AI in Histopathology: Prototypical Few-Shot Classifiers
Author Information
Author(s): Kuritcyn Petr, Kletzander Rosalie, Eisenberg Sophia, Wittenberg Thomas, Bruns Volker, Evert Katja, Keil Felix, Ziegler Paul K., Bankov Katrin, Wild Peter, Eckstein Markus, Hartmann Arndt, Geppert Carol I., Benz Michaela
Primary Institution: Fraunhofer IIS
Hypothesis
Can prototypical few-shot classification models and data augmentation effectively address the challenges of annotated images and data heterogeneity in histopathology?
Conclusion
Prototypical few-shot models can be effectively adapted to new tasks with minimal annotations, achieving high accuracy in histopathology classification.
Supporting Evidence
- Models achieved around 90% classification performance on a multiscanner and multicenter database.
- Only three annotations per subclass were needed to adapt the few-shot model for urothelial carcinoma classification.
- Data augmentation techniques were effective in improving model robustness against domain shifts.
Takeaway
This study shows that AI can help doctors classify tissue samples with just a few examples, making it easier to use in hospitals.
Methodology
The study used prototypical few-shot classification models trained on annotated datasets from multiple centers and scanners, applying domain-specific data augmentation.
Potential Biases
Potential bias due to the limited number of annotations per class and the specific types of tumors studied.
Limitations
The study's results may not generalize to all types of histopathological tasks or different staining methods.
Participant Demographics
Data included tissue sections from patients with colon adenocarcinoma and urothelial carcinoma.
Statistical Information
P-Value
0.001
Confidence Interval
95%
Statistical Significance
p<0.05
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website