Semi-Supervised Segmentation of COVID-19 Infection on CT-Scan Lung Using Dynamic Mutual Training

Corona virus disease 2019 (COVID-19) 's global pandemic has caused the world to face a health crisis. Automated detection of COVID-19 infection from computed tomography (CT-scan) images has improved healthcare for treating COVID-19. However, segmentation of infected areas on CT-scan images of the lungs faces several challenges: detailed infection characteristics and low contrast differences between CT scans of infected lungs. It has a low data scale with a doctor's statement because it is still a new case, with a lot of data with pseudo labels, while pseudo labels have a low confidence level and a high error rate. Therefore, using the data of 1600 pseudo label images and 50 doctor label images, we apply pseudo supervision as the core idea, mutual training between two different models with a dynamic loss function called dynamic mutual training (DMT). DMT will do multi-training on pseudo labels with doctor's labels to be trusted in area segmentation. The results obtained are the most superior value of 91.32% with a loss value of 0.19 dice score 0.23, IOU 0.781, precision 0.843, sensitivity 0.753, and specificity 0.845. We also compare our method with other segmentation methods such as UNET, which is highly preferred in terms of medical images, and mask RCNN, which shows the best method in terms of segmentation. This comparison indicates that DMT provides the best experimental incentive with a dice score value of 2-30%, superior to cases segmentation areas affected by COVID-19 on CT scans of the lungs.