Domain adaptation using optimal transport for invariant learning using histopathology datasets

Retraining Representation Domain Adaptation
DOI: 10.48550/arxiv.2303.02241 Publication Date: 2023-01-01
ABSTRACT
Histopathology is critical for the diagnosis of many diseases, including cancer. These protocols typically require pathologists to manually evaluate slides under a microscope, which time-consuming and subjective, leading interest in machine learning automate analysis. However, computational techniques are limited by batch effects, where technical factors like differences preparation protocol or scanners can alter appearance slides, causing models trained on one institution fail when generalizing others. Here, we propose domain adaptation method that improves generalization histopathological data from unseen institutions, without need labels retraining these new settings. Our approach introduces an optimal transport (OT) loss, extends adversarial methods penalize if images different institutions be distinguished their representation space. Unlike previous methods, operate single samples, our loss accounts distributional between batches images. We show Camelyon17 dataset, while both adapt global color distribution, only OT reliably classify cancer phenotype during training. Together, results suggest rare but phenotypes may make up small fraction total tiles variation slide.
SUPPLEMENTAL MATERIAL
Coming soon ....
REFERENCES ()
CITATIONS ()
EXTERNAL LINKS
PlumX Metrics
RECOMMENDATIONS
FAIR ASSESSMENT
Coming soon ....
JUPYTER LAB
Coming soon ....