Research On COVID-19 Lung Infection Segmentation From CT Images Based On Multi-scale Wavelet Guidance Network

The coronavirus disease 2019(COVID-19)has seriously threatened human health since the end of 2019.Computed Tomography(CT),an important complement for the reverse transcription-polymerase chain reaction(RT-PCR),has helped clinicians diagnose COVID-19.Accurate segmentation of COVID-19 infection from CT scans provides an important basis for diagnosis and assessment of COVID-19.However,it is a challenging task because of COVID-19 infection with various textures,sizes,locations,low contrast,and blurred boundaries.To solve these problems,this paper proposes a method by combining wavelet transform and convolutional neural network(CNN)to achieve accurately COVID-19 lung infection segmentation from CT images.The main work is as follows:1)We propose a new Multi-scale Wavelet Guidance Network(MWG-Net)for COVID-19 lung infection segmentation.Wavelet transform is used to decompose CT images into sub-images containing rich spatial and frequency information,and then they are embedded into the encoder and decoder of a convolutional neural network.Moreover,multi-scale wavelet features and supervised edge information are used to jointly guide the network to extract discriminant features and focus on the infection details and edges.Our MWG-Net mainly includes three modules: Wavelet Guidance Module(WGM),Wavelet & Edge Guidance Module(WEGM),and Progressive Fusion Module(PFM).To overcome the various textures,sizes,and locations,WEG is proposed to guide the network in extracting multi-scale detailed features of infection at multiple stages of encoder.WEGM is proposed to solve the problem of blurring edges.It uses multi-scale wavelet features and edges features to guide the decoder to generate a mask with complete details and edges.Besides,PFM is proposed to fully integrate and utilize multi-scale context information in different stages of encoder and decoder.2)We establish an available dataset for COVID-19 lung infection segmentation(named COVID-Seg-100),which contains more than 5,800 annotated CT slices from100 cases.After multiple CT scans,each case contains CT images of different stages and degrees of infection.Therefore,COVID-Seg-100 with complex and diverse clinical features poses a great challenge to segmentation algorithms.This dataset will be available online for all researchers.3)We conduct extensive experiments on COVID-Seg-100 to verify the validity of our MWG-Net.Moreover,comparative experiments between our MWG-Net and several advanced methods are conducted on two public datasets with different characteristics.The results show that comprehensive performance of our MWG-Net is superior to other methods,which can pay more attention to the details and boundaries of COVID-19 infection and achieve accurate COVID-19 infection segmentation from complex and varied CT images.