In PREDICTioN2020 we have been developing a comprehensive decision support platform for stroke outcome prediction. The project comprises several sub-projects targeting singular use-cases / research problems. This research has been funded by the German Federal Ministry of Education and Research (BMBF) within GO-Bio with €2.2m.
You are here:
To date, most prognostic tools used in clinical practice utilize purely clinical parameters. Additionally, treatment decisions are complemented with image-derived manual scores that are highly time-consuming and require expertise thus limiting their applicability and objectivity. In contrast, our Hybrid Modelling framework provides an observer-independent, robust solution to combine different modalities of imaging data and clinical parameters in the same time.
By leveraging the flexibility of deep learning, we are able to combine various sources of data through sub-models and a Central Intelligence module. In case of stroke outcome prediction, image analysis pipelines can augment clinical data analytics as well as biophysiological simulation. Such a framework not only supports the full range of clinical workflow (image inspection, prognostic analysis of clinical parameters) but allows simple adjustment to new datasets, inputs or image modalities. Furthermore, single module contributions and data-specific explanations (explainable AI; xAI) are developed to open up the integrated deep learning “black boxes”, facilitate interpretation of predictions and make our framework more robust and more reliable.
Blood flow simulation
We developed the first fully implemented brain circulation simulation model that integrates individual stroke patient data and enables simulation of brain blood flow and perfusion based on vascular pathology such as stenosis and occlusion. It enables computation of flow rates, velocities, and perfusion pressures for different brain areas. The model can be adapted to individual neuroimaging patient data and can simulate different blood pressure conditions that can crucially affect the risk assessment. The automated simulation pipeline consists of 4 sub-models (thus sub-projects): Vessel segmentation, Vessel annotation, Hemodynamic modelling and Validation.
For the vessel segmentation sub-model we developed a fully automated deep learning model, trained on more than 200 patients from 3 different clinical cohorts. Our model claims state-of-the-art brain vessel segmentation performance with respect to multiple metrics and provides excellent binary segmentation of the brain vasculature, visually assessed by medical experts.
After segmentation, brain vessels are categorized into anatomical vessel classes by another deep learning model. This is a much more challenging task given the high amount of classes. Our model is currently trained on 100+ patients and delivers high performance classification. We are continuously improving the model and including more patients with expert labels.
Taking the annotated vessels and based on the set blood pressure conditions the simulation engine calculates the blood flow rates in vessels and - crucially - the critical perfusion pressure of the brain supply areas.
Simulation outputs are cross-validated with perfusion maps of time-to-peak (TTP), cerebral-blood-flow (CBF) and time-to-maximum (Tmax) to ensure our hemodynamic simulation model performs on par with perfusion imaging in revealing potential brain areas at risk.