Neurovascular diseases can lead to significant limitations and disabilities in affected individuals and are also one of the most common causes of death in Germany. Patient-specific changes in the
cerebral vessels, which manifest themselves in the form of intracranial aneurysms (permanent, balloon-like vascular bulges), are particularly severe. Although medical imaging techniques are constantly evolving and enable reliable diagnosis, individual risk assessment is extremely complex and subject to numerous influencing factors, which are therefore often simplified in clinical practice.
The AI4IA project therefore aims to develop a robust methodology for computing morphologic and hemodynamic rupture risk parameters for intracranial aneurysms. The project aims to improve the detection, decision-making and treatment planning of intracranial aneurysms (IAs) by developing standardized operating procedures (SOPs) so that both clinicians (neuroradiologists and neurosurgeons) and patients benefit from these automated solutions.
This project includes automated solutions for geometry segmentation with neural networks (e.g. 3D U-Net), followed by automated analysis of morphological parameters and automated computer-aided analysis of hemodynamic parameters. The standardization and automation of these processes should minimize the uncertainty that is transferred from routine clinical image data to the derived rupture risk parameters.
Therefore, the main objectives of this project are to evaluate the uncertainty of rupture risk parameters derived by machine learning (ML) compared to manual segmentation and to identify robust rupture risk parameters to differentiate between ruptured and unruptured IAs.
The development of ML solutions for the automated detection and segmentation of intracranial aneurysms requires a robust database. Therefore, the goal is to generate a standardized data set through the two clinical partner sites in Berlin and Magdeburg. The existing data sets show inconsistencies and pose challenges, such as different image acquisition devices and qualities. To overcome these problems, standards for image acquisition, reconstruction of volume data and manual segmentation will be defined. The standardized data set will then be used to train neural networks for automatic segmentation and rupture state differentiation.
In summary, the project aims to provide significant insights into the robustness of such risk parameters as well as the automation and optimization of their calculation through the multi-layered approach. This will help maximize the usefulness of these risk parameters for clinical decision making and help future clinical translation.