TPI: Tumor Behavior Prediction Initiative

Funded by the BMG from 2020 to 2022
Project leader: Dr. Eva Krieghoff-Henning

One of the declared goals of oncology is to enable an individually tailored, "personalized" therapy for every cancer patient in the near future. This requires new biomarkers that allow a better assessment of the prognosis and the expected response to therapy in each individual case. For the development of such biomarkers, medical data can be used that is already available in Germany due to the documentation and archiving obligations.

A particularly valuable resource in this context are archived slides with tissue sections that are routinely prepared as part of cancer diagnostics. Image analyzes of such tissue sections using artificial intelligence, in particular with neural networks (CNNs), can be used to classify and characterize suspected cancerous lesions at a high level - similar to "classic" pathology. The CNN-based quantitative and qualitative analyses can very probably also identify new biomarkers on tissue sections that are difficult or impossible to detect by eye. The aim of our working group is to develop such new "digital" biomarkers in cooperation with our clinical partners in order to support the work of pathologists and to achieve the most effective therapy possible with the least possible side effects for cancer patients.

In order to minimize the transmission of sensitive health data, we are working on training our models using the principle of federated learning, in which the data remains with the cooperation partners and we only receive model updates.

In cooperation with the Cancer Information Service (KID), we are also setting up the Internet-based communication platform fragdiepatienten.de, which will incorporate the patient's perspective more strongly into oncology projects through short surveys on cancer topics.