Coronavirus Disease 2019 requires chest computed tomography (CT) imaging data for early diagnosis, treatment, and prognosis (COVID-19). Artificial intelligence has been triedto assist physicians in enhancing the diagnosisaccuracy and operating efficiency of CT. Existing supervised techniques on CT images of COVID- 19 pneumonia require voxel-based annotations for training, which takes a long time and effort. This research developed a weakly-supervised technique for COVID-19 lesion localization based on generative adversarial networks (GAN) using just image-level labels. We originally presented a GAN-based framework for generating normal- looking CT slices from CT scans with COVID-19 lesions. We then devised a unique feature matching technique to enhance the realism of produced pictures by directing the generator tocapture the intricate texture of chest CT images. Finally, by subtracting the output picture from its matching input image, the localization map of lesions may be simply generated. We can increasethe classification accuracy of our diagnostic system by adding a classifier branch to the GAN- based architecture to identify localization maps.In this paper, three CT datasets regarding COVID-19 were obtained for examination from hospitals in Sao Paulo, the Italian Society of Medical and Interventional Radiology, and China Medical University. Our method of weakly supervised learning yielded AUC of 0.883, dice coefficient of 0.575, accuracy of 0.884, sensitivityof 0.647, specificity of 0.929, and F1-score of0.640 significantly outperformed other frequently used weakly supervised object localization algorithms. In addition, we compared the suggested technique to fully supervised learning methods in the COVID-19 lesion segmentationproblem, and the proposed weakly supervised method still produces a competitive performance with a dice coefficient of 0.575. Furthermore, we examined the relationship between illness severityand visual score and discovered that the common severity cohort had the largest sample size as well as the highest visual score, implying that our method can aid in the rapid diagnosis of COVID- 19 patients, particularly in the massive common severity cohort .Finally, we argued that this uniquetechnique may be used as an accurate andefficient tool to remove the bottleneck of expert annotation costs and promote the advancement of computer-aided COVID-19 diagnosis.

Keywords : Coronavirus Disease 2019, Generative Adversarial Network, Lesion Location, and Lesion Segmentation