Research On Lesion Segmentation Method Of COVID-19 In CT Images

Since the outbreak of COVID-19 at the end of 2019,hundreds of thousands of new cases have been diagnosed confirmed globally every day.Computed tomography(CT)examination is a highly sensitive and highly efficient COVID-19 screening method.The segmentation of the lesion area is an important step when the CT examination is applied to the diagnosis and monitoring of COVID-19.However,due to the lack of experienced radiologists,it is difficult to segment CT images in large scale.Therefore,it is necessary to develop an efficient COVID-19 lesion segmentation method.The current research on COVID-19 lesion segmentation method prefers CNN-based methods.However,conventional CNN model usually has many parameters,its generalization ability is limited by the number of training samples.the UNet model is commonly used in medical image segmentation tasks,however,when applied to COVID-19 lesion segmentation in CT images,there are many redundant features,resulting in a large amount of redundant calculation,and UNet model has insufficient feature receptive field size,thus cannot extract the global information of the input image well.Therefore,based on UNet,L-UNet-GCCA model is proposed as a lightweight and efficient model for COVID-19 lesion segmentation in CT images.In order to meet the requirement of prediction speed for the model,a redundant feature analysis method is proposed.With the result of redundant feature analysis as guide,the improved L-UNet model is designed,which reduces redundant features and redundant calculations.For the problem of insufficient model feature receptive field size,a global information fusion module GCCA is designed,which is based on attention mechanism,with improved global information fusion process to adapt to models with fewer redundant features.In addition,aiming at the problem of insufficient labeled data,a data augmentation method is designed based on the gray-scale perturbation on the characteristics of CT images to further improve the accuracy of lesion segmentation.The L-UNet-GCCA model trained in a fully supervised manner.When the model is used for prediction,it uses the two-dimensional CT cross-sectional image as input,and outputs the segmentation result of the lesion end-to-end.The parameter amount of the L-UNet-GCCA model is 0.51 M.When the input image size is 512×512 pixels,the calculation amount of the model is 8.84 GFlops,which reduces the parameters and calculations to 1/38 and 1/18 comparing to the original UNet model.For the model tested on the COVID-19-P20 dataset,the SEN is 0.8245,the SPC is 0.9972 and The DSC is 0.8182;and for the model tested on the COVID-19-P1110 dataset,the SEN is 0.7207,the SPC is 0.9979 and the DSC is 0.6811.Compared with other models and methods,the proposed method has a greater advantage in prediction accuracy and prediction speed.