Multi-label Abdominal Image Segmentation with Partially Labeled Data: A Prototypical Consistent Learning Perspective

Guoxia Xu, Hao Wang, Marius Pedersen, Hu Zhu, Meng Zhao

Abstract: Recently, accurate automatic computed tomography (CT) segmentation of organs and tumors has the potential to facilitate clinical diagnosis and therapy. However, the automatics segmentation of multiple organs and tumors (MOTs) is a complex task since they present variability in the partially labeled data due to limited manpower and resources. The most prevalent techniques are committed to proposing a unified framework for the multi-task segmentation problem while suffering from the domain gap and discrepancy caused by the imbalance of data distribution. To handle the aforementioned imbalance challenges, we introduce a novel prototype assignment strategy as a weak enhancement information for a compact intra-class feature representation. Moreover, an exponential-based probability regularization term is proposed to avoid the inter-class imbalance problem caused by forcing the network to provide a consistent prototype label for adjacent features. Experiments comprehensively illustrate the performance of the proposed method compared with other state-of-the-art (SOTA) approaches both qualitatively and quantitatively.