Hierarchical Segmentation of the Female Pelvis: A Novel Framework for Uterus,Healthy Ovaries, and Ovarian Lesions

The main objective of this study is the segmentation of anatomical structures, with particular emphasis on the uterus—specifically the myometrium and endometrium—as well as healthy ovaries and adnexal masses. A baseline segmentation model based on a U-Net architecture and focal loss is first developed. Building on this, we investigate the transition from standard segmentation frameworks—where relationships between structures are not explicitly considered—to a hierarchical segmentation framework in which anatomical structures are organized according to an explicit hierarchy. Within this perspective, a hierarchical model based on a predefined tree structure is introduced to capture and exploit the relationships among anatomical structures. Two different hierarchical strategies are proposed. The first is a hybrid method based on a U-Net architecture, in which a focal hierarchical loss is designed to incorporate hierarchical relationships only during the optimization process, while preserving the same prediction structure as the baseline model. The second is a fully hierarchical approach, referred to as HSSN, in which the hierarchy is explicitly embedded in the network architecture, enabling multi-level predictions that reflect the relationships between anatomical structures. Experimental results demonstrate that incorporating hierarchical relationships significantly improves segmentation performance compared to the baseline model. In particular, the hybrid approach outperforms the fully hierarchical model and achieves improvements exceeding 10% in the segmentation of structures such as the uterus and healthy ovaries. These findings highlight the effectiveness of hierarchical modeling in enhancing the accuracy and consistency of medical image segmentation.