The　effective　segmentation　of　lesion　regions　in　gastric　cancer　images　provides　a　reliable　complementary　basis　for　clinical　analysis　and　diagnosis　of　gastric　cancer.　Deep　learning　algorithms　for　medical　image　segmentation　effectively　learn　high-level　abstract　features　of　the　lesion　region,　resulting　in　good　segmentation　results.　However,　traditional　deep　learning　algorithms　cannot　meet　the　demand　due　to　the　inherent　characteristics　of　blurred　lesion　boundaries　and　lesion　diversity　in　gastric　cancer　images.　Therefore,　this　paper　proposes　a　U-shaped　network　for　lesion　region　segmentation　of　gastric　cancer　images　based　on　depthwise　separable　convolution　and　an　improved　ASPP　module　for　multi-scale　feature　extraction.　The　improved　ASPP　module　(I-ASPP)　concatenates　the　dilated　convolution　with　different　dilation　rates　to　extract　multi-scale　feature　information　and　introduces　a　boundary　feature　calculation　branch　to　improve　the　loss　of　edge　information　due　to　multiple　down-sampling　and　up-sampling　operations　in　U-Net.　Besides,　depthwise　separable　convolution　is　used　to　replace　the　original　convolutional　structure　to　reduce　the　model＇s　complexity.　Experimental　results　show　that　our　proposed　method　has　an　IOU　of　74.0%,　a　Dice　of　84.9%,　and　an　Accuracy　of　91.3%.　These　metrics　demonstrate　that　our　model　yields　high　quality　segmentation　results　that　meet　the　needs　of　clinical　analysis　and　diagnosis.　©　2023　IEEE.