Soni Mukesh, Singh Ajay Kumar, Babu K Suresh, Kumar Sumit, Kumar Akhilesh, Singh Shweta
Department of CSE, University Centre for Research & Development Chandigarh University, Mohali, Punjab, 140413, India.
Mody University of Science and Technology, India.
Smart Health (Amst). 2022 Sep;25:100296. doi: 10.1016/j.smhl.2022.100296. Epub 2022 Jun 11.
Given the novel corona virus discovered in Wuhan, China, in December 2019, due to the high false-negative rate of RT-PCR and the time-consuming to obtain the results, research has proved that computed tomography (CT) has become an auxiliary One of the essential means of diagnosis and treatment of new corona virus pneumonia. Since few COVID-19 CT datasets are currently available, it is proposed to use conditional generative adversarial networks to enhance data to obtain CT datasets with more samples to reduce the risk of over fitting. In addition, a BIN residual block-based method is proposed. The improved U-Net network is used for image segmentation and then combined with multi-layer perception for classification prediction. By comparing with network models such as AlexNet and GoogleNet, it is concluded that the proposed BUF-Net network model has the best performance, reaching an accuracy rate of 93%. Using Grad-CAM technology to visualize the system's output can more intuitively illustrate the critical role of CT images in diagnosing COVID-19. Applying deep learning using the proposed techniques suggested by the above study in medical imaging can help radiologists achieve more effective diagnoses that is the main objective of the research. On the basis of the foregoing, this study proposes to employ CGAN technology to augment the restricted data set, integrate the residual block into the U-Net network, and combine multi-layer perception in order to construct new network architecture for COVID-19 detection using CT images. -19. Given the scarcity of COVID-19 CT datasets, it is proposed that conditional generative adversarial networks be used to augment data in order to obtain CT datasets with more samples and therefore lower the danger of overfitting.
鉴于2019年12月在中国武汉发现的新型冠状病毒，由于逆转录聚合酶链反应（RT-PCR）的假阴性率高且获取结果耗时，研究证明计算机断层扫描（CT）已成为新型冠状病毒肺炎诊断和治疗的辅助必要手段之一。由于目前可用的COVID-19 CT数据集很少，因此建议使用条件生成对抗网络来增强数据，以获得更多样本的CT数据集，从而降低过拟合风险。此外，还提出了一种基于BIN残差块的方法。改进的U-Net网络用于图像分割，然后与多层感知相结合进行分类预测。通过与AlexNet和GoogleNet等网络模型进行比较，得出所提出的BUF-Net网络模型性能最佳，准确率达到93%。使用Grad-CAM技术对系统输出进行可视化可以更直观地说明CT图像在诊断COVID-19中的关键作用。使用上述研究提出的技术在医学成像中应用深度学习可以帮助放射科医生实现更有效的诊断，这是该研究的主要目标。基于上述内容，本研究建议采用CGAN技术扩充受限数据集，将残差块集成到U-Net网络中，并结合多层感知，以便构建用于使用CT图像检测COVID-19的新网络架构。鉴于COVID-19 CT数据集稀缺，建议使用条件生成对抗网络来扩充数据，以获得更多样本的CT数据集，从而降低过拟合的风险。
