Pathology Foundation Model
A foundation model pre-trained in a self-supervised manner on large-scale whole-slide image datasets to learn general tissue morphology representations. These models can be applied to downstream pathological analysis tasks such as cancer grading and genomic mutation prediction.
Timeline (last 90 days)
When the same tissue is imaged on scanners of different models, pathology foundation models generate different feature representations (embeddings)
Existing research has shown that the source scanner can be predicted with high accuracy from embeddings alone, indicating that device fingerprints have been deeply encoded by the model
Data augmentation alone is insufficient to completely eliminate scanner shift, because augmentation can only cover known transformation spaces, while real device differences are more complex and difficult to enumerate in advance
All Facts (3)
When the same tissue is imaged on scanners of different models, pathology foundation models generate different feature representations (embeddings)
50%UnverifiedExisting research has shown that the source scanner can be predicted with high accuracy from embeddings alone, indicating that device fingerprints have been deeply encoded by the model
50%UnverifiedData augmentation alone is insufficient to completely eliminate scanner shift, because augmentation can only cover known transformation spaces, while real device differences are more complex and difficult to enumerate in advance
50%