Multi-Organ Segmentation Based on Multi-Granularity Prototype Learning

The scarcity of annotated data is an important factor restricting the multi-organ segmentation technology. To address this issue, recent research has introduced prototype-based method into the realm of multi-organ segmentation. However, existing prototype-based method have focused only on specific ranges of local feature information, neglecting the impact of local feature information across different ranges on segmentation results. In this paper, we proposed a multi-organ segmentation model based on multi-granularity prototype learning (MGPL_Net), which can make full use of different local feature information by obtaining multi-granularity prototypes. Firstly, we proposed a multi-granularity prototype extraction and fusion (MGPEF) mechanism to effectively utilize different local feature information of support image. Secondly, we proposed a multi-granularity prototype alignment (MGPA) mechanism to enhance the representation capability of local prototypes at different granularities by establishing information closed-loop between support and query. We conducted experiments on two datasets, validating the effectiveness of this approach. The segmentation accuracy improved by 3.91% on the ABD-MRI dataset and by 7.28% on the ABD-CT dataset.