Abstract:Objective To study the superiority of Swin Transformer (SwinT) network in automatic classification for X-ray images of
pulmonary tuberculosis patients by comparing with Vision Transformer (ViT), ConvNeXt and VggNet network. Methods Creating
the hybrid datasets by mixing the public datasets of pulmonary tuberculosis in Montgomery and Shenzhen hospital, which was
expanded to 3940. Then, randomly generated training set, validation set and test set in the ratio of 60∶20∶20. Firstly, we trained
VggNet16 and VggNet19 network as the benchmark by using Fine-tune method of transfer learning, and then trained the Base
network and Large network of SwinT, ViT and ConvNeXt separately in the same way. Finally, the effect of automatic classification
for pulmonary tuberculosis in terms of accuracy, confusion matrix, receiver operating characteristic curve and Grad-CAM were
evaluated. Results The SwinT-Large network had the best classification effect with 98.85% accuracy. There were only 2 false
negative images and 7 false positive images in total 788 test datasets, and the Grad-CAM showed that attention was almost focused
on the lungs, and the accuracy of feature extraction was the best. Conclusion The SwinT network performs well in the task of
classification of pulmonary tuberculosis images, which can be used as a new auxiliary diagnosis method in addition to the traditional
deep learning convolutional neural networks to reduce the rate of missed diagnosis of pulmonary tuberculosis.
刘学思,聂瑞,张和华,杨利,段傲文. 基于Swin Transformer网络的肺结核影像自动分类效果评价[J]. 中国医疗设备, 2022, 37(8): 25-31.
LIU Xuesi, NIE Rui, ZHANG Hehua, YANG Li, DUAN Aowen. Effect Evaluation of Automatic Classification of Pulmonary Tuberculosis Images Based on
Swin Transformer Network. China Medical Devices, 2022, 37(8): 25-31.