A COVID-19 Lesion Segmentation Algorithm Based On Transfer Learning

COVID-19 epidemic,which occurred in late 2019,had a profound impact on all over the world,thousands of people are infected with the virus,all countries in the world has taken a different strategy to deal with the outbreak of: keep social functioning,trade flow,the transformation of industrial production,the enterprise bankruptcy lead by demand reducing,unemployment,etc.,but one of the most important task is the treatment of the patients of COVID-19.Lung CT scan has become an important information basis for the diagnosis of COVID-19 patients.How to effectively analyze the images of the patients’ lung CT scan and segment the lesions is of great significance to the judgment of the patient’s condition.Recently segmentation algorithm based on deep learning in numerous segmentation task performance,but depend on a mass of labeling,high quality data for training,however,the high cost of medical image has marked,it is difficult to get the characteristics of labeled data,which restrict the deep learning applied in medical image analysis,and the outbreak of a sudden,is exacerbated by the phenomenon of lack of labeled data,so exploring how to in the absence of labeled data,a small amount of labeled training data circumstances be able to assist doctors diagnosed image segmentation model become a urgent problem.This paper first introduces the basic knowledge and research progress of transfer learning and image segmentation,and further introduces the main methods combining them.This paper presents a algorithm of image segmentation based on tranfer learning,use of already labeled CT data sets of pcommunity-acquired pneumonia to assist in training COVID-19 lesions segmentation model,the model is an adversarial network,using two separate domain discrimination in focus on the positive and negative,prompting segmentor can extract the feature of invariance of domain,and implements the domain tranfer.At the same time,inspired by Adapt Seg,we propose an algorithm to perform domain adaptation in the output space,and solve the problem that AdaptSeg can’t deal with the sample imbalance on medical images,which is easy to cause the pattern collapse.The effectiveness of the transfer segmentation model proposed in this paper is proved through experiments.Moreover,for further exploring the applicable scope of transfer learning,we confirms that transfer learning is more effective in few-shot learning and reveals the inherent correlation between transfer learning and meta-learning.