The DeepTos project is an exploratory project funded by the GDR IASIS which aims to improve 3D vascular segmentation by leveraging the Tree of Shapes (ToS), a hierarchical image representation from mathematical morphology. Current deep learning approaches often produce fragmented segmentations that fail to reflect biological continuity and hinder applications such as flow simulation. In DeepToS, the segmentation task is reformulated as a classification problem on the nodes of the Tree of Shapes, allowing for direct control over the connectivity of the resulting segmentation. The project is structured in two phases: first, learning descriptors for each node using a deep autoencoder such as a UNet, and second, classifying the nodes with a Graph Convolutional Network (GCN) to isolate those representing the vascular structure.