【目的】提出一种时间序列下双流增强的肺癌生长演变预测模型(DSGNet)。【方法】DSGNet充分利用CNN以及Transformer的优势,通过基于CNN的分支来提取肿瘤的静态特征,并用多尺度的方式加强所提取到的特征表示。用基于Transformer的分支来获取肿瘤序列图像间的顺序依赖关系,该分支将病灶序列图像映射成一个特征映射序列,然后将特征映射序列输入包含有多头自注意力的深度网络中,并从多期病灶序列图像的特征映射中,提取完整的肿瘤间生长关系。【结果】在肺癌NLST以及合作医院的数据集上评估该算法,实验结果表明,DSGNet对肿瘤生长预测的Precision达到了92.45%、Dice系数为82.78%,相比于其他的肿瘤预测算法,本文所提DSGNet在各方面都有一定程度提升,且在多方面被证明能够应用于临床研究。

【Purposes】 In this paper, a dual stream enhanced lung cancer growth evolution pre⁃ dictive network under time series (DSGNet) is proposed. 【Methods】 With full use of the advantages  of CNN and Transformer, PSGNet can be used to extract the static features of tumors through CNN- based branches, and strengthen the extracted feature representations in a multi-scale manner. The  Transformer-based branch helps obtain sequential dependencies between tumor sequence images,  with its core component being the multi-head self-attention layer proposed in this paper. This branch  maps lesion sequence images into a feature map sequence and then inputs the sequence into a deep net⁃ work with multi-head self-attention, from which the complete inter-tumor growth relationship is ex⁃ tracted. 【Findings】 The proposed algorithm is evaluated on the lung cancer NLST dataset and a data⁃ set from cooperative hospitals. The experimental results show that DSGNet achieves a precision of  92.45% and a Dice coefficient of 82.78% in predicting tumor growth. Compared with other tumor pre⁃ diction algorithms, DSGNet proposed in this paper has been improved to a certain extent in all aspects  and has been proven to be applicable to clinical research in many ways.