Chauhan Hetal, Modi Kirit
Ganpat University, Kherva, Mahesana, 384012 India.
Sankalchand Patel University, Visnagar, 384315 India.
Procedia Comput Sci. 2023;218:1394-1404. doi: 10.1016/j.procs.2023.01.118. Epub 2023 Jan 31.
A serious medical issue reported at the center of media worldwide, Since December, 2019 is the Covid19 pandemic. As declared by World Health Organization, confirmed cases of Covid19 have been 579,893,790 including 6,415,070 deaths as of 29 July 2022. Even new cases reported in last 24 hours are 20,409 in India. This needs to diagnose and timely treatment of Covid-19 is essential to prevent hurdles including death. The author developed deep learning based Covid19 diagnosis and severity prediction models using x-ray images with hope that this technology can increase access to radiology expertise in remote places where availability of expert radiologist is limited. The researchers proposed and implemented Attentive Multi Scale Feature map based deep Network (AMSF-Net) for x- ray image classification with improved accuracy. In binary classification, x-ray images are classified as normal or Covid19. Multiclass classification classifies x-ray images into mild, moderate or severe infection of Covid19. The researchers utilized lower layers features in addition to features from highest level with different scale to increase ability of CNN to learn fine-grained features. Channel attention also incorporated to amplify features of important channels. ROI based cropping and AHE employed to enhance content of training image. Image augmentation utilized to increase dataset size. To address the issue of the class imbalance problem, focal loss has been applied. Sensitivity, precision, accuracy and F1 score metrics are used for performance evaluation. The author achieved 78% accuracy for binary classification. Precision, recall and F1 score values for positive class is 85, 67 and 75, respectively while 73, 88 and 80 for negative class. Classification accuracy of mild, moderate and sever class is 90, 97 and 96. Average accuracy of 95 % achieved with superior performance compared to existing methods.
自2019年12月以来，全球媒体关注的一个严重医学问题是新冠疫情。世界卫生组织宣布，截至2022年7月29日，新冠确诊病例达579,893,790例，其中死亡6,415,070例。仅在过去24小时内，印度就报告了20,409例新增病例。因此，对新冠病毒进行诊断并及时治疗对于预防包括死亡在内的各种障碍至关重要。作者利用X光图像开发了基于深度学习的新冠病毒诊断和严重程度预测模型，希望这项技术能够在专家放射科医生数量有限的偏远地区增加获取放射学专业知识的机会。研究人员提出并实施了基于注意力多尺度特征图的深度网络（AMSF-Net）用于X光图像分类，提高了准确率。在二分类中，X光图像被分类为正常或新冠病毒感染。多分类则将X光图像分为新冠病毒轻度、中度或重度感染。研究人员除了利用最高层不同尺度的特征外，还利用了较低层的特征，以提高卷积神经网络学习细粒度特征的能力。还引入了通道注意力来增强重要通道的特征。基于感兴趣区域的裁剪和自适应直方图均衡化用于增强训练图像的内容。利用图像增强技术来增加数据集的大小。为了解决类别不平衡问题，应用了焦点损失。使用灵敏度、精度、准确率和F1分数指标进行性能评估。作者在二分类中达到了78%的准确率。阳性类别的精度、召回率和F1分数值分别为85、67和75，而阴性类别的分别为73、88和80。轻度、中度和重度类别的分类准确率分别为90%、97%和96%。与现有方法相比，平均准确率达到95%，性能优越。
