0. Original Xenium data from 10x Genomics public datasets.