Datasets:
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,74 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: cc-by-4.0
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
task_categories:
|
| 4 |
+
- image-classification
|
| 5 |
+
tags:
|
| 6 |
+
- pathology
|
| 7 |
+
- histopathology
|
| 8 |
+
- prostate-cancer
|
| 9 |
+
- gleason-grading
|
| 10 |
+
- foundation-models
|
| 11 |
+
- benchmark
|
| 12 |
+
size_categories:
|
| 13 |
+
- 1K<n<10K
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
# PANDA-PLUS-Bench
|
| 17 |
+
|
| 18 |
+
A benchmark dataset for evaluating WSI-specific feature collapse in pathology foundation models.
|
| 19 |
+
|
| 20 |
+
## Dataset Description
|
| 21 |
+
|
| 22 |
+
PANDA-PLUS-Bench contains expert-annotated prostate biopsy patches from 9 whole slide images (9 unique patients) with pixel-level Gleason pattern annotations.
|
| 23 |
+
|
| 24 |
+
### Dataset Summary
|
| 25 |
+
|
| 26 |
+
- **Patches**: ~2,770 per augmentation condition
|
| 27 |
+
- **Resolution**: 224×224 pixels at 20× magnification
|
| 28 |
+
- **Classes**: Benign (0), GP3 (1), GP4 (2), GP5 (3)
|
| 29 |
+
- **Slides**: 9 (one per patient)
|
| 30 |
+
- **Augmentations**: 8 conditions
|
| 31 |
+
|
| 32 |
+
### Augmentation Conditions
|
| 33 |
+
|
| 34 |
+
| Split | Description |
|
| 35 |
+
|-------|-------------|
|
| 36 |
+
| baseline | ImageNet normalization only |
|
| 37 |
+
| color_jitter | Brightness, contrast, saturation, hue |
|
| 38 |
+
| grayscale | Complete color removal |
|
| 39 |
+
| gaussian_noise | Additive noise (σ=0.05) |
|
| 40 |
+
| heavy_geometric | Rotation ±180°, flips |
|
| 41 |
+
| combined_aggressive | All augmentations combined |
|
| 42 |
+
| macenko_normalization | Stain normalization |
|
| 43 |
+
| hed_stain_augmentation | H/E channel perturbation |
|
| 44 |
+
|
| 45 |
+
## Usage
|
| 46 |
+
```python
|
| 47 |
+
from datasets import load_dataset
|
| 48 |
+
|
| 49 |
+
# Load baseline patches
|
| 50 |
+
dataset = load_dataset("dellacortelab/PANDA-PLUS-Bench", split="baseline")
|
| 51 |
+
|
| 52 |
+
# Access a sample
|
| 53 |
+
sample = dataset[0]
|
| 54 |
+
image = sample['image'] # PIL Image
|
| 55 |
+
label = sample['label'] # 0-3
|
| 56 |
+
slide_id = sample['slide_id'] # Slide identifier
|
| 57 |
+
```
|
| 58 |
+
|
| 59 |
+
## Evaluation
|
| 60 |
+
|
| 61 |
+
See our [Colab notebook](https://colab.research.google.com/github/dellacortelab/PANDA-PLUS-Bench/blob/main/PANDA_PLUS_Bench_Evaluation.ipynb) for standardized evaluation.
|
| 62 |
+
|
| 63 |
+
## Citation
|
| 64 |
+
```bibtex
|
| 65 |
+
@article{ebbert2025pandaplusbench,
|
| 66 |
+
title={PANDA-PLUS-Bench: A Benchmark for Evaluating WSI-Specific Feature Collapse},
|
| 67 |
+
author={Ebbert, Joshua and Della Corte, Dennis},
|
| 68 |
+
year={2025}
|
| 69 |
+
}
|
| 70 |
+
```
|
| 71 |
+
|
| 72 |
+
## License
|
| 73 |
+
|
| 74 |
+
CC-BY-4.0
|