Most recent histopathology foundation models have been pushed forward by training on ever-larger numbers of image patches. This study takes the opposite approach and asks what happens when the priority is the variety of the slides rather than the quantity of patches cut from them.
What we did
Athena was initialised from a pretrained model and trained on 115 million tissue patches drawn from 282,000 whole-slide images, several times fewer patches than recent histopathology foundation models use. Instead of elaborate sampling heuristics, we simply took a moderate number of patches at random from each slide in an internal repository spanning 52 contributing centres across 25 countries, together with slides from GTEx. That repository varies widely in staining, scanning protocol, organ and tissue preparation, which is precisely the point.
What we found
Evaluated on one patch-level benchmark and four slide-level downstream tasks (two molecular, two morphological), Athena approaches the state of the art and surpasses several models trained on substantially larger datasets.
The practical implication is that a diverse slide repository lets a model reach strong performance for a fraction of the compute, which lowers the bar for building pathology foundation models outside the largest labs.