移窗自注意力与卷积融合的医学图像分割网络

Medical Image Segmentation Network Based on Shifted Window Self-Attention and Convolution Fusion

医学图像分割是临床诊疗中的关键技术,为疾病诊断提供可靠依据。由于病灶或器官等区域尺度不一、小目标难以辨识,且边界信息较弱,容易导致过分割或欠分割问题。提出一种移窗自注意力与卷积融合的医学图像分割网络STrongUNet,使用卷积提取浅层特征,小感受野获得细粒度信息;通过移窗自注意力机制对局部信息进行增强提取,解决长距离依赖问题;编码器和解码器结构对称,并用两种跳跃连接方式融合高级和低级特征,实现多尺度融合的精准分割。在多器官分割数据集Synapse上的实验表明,在Dice相似系数(DSC)和Hausdorff距离(HD)评估度量上至少提高了1.94%和4.99%。

Medical image segmentation is a key technology in clinical diagnosis and treatment,which provides a reliable basis for disease diagnosis.Due to different regional scales of lesions or organs,unrecognisable small targets,and the weak boundary information,it is easy to lead to over segmentation or under segmentation.In this paper,a medical image segmentation network STrongUNet is proposed,which combines shifted window self-attention and con-volution.It uses convolution to extract shallow features and small receptive fields to obtain fine-grained information.The local information is extracted by shifted window self-attention mechanism to solve the problem of long-distance dependence.The encoder and decoder are symmetrical in structure,and high-level and low-level features are fused by two kinds of skip connection methods to achieve accurate segmentation of multi-scale fusion.The experimental re-sults on Synapse,a multiple organ segmentation dataset,have improved by 1.94%and 4.99%in the evaluation meas-ures of Dice similarity coefficient(DSC)and Hausdorff Distance(HD),respectively.