用于多器官分割的多尺度聚合网络研究

Research on multi-scale aggregation network for multi-organ segmentation

多器官分割在病理分析、手术方案制定以及临床诊断上都具有重要的应用价值.但是,一些器官形变较大、尺寸较小且组织边缘模糊,导致分割效果较差.为了解决该问题,提出了一种用于多器官分割的多尺度聚合网络(MSANet).MSANet有两个优势:首先,设计了多尺度边界提取模块,使用多尺度卷积核提取多个特征图,将不同尺度的特征图互相结合,从而聚合全局上下文信息,并提取不同器官的边界和细节信息;其次,设计了聚焦式注意力模块,通过学习的注意力权重来调节特征图的重要性,从而聚焦感兴趣的多器官区域并捕捉不同器官的关键特征,进一步提高分割性能.在两个公开数据集CHAOS和MS-CMRSeg上进行了大量实验.实验结果表明:MSANet在两个数据集上的分割效果均优于当前主流的多器官分割方法,显著提高了多器官分割精度.

Multi-organ segmentation has important application value in pathological analysis,surgical planning,and clinical diagnosis.However,some organs have significant deformation,small size,and blurry tissue edges,resulting in poor segmentation performance.To address this issue,a multi-scale aggregation network(MSANet)for multi-organ segmentation is proposed.MSANet has two advantages.First,a multi-scale boundary extraction module is designed,which uses multi-scale convolutional kernels to extract multiple feature maps.Different scale feature maps are combined with each other to aggregate global context information and extract boundary and detail information of different organs.Secondly,a focused attention module is designed to adjust the importance of feature maps through the learned attention weights,thereby focusing on multiple organ regions of interest and capturing key features of different organs,further improving segmentation performance.Extensive experiments are conducted on two public datasets,CHAOS and MS-CMRSeg.The experimental results demonstrate that the segmentation performance of MSANet is superior to the current mainstream multi-organ segmentation methods on both datasets,improving the accuracy of multi-organ segmentation significantly.