基于双注意力的肺癌半监督学习分割网络

Semi-supervised learning lung cancer segmentation network based on dual attention

肺癌计算机断层扫描(computed tomography, CT)中对病灶的分割目前存在两个问题:一是病变区域的大小和形状差异大,二是标注数据量少。为了应对上述问题,提出一种用于肺癌分割的双注意力半监督学习网络(dual attention semi-supervised learning network, SDA-Net)。首先,在U-Net的编码阶段加入残差-密集块(residual-dense block, RDB)进行特征提取,尽可能多地保留浅层特征。其次,在编码阶段末端利用包含位置注意力和通道注意力的双注意力机制整合同一类别特征的语义相关性,增强目标的特征表达。最后,针对标注数据量少的问题,采用双路一致性半监督学习(semi-supervised learning, SSL)的方法,使得双注意力网络同时利用标注数据和未标注数据进行训练,大幅提高了网络分割的性能。测试结果表明,所提方法的Dice相似系数、杰卡德系数、灵敏度和精确度分别达到了0.843 2、0.733 1、0.809 2和0.886 1,优于当前典型的分割算法。

Currently, there are two problems in segmenting lung cancer from computed tomography(CT) images.One is that the size and shape of the lesions are different, and the other is that the quantity of labeled data is small.To figure out those above issues, this paper proposes a dual attention semi-supervised learning network(SDA-Net) for lung cancer segmentation algorithm.Firstly, the residual-dense blocks(RDBs) are added to the encoding stage of U-Net for feature extraction to retain more shallow features as much as possible.Secondly, the semantic relevance of the same class features is integrated at the end of the encoding stage using a dual attention mechanism containing position attention and channel attention to enhance the feature representation of the target.Finally, to address the problem of small quantity of labeled data, a two-path consistent semi-supervised learning(SSL) method is used to enable the dual attention network to be trained with both labeled and unlabeled data, which significantly improves the segmentation performance of network.The test results show that the Dice similarity coefficient, Jaccard index, sensitivity and precision of the proposed method achieved 0.843 2,0.733 1,0.809 2 and 0.886 1,respectively, which outperforms the current typical segmentation algorithms.