HistoEncoder
An open-source foundation model for prostate cancer digital pathology, trained without labels on 48 million tissue tiles.
I co-authored HistoEncoder, an open-source foundation model for prostate cancer digital pathology. We pre-trained it without labels on 48 million tissue tiles from 1,307 patients, using DINO self-supervised learning with an XCiT backbone.
What we found
- Efficient feature extraction: in our benchmarks, HistoEncoder achieved comparable Gleason pattern classification performance to larger pan-cancer foundation models while running 3–5× faster.
- Learning from limited annotations: fine-tuning on four slides reached performance comparable to a natural-image model fine-tuned on 37.
- Large-scale tissue annotation: clustering the learned features grouped similar tissue morphology, supporting pre-annotation of large datasets.
Histology clusters also carried prognostic information on prostate cancer-specific mortality of a similar magnitude to Gleason grading in internal cross-validation. Adding them to Gleason did not improve prognostic performance in that analysis.
The clinical proof-of-concept analyses ran on a single laptop. Code and model weights are openly available in the HistoEncoder repository.
The study is published in the Journal of Pathology Informatics (2026). This summary follows my publication announcement and the published abstract.
Publication announcement
Original poster from my LinkedIn post.