Semi-supervised　medical　image　segmentation　(SS-MIS)　has　gained　growing　interest　for　its　ability　to　mitigate　costly　annotation.　However,　existing　solutions　struggle　in　this　field,　primarily　for　neglecting　challenging　boundary　regions　and　the　similarity　between　target　structures.　These　issues　result　in　imprecise　boundary　segmentation　and　unreasonable　predictions.　To　this　end,　this　paper　presents　a　novel　SS-MIS　framework,　integrating　Boundary-aware　Multi-Task　(BMT)　strategy　and　Dynamic　Competitive　Contrastive　Learning　(DCCL).　BMT　employs　a　boundary-aware　multi-task　strategy　to　focus　the　model　on　boundary　regions,　and　the　extracted　boundary　features　are　further　integrated　with　the　segmentation　features　to　achieve　more　precise　predictions.　Additionally,　to　promote　compact　distribution　for　identical　classes　in　the　feature　space,　DCCL　adopts　a　dynamic　competition　strategy　to　generate　more　reliable　feature　prototypes.　The　model　then　performs　contrastive　learning　by　minimizing　the　distance　between　its　features　and　the　corresponding　feature　prototypes.　This　strategy　further　enhances　the　model＇s　ability　to　discriminate　between　different　classes.　Extensive　experiments　on　three　public　datasets,　i.e.,　ACDC,　PROMISE12,　and　BUSI,　demonstrate　that　our　method　achieves　promising　results,　particularly　regarding　boundary　regions　and　class　discrimination.　Specifically,　with　only　10%　labeled　data,　we　achieved　the　Dice　scores　of　87.89%,　74.92%,　and　65.68%　on　ACDC,　PROMISE12,　and　BUSI,　respectively.　These　results　notably　outperform　the　comparative　CPS　by　2.36%,　18.64%,　and　5.48%,　respectively.　The　code　is　publicly　available　at:　https://github.com/linzk99/BMT-DCCL.　©　2025　Elsevier　Ltd