Vision Transformer adapted for neuropathology, pretrained using self-supervised learning (DINO + iBOT objectives) on real-world neuropathology data. This initial test model was developed as part of the Federated Brain Digital Slide Archive project.
The model supports downstream tasks including transfer learning, tissue segmentation, patch-level classification, and similarity search for large image repositories.
| Metric | Value |
|---|---|
| Linear Probe Accuracy | 80.17% |
| KNN Accuracy | 83.76% |
Outperforms UNI, UNI2-h, prov-gigapath, and base dinov2-giant on the same dataset.
The model is publicly available on Hugging Face.
Trained on NVIDIA DGX H100 cluster (32 GPUs). Part of the Brain Digital Slide Archive project funded by NIH NINDS (U24NS133945-01).
Hosted on Hugging Face.