Abstract: Due to the small difference between the focus area of pulmonary tuberculosis and the normal lung area, it is difficult to accurately detect the pulmonary tuberculosis disease. To solve this problems, a tuberculosis disease detection algorithm based on the combination of deep separable convolution and graph convolution is proposed. Firstly, the depth separable convolution module is used to extract the local features of the image. Secondly, the graph convolution module is used to obtain the global features of the image. Then, using a shortcut branching operation, the extracted local and global features are fused. Finally, the fused features are output through the linear layer to output the detection results. The algorithm model has been fully experimented and verified on the publicly available positive chest X-Ray dataset collected by the Third Hospital of Shenzhen City, Guangdong Province, China. Experimental results show that compared with the benchmark model, the proposed algorithm model improves by 2.98, 3.23, 2.94 and 3.08 percentage points on the four evaluation indicators of Accuracy, Precision, Recall and F1-Score, respectively, which proves the effectiveness of the proposed method.

Key words: X-Ray, depth separable convolution, graph convolutional network, tuberculosis detection

摘要： 由于肺结核病灶区域与正常肺部区域之间的差异微小，导致肺结核疾病难以准确检测。针对此问题，提出了一种基于深度可分离卷积和图卷积相结合的肺结核疾病检测算法。使用深度可分离卷积模块提取图像的局部特征；使用图卷积模块提取图像的全局特征；通过一个捷径分支操作，将提取的局部特征和全局特征进行融合；将融合后的特征经过线性层输出检测的结果。算法模型在中国广东省深圳市第三医院收集的公开可用的正面胸部X光片数据集上进行了充分的实验与验证。实验结果表明，所提算法模型与基准模型相比，在Accuracy、Precision、Recall和F1-Score四个评价指标上分别提高了2.98、3.23、2.94和3.08个百分点，从而证明了所提方法的有效性。

关键词: X光片, 深度可分离卷积, 图卷积网络, 肺结核检测