Project announcement 🏗️: "PROSurvival" builds a collaborative federated learning framework to predict survival in prostate cancer patients.
Partners are OFFIS - Institute for Information Technology, Charité, Goethe University Frankfurt, and Fraunhofer MEVIS.
Here's the gist: In the long run, we want to find predictive image features that can be identified in tissue sections from routine diagnostics.
#federatedlearning #ai #foundationmodels #ComputationalPathology
1/3
The problem: multi-centric data 🏥 is essential to train a robust model, but is rarely accessible. Science needs more public datasets! If data cannot be shared, we need collaborative tools.
Our solution: In the project, we build a cooperative federated learning platform that allows scientists at the different sites to train a joint model together.
2/3
A core building block is a foundation model ("TissueConcepts") that we train supervised specifically for prostate cancer for (hopefully) better generalization on external data. The TissueConcepts-model also is a feature extractor such that only compressed representations of the data are federated. Training becomes faster and requires less data exchange.
More at https://prosurvival.org or in the preprint on supervised foundation model training -> https://arxiv.org/abs/2311.09847.
3/3
Johannes Lotz