Site-Specific Digital Histology Signatures
Evidence that models can learn the submitting site instead of the biology.
The most important cautionary result in computational pathology. It explains why a model reporting excellent accuracy on TCGA may fail completely on your slides.
- Cost
- Free
- Hardware
- Runs on any laptop
- Scanner
- Not required
- Your data
- Stays on your machine
- Licence
- unknown
- Maturity
- production
What it is
Shows that images from different TCGA submitting sites are distinguishable by deep learning, that stain normalisation and augmentation do not remove the signal, and that this produces biased accuracy for survival, mutation and stage prediction.
Licence notes
Published document; check the publisher's terms for reuse.
Related
Catalogued 2026-08-02, last checked 2026-08-02. View the source record · Report an error