Research On COVID-19 Segmentation And Data Amplification Method Based On Deep Learning

The outbreak of novel coronavirus pneumonia occurred worldwide in early 2020,which brought severe health risks to human beings,and brought hitherto unknown pressure to the medical diagnostic systems in the world.Using the computer-aided diagnosis system based on deep learning,we can detect whether there is covid-19 characterization information from the CT image of the lung,and further give the accurate diagnosis of infection,which plays a vital role in covid-19 diagnosis.However,in covid-19 segmentation,the lung infection area is relatively scattered,has complex and irregular texture performance,and accounts for a small proportion in the whole image,which increases the difficulty of segmenting the infection area.In addition,the scarcity of data easily leads to over fitting in model training,which affects the generalization ability of the model and reduces the accuracy of segmentation.In order to improve the segmentation accuracy of covid-19 in CT image and alleviate the problem of data scarcity,this paper proposes a method to segment covid-19 infection area from chest CT image and amplify the diversity of infection area CT image by using existing data.The main contents of the two-part method proposed in this paper are as follows:(1)In the segmentation method,a hierarchical segmentation strategy and a new segmentation Network MA-Net are proposed.The hierarchical segmentation strategy is to segment the lung region from the chest CT image,and then segment the infected region from the lung region.This method can effectively avoid the interference of areas outside the lung.The new segmentation Network MA-Net,combined with multi-scale feature extraction module,context based attention mechanism and depth residual structure,can effectively learn the deep multi-scale feature information,increase the attention to the target infected area and improve the segmentation accuracy.(2)In the image generation method,a method of using two Generative Adversarial Network models to expand the chest CT image is proposed.Firstly,the diversity of the infected area is increased by expanding the segmentation label,the diversity of the generated CT image is increased by editing the fusion label,and the quality of the generated CT image is improved by using the improved generator.Experiments show that this method can effectively improve the quality of the generated image,and can expand a large number of CT images of the diversity of infection areas.This method is tested on the open covid-19 dataset.The comparative experiment proves the effectiveness of the segmentation and amplification method proposed in this paper,which can effectively improve the accuracy of covid-19 segmentation and the diversity of amplified data.There are 37 figures,5 tables and 105 references.