AI for Precision Medicine: From Foundation Models to Tumor Tissue Profiling in Computational Pathology

Christina Embacher is a computer scientist with seven years of experience applying AI to translational research in medicine and biotechnology. As a Senior Data Scientist at Aignostics, she works across deep learning, computer vision, and statistics to develop applications for cancer research and clinical diagnostics. Jonas Dippel is a Senior Data Scientist at Aignostics. He holds a PhD from the machine-learning group at TU Berlin and BIFOLD, where he worked on self-supervised learning, robustness, anomaly detection, and explainable AI. At Aignostics, he focuses on training foundation models for pathology and building search capabilities based on them.

Pathology is central to cancer diagnosis, biomarker assessment, and treatment decisions, but the analysis of tumor tissue is still largely manual, semi-quantitative, and subjective. AI can make pathology more quantitative and scalable, but challenges remain: whole-slide images are extremely large, expert labelling is costly, cancer subtypes are unevenly represented, and differences in staining and scanners can cause models to learn hospital-specific signatures instead of biological patterns.

This workshop explores the robustness of pathology foundation models and the analysis of the tumor microenvironment using open-access, de-identified data.

In the session

Participants will work through two hands-on parts using Python notebooks in the browser, including Google Colab or Molab.

Part I – Robust foundation models: Participants will measure how strongly pathology foundation-model embeddings encode the medical centre rather than biological information, compare models, and prototype possible mitigations such as stain augmentation, feature normalisation, and batch correction.

Part II – Tumor microenvironment insights: Participants will explore OpenTME readouts for 4,691 TCGA slides across eight cancer types to identify tumor-microenvironment patterns that distinguish cancer types or relate to clinical and molecular annotations.

What to expect

Participants can expect an introduction to the AI challenges of digital pathology and current approaches to pathology foundation models. They will gain hands-on experience with foundation-model features and quantitative tissue analysis and receive working notebooks that they can reuse in their own research.

Participant requirements

Participants should:

  • Bring a laptop.
  • Have a Google account for Colab.
  • Have a Hugging Face account using an academic or non-profit email address.
  • Request access to the OpenTME dataset in advance.
  • Be aware that OpenTME is available for non-commercial academic research; commercial use and AI model distillation are not permitted.
The workshop uses open-access, de-identified data, including H&E-stained tissue-image patches, pre-computed tabular readouts, and public datasets such as TCGA and CAMELYON16/17.
Zurück zur Übersicht