Paper Recipes
wsistream can approximate the data pipelines of several pathology foundation models. This is a selected (not exhaustive) collection -- each recipe shows how to configure the pipeline to match a specific paper's preprocessing as closely as possible.
| Paper | Patch size | Resolution | Tissue detection | Tissue threshold | Training data |
|---|---|---|---|---|---|
| Midnight | 256x256 | 0.25, 0.5, 1.0, 2.0 mpp | U-Net (not open-sourced) | 40% | 384M tiles from 92K WSIs |
| UNI | 256x256 / 512x512 | 20x (0.5 mpp) | CLAM | not specified | ~100M tiles from 100K WSIs |
| Virchow | 392x392 | 5x/10x/20x/40x mixed | Trained FCN (not open-sourced) | 65% | 3.1M WSIs |
| GPFM | 512x512 | Native (level 0) | CLAM (dataset-specific presets) | contour-based | 190M tiles from 72K WSIs |
| Prov-GigaPath | 256x256 | 20x (0.5 mpp) | Otsu on luminance | 10% | 1.38B tiles from 171K WSIs |
| Phikon | 224x224 | 20x (0.5 mpp) | U-Net (not open-sourced) | 60% | 456M tiles from 58K WSIs |
