Small Datasets in Personalized Medicine—Integrating Cross-Center Knowledge in Foundation Models

Challenge: Extracting the Needle from the Haystack Small Datasets

• Small patient collectives occur in personalized medicine and in rare diseases. Powerful deep learning methods require large amounts of data.
• Additionally, images in histology vary between clinics and this heterogeneity needs to be covered during training. 
• Find out more in our podcast episode: AI-Based Precision Pathology: Learning with Small Amounts of Data

Training of Use-Case-Specific Foundation Models

• Models for routine use need to be robust even if only small datasets are available for training.
• Foundation models trained on large cohorts provide a basis for covering the required data diversity.
• We build foundation models specifically for applications in pathology and radiology. [1]
• Tissue Concepts is a specialized model for pathology aiming at cross-center generalizability and interactive training of AI models. [2]

References

[1] Overcoming Data Scarcity in Biomedical Imaging with a Foundational Multi-Task Model. Schäfer et al. 2023. article

[2] Tissue concepts: Supervised foundation models in computational pathology. Nicke et al. 2025. article

Quick Training of Custom Models with Small Datasets

• We build robust models based small datasets data or few annotations
• Models can be trained interactively based on pre-computed image features 

(Demo Video: A foundation-model-based few-shot segmentation),  (DGP 2024)

Predicting Outcome from Patient-Level Labels

• We build models based on clinical data to detect new biomarkers and to predict outcome
• Most accurate detector of colorectal cancer in the Semicol-challenge and overall second place!

[1] prosurvival.org

[2] semicol.org

[3] Tissue concepts: Supervised foundation models in computational pathology. Nicke et al. 2025. article

Prostate Survival Prediction and Biomarker Discovery

• We predict recurrence-free survival time in prostate cancer and analyze predictive patterns in tissue morphology.
• We develop a cooperative federated learning platform that allows scientists to collaborate between clinical centers.
• We collaborate on the standardization of data exchange formats for survival studies [1].
• 2nd place in the LEOPARD challenge for prostate cancer risk prediction [2].
   

References

[1] PROSurvival: A Technical Case Report on Creating and Publishing a Dataset for Federated Learning on Survival Prediction of Prostate Cancer Patients. Xu et al. 2021. article

[2] Leopard

[3] prosurvival.org