The analysis of cell nuclei and their spatial distribution is crucial for clinical diagnosis and treatment. While current state-of-the-art methods focus on generating pixel-perfect masks which perform well on benchmarks, they often fail to capture the true variability of nuclei shapes. Additionally, pixel-level masks pose challenges for downstream processing by graph neural networks, as they require approximation using shape descriptors, which can further reduce accuracy. This thesis aims to design, train, and evaluate a neural network that predicts cell nuclei instances using shape descriptors, with the potential to better model overlapping nuclei and better reflect their true shapes in histopathological tissue images. The work will utilize open-source datasets such as PanNuke and MoNuSeg. The neural network model will be developed, trained, and evaluated within the cloud infrastructure of the RationAI group.