多尺度融合编码与自注意力的肺部CT分割算法

Lung CT Image Segmentation Algorithm with Multi-Scale Fusion Encoder and Self-Attention

摘要: 计算机断层扫描(CT)在当前是一种辅助检测肺炎的有效手段，但病理表征的复杂性给医生诊断时带来不便，难以准确地对图像进行分割。为进一步辅助医生根据病理表征诊断病情，本文基于U-net提出了一种多尺度融合编码网络，并结合自注意力机制，力图在辅助医生判断的角度提供可行性方案。为了获取不同尺度的语义信息，首先在编码器部分设计了一种多尺度融合编码器模块，提取不同尺度的特征，充分感知语义信息；同时在编码器和解码器之间的跳连部分引入了改进的自注意力机制，使得网络更好地关注不同语义特征的相关性；最后，采用融合Dice损失函数，Focal损失函数，交叉熵损失函数构建的多级损失函数，更好地约束训练。通过训练公开的数据集，得到分割结果表明Dice相似系数、精确率(Precision)、召回率(Recall)分别为75.37%、77.03%、71.87%，优于其他的模型。我们验证了改进的网络能够在图像分割过程中提升网络性能的可能性。

Abstract: Computed tomography (CT) is currently an effective means to assist in the detection of new coro-nary pneumonia, but the complexity of pathological representations brings inconvenience to doc-tors when diagnosing lesions, and it is difficult to accurately segment the images. In order to further assist doctors in diagnosis based on pathological representations, we propose a multi-scale fusion coding network based on U-net, combined with the self-attention mechanism, trying to provide a feasible solution from the perspective of assisting the doctor’s judgment. In order to obtain seman-tic information of different scales, firstly, a multi-scale fusion encoder module is designed in the en-coder part to extract features of different scales and fully perceive the semantic information; at the same time, a skip connection between the encoder and decoder is introduced. The improved self-attention mechanism makes the network pay more attention to the correlation of different se-mantic features; finally, the multi-level loss function of the fusion Dice loss function, Focal loss func-tion, and cross-entropy loss function is used to better constrain training. Through training the pub-lic data set, the results show that the Dice similarity coefficient, precision rate and recall rate are 75.37%, 77.03%, and 71.87%, respectively, which are better than other existing network models. We have verified the feasibility of the improved network being able to improve network perfor-mance during image segmentation.