Clinical Applications

Medical images are crucial for diagnostics, but frequently distributed across various centers, precluding direct data sharing. Federated analyses and learning, in which algorithms are distributed instead, are a possible solution to this problem. In this project, our objective is to assess the robustness of segmentation algorithms by validating them on a large-scale federation. Furthermore, we aim to enhance the trustworthiness of these algorithms by developing methods capable of detecting potential inaccuracies in segmentation results. This proactive approach ensures the identification of instances where segmentation may be erroneous, contributing to the overall reliability of medical image analysis in distributed environments.

Contact: Maximilian Zenk

Breast cancer is the most invasive cancer for women throughout the world, in both developed and developing countries. Diffusion-Weighted Imaging (DWI) has potential in breast cancer screening, as it is a fast, safe, and accurate acquisition technique. This project aims to create robust deep learning models for breast lesion detection and classification using DWI and to produce a software platform that can assist clinical decision making by providing AI-based diagnostic suggestions.

Contact: Dimitrios Bounias

CT-navigated spinal instrumentation requires intraoperative screw trajectory planning in CT volumes. In the current clinical routine this is often performed manually, which is prone to error and time-consuming. This project focuses on the development of deep learning-based methods for automatic image-based spine screw planning. Leveraging a large intraoperative planning dataset the screw planning task is interpreted as a segmentation task and screw dimensions, location and orientation is automatically predicted based on the image context.

Contact: Alexandra Ertl

Multiple Myeloma (MM) is a malignancy of bone marrow plasma cells, so-called myeloma cells, which disrupt the production of new blood cells and cause bone breakdown. In recent years, modern imaging technologies such as Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) have gained a lot of attention in diagnosis and staging of MM and the standardized, comprehensive evaluation of whole-body imaging is of great interest. In this project, we investigate fully automatic image analysis methods for the diagnosis and image staging of myeloma patients. The investigation encompasses tasks like bone marrow segmentation, lesion detection and subsequent analysis utilizing radiomics and deep learning methodologies.

Contact: Jessica Kächele

To develop a research-supporting risk prediction platform for the National Cancer Prevention Center, we are currently assembling an interdisciplinary digital cancer prevention team. The focus of the working group is on the development of a specific and evidence-based portal for the individual calculation of personal cancer risk and lifestyle recommendation. In doing so, existing prediction models will be validated, curated and merged according to a standardized procedure.

Through this portal, interested citizens should be able to easily evaluate their individual cancer risk, receiving tailored information based on their personal profile. The calculation incorporates key factors such as demographic data, lifestyle information, family history, and past test results. Simultaneously, user data is instrumental in refining and optimizing our prediction models, ensuring a consistently high level of performance.

Our long-term vision is to establish a research-capable platform that facilitates sustainable data collection and provides access to research data in the field of modern prevention research.

Contact: Odile Elias

With recent advances in deep learning, there is a promising foundation for achieving the fast detection of aortic dissections (AD) in clinical CT data acquired under emergency conditions. Such computerized detection can help drastically decrease the reaction times for urgent cases, which are regularly delayed due to the absence of clear symptoms. This project is focused on the development of a fast and accurate approach for AD detection in emergency CT data. Our goal is to alert clinicians in the event of an acute and life-threatening AD, and more generally to provide an initial assessment of the specific AD type (after Stanford classification), and of any coronary or carotid artery involvement. To this end, we perform in-depth analyses on segmentation and detection approaches for solving this task, with a specific focus on the generalizability of the employed techniques across heterogeneous multi-centric data.

Contact: Tobias Norajitra

Detecting and segmenting colorectal cancer (CRC) and adenomas remains a significant challenge in medical image processing due to the complexity of imaging in the colon. To address this, our project shifts the focus to the liver, which is easier to segment and serves as a critical site of interaction with the colon. We use deep learning approaches to stratify healthy, adenoma, and CRC patients. Radiomics features extracted with RPTK serve as an integrative approach to include clinical parameters, and conduct a comprehensive analysis to differentiate between healthy individuals, adenoma patients, and CRC cases. This approach aims to provide a novel, reliable diagnostic solution while bypassing the challenges of direct colon segmentation.

Contact: Jonas Bohn, Darya Trofimova

Medical Imaging is prone to domain shifts (changes in acquisition protocols, diverse scanners, variable population, etc.), often leading to poor generalizability while impeding AI model deployment in real-world systems. The Medical Image Computing division in DKFZ has developed an algorithm for vessel occlusion detection from contrast-enhanced Computer Tomography (CT) images known as "angiographies" (https://www.nature.com/articles/s41467-023-40564-8), across several clinics in Germany. The algorithm performs well in general, but undesirable performance drops may happen while dealing with out-of-distribution data. We investigate techniques that help guarantee model robustness against shifts (domain generalization, incorporation of prior information from brain vasculature, blood circulation dynamics, etc.). Data from more clinics will be made available in the near future, so having a robust object-detection algorithm in 3D medical images against potential domain shifts is crucial.

This project is also part of the European Laboratory for Learning and Intelligent Systems (ELLIS), an effort to bring together the top AI-research labs into a joint environment (ELLIS, ELLIS Life Heidelberg, Andrés Martínez Mora's ELLIS profile). Part of the Project is being conducted together with researchers from the University of Amsterdam and Amsterdam University Medical Centers.

Contact: Andrés Martínez

This project focuses on the non-invasive characterization of intratumoral heterogeneity in sarcoma patients, with a particular emphasis on young and adolescent cases. Utilizing radiomics features extracted and selected through RPTK, we analyze data derived from PET-MRI scans to uncover critical insights into tumor composition. By quantifying heterogeneity, our approach aims to improve understanding of tumor dynamics, enabling personalized treatment strategies and enhancing prognostic accuracy in this vulnerable patient group.

Contact: Jonas Bohn