基于全局注意力引导的肺结节分割方法

肺结节的精确分割能有效地辅助医生的治疗诊断工作，但由于不同患者所呈现的肺结节病灶形式多种多样，基于传统专家系统和统计学习的方法难以获得准确的肺结节分割结果。针对这种情况，提出一种由全局注意力引导的注意力机制，达到了从一张完整的胸部影像切片中自动定位并分割出肺结节的效果。该方法首先对目标区域进行肺实质分割，再利用区域建议网络（region proposal network，RPN）进一步缩小感兴趣区域，并生成注意力权重图，最后使用融合了残差网络（residual network，ResNet）与卷积长短期记忆网络（convolutional long short-term memory，ConvLSTM）的结构结合注意力权重进行肺结节分割。将所提方法在肺图像数据库联盟（lung image database consortium，LIDC-IDRI）数据集上进行了全面的评估，结果表明，本文方法分割结果的平均dice得分（标准差）为89.97%（8.9%），具有出色的分割性能，精度相较其他方法取得一定提升。进一步在相同数据集上将所提方法的肺结节分割结果与4位放射科医生的手工标注结果进行了比较，结果表明本文方法的分割结果与医生们的标注结果的一致性达到了85.81%，相较于医生们手工标注之间的一致性高出了3.39%。

A method for segmentation of lung nodules based on global attention guidance

Precise segmentation of lung nodules can effectively assist doctors in treatment and diagnosis. However, due to the various forms of lung nodules presented by different patients, it is difficult to obtain accurate lung nodule segmentation results based on traditional expert systems and statistical learning methods. In response to this situation, an attention mechanism guided by global attention has been proposed in order to automatically locate and segment lung nodules from a complete chest image slice. This method first segments the lung parenchyma of the target area, then uses a region proposal network (RPN) to further reduce the region of interest, and generate an attention weight map, and finally uses the residual network (ResNet)+convolutional long short-term memory (ConvLSTM) structure combined attention weights to segment the lung nodules. The proposed method has been extensively evaluated using the lung image database consortium (LIDC-IDRI) data set. The results show that the average dice score (standard deviation) of the proposed method on the LIDC-IDRI data set is 89.97% (8.9%), with excellent segmentation performance. Furthermore, the segmentation results of lung nodules of the proposed method were compared with the manual annotation results of four radiologists using the same data set, and the results showed that the consistency between the segmentation results of this method and the annotation results of doctors reached 85.81%, which is 3.39% higher than the agreement between the doctors’ manual annotations.