Foundation models for digital pathology: most are not rotation-robust unless trained with rotation

augmentation ## main finding

Researchers evaluated twelve foundation models that encode patches from whole slide H&E images into latent representations. They found that rotation robustness is not automatic. Only models trained with self-supervised learning that included rotation augmentation showed significant invariance: the representation of a patch remained similar after rotations of 90, 180 or 270 degrees. The remaining models produced markedly different vectors for the same content when rotated, which could impair downstream tasks such as classification or similarity search.

The team measured robustness with two metrics. First, they computed cosine distance between the representations of an original patch and its rotated version. Second, they used mutual k-nearest neighbours to assess whether the two representations pointed to the same neighbors in latent space. Tests were run on identical patches rotated at various angles, without providing the angle as input to the model. Results showed a clear gap: models that received rotation augmentation achieved substantially higher invariance scores, while models without it dispersed.

The researchers attribute the failure of Transformer architectures to a lack of an inductive bias for rotation, unlike CNNs that employ weight-shared convolutions and embed partial rotation invariance. Without seeing rotated examples during training, a Transformer does not learn that rotation does not alter the semantic meaning of tissue. Rotation augmentation supplies the missing examples and forces the model to learn direction-independent representations.

Practitioners deploying foundation models in digital pathology should verify that the model was trained with rotation augmentation, or add such augmentation during fine-tuning, especially when scans may be captured at different angles or patches are cut at arbitrary orientations. The research code is publicly available, allowing replication of the test on additional models before selecting one for production use.