Enhancing Unsupervised Segmentation Frameworks for Volumetric Medical Images via Superpixel Segmentation and Agglomerative Clustering.

Medical image segmentation plays a crucial role in precise diagnosis and disease monitoring. Current state-of-the-art (SOTA) segmentation methods, such as nnUNet [1], require a large amount of human-annotated segmentation ground truths, which are time-consuming to produce. Furthermore, the quality of such ground-truth depends on clinician's level of expertise and imperfect segmentation may affect the performance of deep learning model. In this paper, we aim to develop a general segmentation framework that can generate precise segmentation results for volumetric images without requiring human-annotated ground truths or domain knowledge of human anatomy.