The thesis examines shape-aware loss functions used for training neural networks for segmentation. It studies their applicability to biomedical objects whose shape is complex to describe explicitly, e.g., cells or organoids. In the theoretical part, the student will introduce essential concepts for deep learning segmentation. He will choose at least three shape-aware loss functions and elaborate on their principles. The experimental part will be done over an in-home dataset of organoids with manual annotations. The student will implement chosen loss functions in a deep-learning API. He will train a vanilla convolutional network using both standard and shape-aware loss functions and study the difference in results. He will use proper techniques to ensure all models are trained well. For evaluation, he will use a wide range of metrics comparing predicted segmentation and reference masks' shape properties.