AI for Multi-center Data

AI for Multi-center Data #

Wilson Silva, Valentina Corbetta, Daan Boeke

Background #

It is extremely common that a deep learning algorithm performs well in one medical dataset but is unable to generalize to a different dataset. This is mainly due to different medical datasets being characterised by different sources (e.g., scanners) and different populations. AI for multi-center data focuses on the development of robust algorithms, techniques to share data, and understanding of AI decisions.

Objectives #

Generalizability #

Make AI-models work across different centers. Development of robust models:

  • Through disentanglement representation learning;

  • Inductive bias (embedding expert knowledge into the models);

Privacy #

Protect patient privacy:

  • Develop visual anonymization models to share realistic synthetic data;

  • Propose federated learning algorithms to work with multi-center data without data sharing;

Explainable AI #

Make it easier to understand and trust AI decisions:

  • Case-based Interpretability (factual and counterfactual explanations);

  • Semantic concept discovery;

People #

  • Wilson Silva, PhD. Wilson is the PI behind the AI for multi-center data research line.

  • Valentina Corbetta (PhD student, University of Maastricht) [Dec 2022 - ongoing]. Valentina is exploring disentanglement learning techniques for segmentation and classification algorithms in the context of medical multi-center data.

  • Daan Boeke (PhD student, University of Maastricht) [March 2023 - ongoing]. Daan is exploring the potential of AI for immune micro-environment characterization and prediction of treatment response.

  • Melanie Groeneveld (MSc student, VU Amsterdam) [Jan 2023 - ongoing]. Melanie’s research focuses on embedding expert knowledge into AI models for increased robustness and interpretability.

  • Filipe Campos (MSc student, University of Porto)

  • Rita Mendes (MSc student, University of Porto)

  • Elena Mordmillovich (BSc student) [Sep 2023 - ongoing]. Elena is developing language models to analyse radiology reports.

  • Daniel Silva (MSc student, University of Porto) [Feb 2023 - Sep 2023]. Daniel worked on disentangling scanner and disease features to come up with robust models for disease diagnosis.

  • Isabela Miranda (MSc student, University of Porto) [Feb 2023 - Sep 2023]. Isabela worked on self-supervised techniques to improve model performance and robustness in breast multi-center data.

  • Laura Latorre (MSc student, VU Amsterdam) [Feb 2023 - Aug 2023]. Laura’s research focused on federated learning AI models and how we move towards federated explainability.

  • Lena Stelter (Intern) [Feb 2023 - July 2023]. Lena developed federated polyp segmentation models for colonoscopy multi-center data.