Generating vascular network structures from real data

• Novel data-oriented model to generate vascular networks • An encoder-decoder architecture to approach the generation of vascular networks in a sequential manner • Model’s hyperparameters allow to regulate the morphology of resulting networks Blood vessels are essential components of the mammalian circulatory system, responsible for delivering nutrients and oxygen to tissue cells. Studying the structural properties of these complex vascular networks is of great importance given their role in several diseases. Despite recent advances in understanding the biological mechanisms underlying vascular growth and the vast number of computational models that use those mechanisms to simulate vascular development, researchers have not yet tackled the generation of synthetic vascular structures using deep learning implementations that are trained with existing datasets. This work aims to address these challenges by introducing a novel computational approach to generate three-dimensional vascular networks. By using an encoder-decoder architecture, we treat these networks as complex spatial graphs in a three-dimensional space, learning sequential moves to create new nodes and edges that mimic real vascular structures. Without replicating the real-truth network, our method can generate new vascular networks with structural characteristics identical to those of existing real data, such as branching pattern, vessel lengths, network density, and degree distribution. The parameters of the model can be adjusted to regulate and fine-tune the morphology of the generated networks. The achieved results not only demonstrate the potential of our approach in accurately modelling vascular structure but also open new avenues for exploring data-driven models in tissue engineering and to complement imaging data collected in clinical and research contexts.