Medical Image Segmentation of Liver Tumors with Multi-phase Deficiency Based on Hierarchical Knowledge Distillation Network

Contrast-enhanced Computed Tomography (CT)-based multi-phase liver tumor medical image segmentation is one of the important methods of liver cancer diagnosis. However, the multiply phases of patients might be incomplete clinically, resulting in the clinical limitation of multi-phase liver tumor medical image segmentation. Recently, the knowledge distillation-based incomplete multi-phase/modal methods conduct the paradigm of teacher-student networks, in which the knowledge of complete modalities are transferred to the student network via teacher network, achieving the contrast-enhanced CT-based incomplete multi-phase liver tumor medical image segmentation. However, such paradigm lacks the effective hierarchical feature representation and the texture information transmission mechanism, resulting in the significant degradation of student network performance against the incomplete modalities. To address this issue, a hierarchical knowledge distillation network (HKD-Net) is proposed to guide the student network to learn the hierarchical knowledge, which consists of teacher-student network, style matching module, context consistency module, hierarchical discriminator, and segmentation consistency module. Experimental results demonstrate the superiority of the proposed method compared with the state-of-the-art incomplete multi-phase/modal medical image segmentation methods, achieving the reliable incomplete multi-phase liver tumor segmentation.