U-Net　is　a　classic　architecture　for　semantic　segmentation.　However,　it　has　several　limitations,　such　as　difficulty　in　capturing　complex　images　detail　due　to　its　simple　U　structure,　long　convergence　time　arising　from　fixed　network　parameters,　and　suboptimal　efficacy　in　decoding　and　restoring　multi-scale　information.　To　deal　with　the　above　issues,　we　propose　a　Multiple　U-shaped　network　(Multi-UNet)　assuming　that　constructing　appropriate　U-shaped　structure　can　achieve　better　segmentation　performance.　Firstly,　inspired　by　the　concept　of　connecting　multiple　similar　blocks,　our　Multi-UNet　consists　of　multiple　U-block　modules,　with　each　succeeding　module　directly　connected　to　the　previous　one　to　facilitate　data　transmission　between　different　U　structures.　We　refer　to　the　original　bridge　connections　of　U-Net　as　Intra-U　connections　and　introduce　a　new　type　of　connection　called　Inter-U　connections.　These　Inter-U　connections　aim　to　retain　as　much　detailed　information　as　possible,　enabling　effective　detection　of　complex　images.　Secondly,　while　maintaining　Mean　Intersection　over　Union　(Mean-IoU),　the　up-sampling　of　each　U　applies　uniformly　small　channel　values　to　reduce　the　number　of　model　parameters.　Thirdly,　a　Spatial-Channel　Parallel　Attention　Fusion　(SCPAF)　module　is　designed　at　the　initial　layer　of　every　subsampling　module　of　U-block　architecture.　It　enhances　feature　extraction　and　alleviate　computational　overhead　associated　with　data　transmission.　Finally,　we　replace　the　final　up-sampling　module　with　Atrous　Spatial　Pyramid　Pooling　Head　(ASPPHead)　to　accomplish　seamless　multi-scale　feature　extraction.　Our　experiments　are　compared　and　analyzed　with　advanced　models　on　three　public　datasets,　and　it　can　be　concluded　that　the　universality　and　accuracy　of　Multi-UNet　network　are　superior.　©　2024