KEMAI - Research Training Group on Medical XAI

We believe that in the field of medical AI, not only physicians need to learn about and adapt to the tools at hand, but also that AI systems need to be developed with a clear understanding of the medical diagnosis process, and the needs of medical experts, as well as ethical challenges. Unfortunately, current research and the thereof derived medical AI systems often rather focus on prediction accuracy, as measured against a chosen test set, than on the understanding and communication of the obtained predictions. Therefore, within KEMAI we aim at combining the benefits of knowledge- and learning-based systems, to not only allow for state-of-the-art accuracy in medical diagnosis, but to also clearly communicate the obtained predictions to physicians, considering ethical implications within the medical decision process.

KEMAI Project Outline

To address the KEMAI research goals, we plan for the following four doctoral research topic areas, of which each covers multiple specific dissertation topics:

• A - Data Exploitation – Harvesting medical guidelines and investigating contrastive pre-training
• B - Knowledge Infusion – Enriching learning models with medical guidelines
• C - Knowledge Extraction – Deriving medical guidelines from trained models
• D - Model Explanation – Communicating the decision process of trained models

The following figure shows an overview of how these research topic areas are related, and how individual dissertation projects (A1…D3) are related to these research topic areas. More information on the individual projects can be found on the KEMAI projects page.

All participating researchers will work cooperatively on these four topic areas, and contribute not only technical and medical expertise, but also consider and address the ethical implications.

Medical Use Cases

To ensure, that the four KEMAI topic areas are also well grounded in the medical application domain, we have identified the following three medical use cases to be worked on.

COVID-19 CT Imaging Diagnosing COVID-19 with chest CT reveals suspicious abnormalities with a 97% sensitivity. Current guidelines highlight CT’s role in detecting, distinguishing, and monitoring COVID-19-related lung signs. However, CT patterns can be similar in COVID-19 and other pneumonia types. Modern CNNs can help differentiate these conditions, achieving 86.27% accuracy and 83.33% specificity. Ulm University Hospital, a core lab for the RACOON trial, utilizes a large COVID-19 patient database for training algorithms to analyze prognostic and predictive information from CT imaging using AI-based software.

Bronchial Carcinoma PET/CT Imaging Lung cancer is a leading cancer type with a poor prognosis. Early-stage NSCLC (Non-Small-Cell Lung Cancer) patients could benefit from better risk stratification using AI techniques on 18F-FDG PET/CT images. At Ulm Medical Center, over 2000 PET/CT scans are performed annually, providing a rich dataset for training algorithms to improve prognostic stratification. The Comprehensive Cancer Center Ulm maintains a well-curated database to enhance predictive information.

Echinococcosis PET/CT Imaging Human alveolar echinococcosis (AE) is a lethal parasitic disease prevalent in central Europe. At Ulm Medical Center, whole-body FDG PET/CT scans are used to diagnose and monitor AE, evaluating disease activity and treatment response. AI-based approaches aim to derive novel imaging biomarkers for better disease characterization. Ulm, a major echinococcus center, handles over 80 AE cases annually, providing a substantial database for training predictive algorithms.