Author:

Keywords:

PSI_4738, PSI_MIC

Abstract:

The cerebral vasculature is pivotal for the correct functioning of brain tissue. It uses a variety of defense mechanisms to keep the blood perfusion within a functional range. An acute ischemic stroke (AIS) puts this machinery to the test and, as a result of ischemia, brain tissue may get irreversibly damaged over time. Due to the significant variation in the exact state of the defense mechanisms, the progression from reversible (i.e., penumbra) to irreversible (i.e., core) tissue damage is non-trivial. Medical imaging can measure the altered perfusion dynamics and give information on the tissue status of each voxel. The focus of this thesis is to improve the estimation of the final infarct (i.e., core when no penumbra remains) using a convolutional neural network (CNN). In particular, we are interested in a detailed quantitative comparison with clinical practice while acquiring a profound understanding of the underlying mechanisms. The goal is to provide the medical community with a superior tool to quantify the mismatch between core and penumbra, resulting in improved treatment and patient well-being. The extensive comparison between the traditional method and the CNN-based method revealed that a CNN has some clear additional benefits, not only in core and penumbra estimation but also regarding the estimation of the final infarct. It became clear that a CNN can process a larger spatial context and combine it with procedural data. With a similar volume estimation performance, the CNN can provide a superior localization and can be used to obtain probabilistic estimates of the tissue status. In clinical practice, we could harvest the total capacity of the CNN by simulating the final infarct under different scenarios (e.g., early and complete reperfusion, no reperfusion at all), potentially tuned to each clinical center individually. By gaining deeper insights into the working of the CNN, we identified three main aspects that led to an improved final infarct prediction. First, the choice of loss function turned out to be a significant factor, highlighting the effects of inherent ambiguities in the task description. We could add a first layer of improvements using recalibration, a different loss function, or a combination of losses. Then, we could add a second layer of enhancements by tuning the CNN architecture to incorporate contra-lateral information. Finally, we showed that the CNN could benefit from and estimate perfusion maps. Remarkably, the proposed CNN is relatively robust in terms of temporal resolution of the perfusion imaging. This finding supports research towards the implementation of a one-stop stroke shop. Last but not least, the focus on comprehensibility has led to developing our coding framework: DeepVoxNet2 (DVN2). In DVN2 we have unified the sampling process and the CNN itself, thereby keeping track of the spatial origin of the data. DVN2 provides users with an intuitive environment containing tools to create and exchange end-to-end processing pipelines and utility and analysis functions to practice medical image analysis.