基于轻量级RG-DenseNet的COVID-19 CT图像分类-《中国医学物理学杂志》

文章信息/Info

作者:: 张子宇1; 赵可辉2; 牛慧芳3; 张志强1; 周连田4; 1.山东中医药大学智能与信息工程学院，山东济南 250000； 2.山东中医药大学第二附属医院特检科，山东济南 250000； 3.山东省药品不良反应检测中心，山东济南 250000； 4.菏泽市中医医院碎石科，山东菏泽 247000

Author(s):: ZHANG Ziyu1; ZHAO Kehui2; NIU Huifang3; ZHANG Zhiqiang1; ZHOU Liantian4; 1. College of Intelligence and Information Engineering, Shandong University of Traditional Chinese Medicine, Jinan 250000, China 2. Special Inspection Department, the Second Affiliated Hospital of Shandong University of Traditional Chinese Medicine, Jinan 250000, China 3. Shandong Province Adverse Drug Reaction Testing Center, Jinan 250000, China 4. Department of Lithotripsy, Heze Traditional Chinese Medicine Hospital, Heze 247000, China

Keywords:: Keywords: RepGhost DenseNet COVID-19 deep learning image classification

摘要:: 【摘要】目的：基于轻量级RG-DenseNet构建COVID-19 CT 图像分类模型。方法：以DenseNet121为基础，添加通道和空间注意力机制模块减少无关特征的干扰，将DenseNet中的Bottleneck模块替换为前激活的RG-beneck2模块减少模型参数的同时保持精度尽可能不变。构建RG-DenseNet模型，在COVIDx CT-2A数据集上进行3分类实验。结果：RG-DenseNet准确率为98.93%、精确率为98.70%、召回率为98.97%、特异性为99.48%、F1分数为98.83%。结论：RG-DenseNet与原模型DenseNet121相比在保持准确度仅降低0.01%的情况下，减少92.7%的参数量和计算量，轻量化效果显著，具有实际应用价值。

Abstract:: Abstract: Objective To construct a COVID-19 CT image classification model based on lightweight RG DenseNet. Methods A RG-DenseNet model was constructed by adding channel and spatial attention modules to DenseNet121 for minimizing the interference of irrelevant features, and replacing Bottleneck module in DenseNet with pre-activated RG beneck2 module for reducing model parameters while maintaining accuracy as much as possible. The model performance was verified with 3-category classification experiments on the COVIDx CT-2A dataset. Results RG-DenseNet had an accuracy, precision, recall rate, specificity, and F1-score of 98.93%, 98.70%, 98.97%, 99.48%, and 98.83%, respectively. Conclusion Compared with the original model DenseNet121, RG-DenseNet reduces the number of parameters and the computational complexity by 92.7%, while maintaining an accuracy reduction of only 0.01%, demonstrating a significant lightweight effect and high practical application value.