Nowadays,　the　pneumonia　has　widely　spread　around　the　world.　CT　images,　as　a　mostly　used　means　to　detect　pneumonia,　take　a　long　time　to　read　manually　by　radiologists.　Instead,　automatic　medical　image　segmentation　technology　can　segment　the　lesion　in　a　few　seconds.　Considering　the　difficulties　in　segmentation　of　pneumonia　lesion,　such　as　the　extreme　variations　in　the　morphology　of　the　lesions　with　noisy　background,　in　this　paper,　we　propose　an　accurate　and　efficient　lesion　segmentation　algorithm,　Adversarial　Residual　U2Net　(ARes-U2Net)　for　COVID-19　infection　segmentation.　The　model　is　comprised　of　two　parts:　segmented　network　(S)　and　critic　network　(C),　which　are　trained　alternately　with　minimizing　and　maximizing　a　multi-scale　L1　function　respectively.　The　segmented　network　(S)　has　U-shaped　structure　with　residual　connections,　which　extracts　the　multi-scale　feature　information.　The　critic　network　(C)　can　evaluate　the　multi-scale　output　of　the　segmented　network　by　FCN　layers.　In　the　experiments,　we　designed　intra-dataset　and　cross-dataset　scenario　experiments　on　a　new　dataset　and　four　public　datasets.　The　Dice　Score,　IoU,　spec　and　prec　of　our　proposed　model　are　64.99%,　54.59%,　99.77%　and　76.73%　on　the　constructed　dataset,　which　significantly　outperform　other　compared　methods.　The　consistent　results　can　also　be　derived　for　cross-data　scenarios.　©　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.　2024.