With stimulating presentations, discussions, networking and insights into the latest developments in the field of AI in healthcare, KI-MED Connect brings together experts from companies, universities and hospitals to discuss various topics relating to AI in healthcare. Topics include data acquisition, compliance/standards and guidelines, ethical aspects, explainable AI, anonymization, validation and verification as well as transparency and high-performance computing.

Fraunhofer MEVIS will be represented with the following contribution:

Talk

Johannes Lotz

"Überlebenszeitvorhersage beim Prostatakarzinom durch föderierte Foundation-Modelle" (Survival time prediction in prostate cancer using federated foundation models)

Session: KI-basierte Prognose-modelle (AI-based forecast models)

Thursday, September 19, 9 am – 10:30 am

With the increasing workload of pathologists, the need for automation to support diagnostic tasks and the quantitative assessment of biomarkers is becoming more apparent. Foundation models have the potential to improve generalizability both within a center and between different centers, and can play a crucial role in the efficient development of specialized AI models. However, training these models often requires an enormous amount of data, computational power and time. We present a supervised multitask learning approach that significantly reduces these requirements, which we used to train the Tissue Concepts Foundation model.

In the ProSurvival project, we are extending the application of the Foundation model to predict the survival of prostate cancer patients based on histopathological image data. Through federated learning, we enable the collaborative use of patient data from multiple clinical sites without the data leaving the clinic. This infrastructure is designed to reduce privacy concerns and train a global model from local data. The Tissue Concepts Foundation model was tested on whole-slide images of breast, colorectal, lung and prostate cancer and achieved comparable performance to other models with only 6% of the training data set. Project website: prosurvival.org.