Publication

Abstract : Image segmentation represents a fundamental technique widely employed in diverse fields, with particular prominence in the biomedical domain for extracting intricate and crucial entities from images. Within the scope of this research, we have harnessed the potent capabilities of the U-Net architecture to perform image segmentation specifically on junctions discernible within neurovascular MRI images of the human brain. This undertaking serves as a crucial stepping stone towards the overarching goal of constructing a comprehensive and finely labeled 3D graph representation of the neurovascular system. The motivation behind developing a neurovascular mapping of the human brain through Convolutional Neural Networks (CNNs) lies in the potential insights we can glean concerning the intricate dynamics of blood flow within this complex organ. By scrutinizing the nodes and connections of this neurovascular mapping, we gain a more profound understanding of the hemodynamic processes, thus paving the way for subsequent investigations into diffusion modeling. The diffusion model derived from these hemodynamic insights holds promise for early stroke detection by identifying regions of insufficient blood flow before the onset of clinical symptoms. In the course of our current study, albeit conducted with a limited set of training examples, we have determined that the InceptionV3 architecture, when employed as the backbone for U-Net-based segmentation, emerges as the optimal choice. This selection is rooted in rigorous empirical analysis and serves as a vital component in our pursuit of advancing the accuracy and effectiveness of neurovascular image segmentation for improved stroke detection and neurovascular system mapping.