Computational Pathology · owkin
plism-dataset-tiles
Histology tile dataset across 7 scanners and 13 staining conditions for benchmarking foundation model robustness.
What plism-dataset-tiles is
plism-dataset-tiles is a computational pathology image dataset curated by Owkin to evaluate how well artificial intelligence models tolerate inter-institutional domain shifts. It builds upon the Pathology Images of Scanners and Mobilephones (PLISM) dataset initially gathered by Ochi et al., re-processing whole slide images to supply standardised image tiles for robustness benchmarks.
The underlying biological material consists of consecutive tissue sections derived from patients who underwent surgery at the University of Tokyo Hospital between 1955 and 2018. Across 91 whole slide images, the collection covers 46 tissue microarrays from diverse organs, imaged under seven separate digital slide scanners and 13 distinct haematoxylin and eosin (H&E) staining conditions.
Because each condition captures the same underlying biological structures across consecutive sections, the dataset allows researchers to evaluate whether feature extraction and classification models produce consistent outputs despite technical variations introduced by slide preparation and digitisation hardware.
How it was made
Unregistered whole slide images were aligned using the Elastix medical image registration software. A slide stained under the GMH condition and digitised with a Hamamatsu Nanozoomer S60 scanner served as the spatial reference.
From the aligned slides, an in-house bidirectional U-Net was applied to extract tissue patches. Each tile measures 224 × 224 pixels at a resolution of 0.5 µm per pixel, representing 20× magnification. This extraction yielded 16,278 tiles for each of the 91 slide conditions, producing a total of 1,481,298 registered histology tiles. Every tile is catalogued with its slide identifier, tile identifier, scanner model, and stainer condition.
Key facts
- Total tiles
- 1,481,298
- Tile dimensions
- 224 × 224 pixels
- Resolution
- 0.5 µm/pixel (20× magnification)
- Total data size
- Approximately 150 GB
- Experimental conditions
- 91 whole slide images (7 scanners × 13 H&E staining conditions)
- Biological coverage
- 46 tissue microarrays (TMAs) across various organs
- Registration software
- Elastix
- Tile extraction method
- Bidirectional U-Net
- Data formats
- Parquet, WebDataset
What you can use it for
- Benchmarking foundation model robustness
- Evaluating the stability of pathology representations across hardware and staining variations using identical biological tissue areas.
- Histopathology feature extraction
- Generating and comparing vector embeddings from standardised 224 × 224 histology patches across multiple digitisation pipelines.
- Domain shift assessment in classification
- Testing whether image classification models trained on specific stain or scanner profiles generalise when exposed to alternative scanning devices and stain preparations.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from owkin/plism-dataset-tiles at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 55,616 rows of default/train
__key__
string_text- min
- 32
- median
- 36
- mean
- 36.25
- max
- 39
png
image- min
- 224
- median
- 224
- mean
- 224
- max
- 224
scanner
string_label- AT229%
- P29%
- GT45029%
slide_id
string_label- GIVH_AT2_to_GMH_S6029%
- GIVH_GT450_to_GMH_S6029%
- GIVH_P_to_GMH_S6029%
stainer
string_label- GIVH100%
tile_id
string_text- min
- 12
- median
- 14
- mean
- 14.13
- max
- 15
Strengths and limits
Good at
- Provides 1,481,298 spatially aligned tiles that maintain identical biological structures across 91 scanner and stainer permutations.
- Supplies tiles formatted at 224 × 224 pixels and 20× magnification, matching the standard input parameters of many vision foundation models.
- Includes granular metadata per tile, explicitly recording slide identifiers, tile coordinates, scanner devices, and stainer protocols.
Watch out for
- All tissue specimens originate from a single centre, the University of Tokyo Hospital.
- Staining variability is limited to haematoxylin and eosin (H&E) protocols and does not encompass immunohistochemical or special stains.
- Tile extraction and alignment depend on algorithmic registration against a single reference slide, making downstream evaluation sensitive to registration accuracy.
Questions
Can I look inside plism-dataset-tiles without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is plism-dataset-tiles under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
How does this version differ from the original PLISM dataset?
The original dataset used AKAZE keypoint matching to extract 310,947 patches of 512 × 512 pixels at 40× magnification (0.22 to 0.26 µm/pixel). This dataset re-registers the slides using Elastix and extracts all 16,278 tiles per slide at 224 × 224 pixels and 20× magnification (0.5 µm/pixel), resulting in 1,481,298 tiles.
What compute is required to run feature extraction across the dataset?
Extracting features from all tiles using a Vision Transformer (ViT-B) model takes approximately two and a half hours and roughly 10 GB of storage using 16 CPUs and a single 16 GB Nvidia T4 GPU.
Which reference slide was chosen for image registration?
The reference slide used during the Elastix registration procedure was stained under the GMH condition and scanned using a Hamamatsu Nanozoomer S60.
What metadata is recorded for individual image tiles?
Each tile entry contains the original slide identifier (slide_id), the tile identifier (tile_id), the specific stainer condition, and the scanner model.
How to cite it
Filiot, A., Dop, N., Tchita, O., Riou, A., Peeters, T., Valter, D., Scalbert, M., Saillard, C., Robin, G., & Olivier, A. (2025). Distilling foundation models for robust and efficient models in digital pathology. arXiv. https://arxiv.org/abs/2501.16239
Written from plism-dataset-tiles's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.