面向CT图像新冠肺炎识别的密集重参轻量化Transformer模型

新冠(COVID-19)肺炎严重威胁人类健康，基于深度学习的计算机辅助诊断方法能有效提高新冠肺炎的诊断效率。但是深度学习模型结构复杂、参数量和计算量大，在保持模型性能的前提下提高网络轻量化的程度具有重要研究意义，因此，该文提出一种面向CT图像新冠肺炎识别的密集重参轻量化Transformer模型(DRLTransformer)。首先，为提高模型的轻量化程度，构造了重参密集块和层次化Transformer，在保持模型精度的同时提高计算速度，降低模型参数量；然后，为充分提取新冠肺炎病灶的全局与局部信息，设计层次化Transformer增强全局注意力对局部特征相关性的关注程度，其中采用分组提取全局特征，在不同组之间进行融合获得多层次信息，并且进行信息融合，进一步提高组内和组间特征的交互能力，此外对所有全局特征进行聚合，实现深浅层特征深度融合。最后，在新冠肺炎CT数据集中进行对比实验，结果表明该模型参数量和计算量分别为1.47 M和81.232 M，相比密集网络(DenseNet)参数量降低29倍、计算量降低23倍，该模型对新冠肺炎计算机辅助诊断具有积极的意义，为深度学习模型轻量化提供了新思路。

Dense Heavy Parameter Lightweight Transformer Model for CT Image Recognition of COVID-19

COrona VIrus Disease 2019(COVID-19) is a serious threat to human health, deep learning computer aided diagnosis method can effectively improve the diagnosis efficiency. But deep learning models have usually complex structure which have large number of parameters and computations, therefore, a Dense Reparameter Lightweight Transformer(DRLTransformer) for COVID-19 CT recognition is proposed. Firstly, reparameter dense block and hierarchical Transformer are proposed to improve lightweight degree of model, which can improve computation speed and reduce parameters without losing model performance. Secondly, in order to fully extract global and local information of lesions, using hierarchical Transformer enhance global attention on local feature relevance, which use grouping to extract global features and fused between different groups to obtain multi-level information, and then information fusion is used to further improve interaction of intra group and inter group features. In addition, all global features are aggregated to achieve deep fusion of deep and shallow features. Finally, comparative experiments in COVID-19 CT dataset, the results show that the parameters and computations of DRLTransformer are 1.47 M and 81.232 M. Compared to Dense Convolutional Network(DenseNet), parameters are reduced by 29 times and computations are reduced by 23 times. The model proposed in this paper has positive implications for computer aided diagnosis of COVID-19 and provides a new idea for lightweight deep learning model.