{"id":1068,"date":"2026-05-19T08:35:33","date_gmt":"2026-05-19T00:35:33","guid":{"rendered":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/19\/2026-05-19-%e5%8c%bb%e5%ad%a6%e5%9b%be%e5%83%8f%e5%88%86%e5%89%b2%e8%ae%ba%e6%96%87%e7%b2%be%e8%af%bb%ef%bc%9asemi-medref-%e4%b8%8e-anatomical-shape-priors\/"},"modified":"2026-05-19T08:35:33","modified_gmt":"2026-05-19T00:35:33","slug":"2026-05-19-%e5%8c%bb%e5%ad%a6%e5%9b%be%e5%83%8f%e5%88%86%e5%89%b2%e8%ae%ba%e6%96%87%e7%b2%be%e8%af%bb%ef%bc%9asemi-medref-%e4%b8%8e-anatomical-shape-priors","status":"publish","type":"post","link":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/19\/2026-05-19-%e5%8c%bb%e5%ad%a6%e5%9b%be%e5%83%8f%e5%88%86%e5%89%b2%e8%ae%ba%e6%96%87%e7%b2%be%e8%af%bb%ef%bc%9asemi-medref-%e4%b8%8e-anatomical-shape-priors\/","title":{"rendered":"2026-05-19 \u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u7cbe\u8bfb\uff1aSemi-MedRef \u4e0e Anatomical Shape Priors"},"content":{"rendered":"<h1>\u4eca\u65e5\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6700\u65b0\u8bba\u6587\u7cbe\u8bfb\u8ffd\u8e2a<\/h1>\n<h2>\u4eca\u65e5\u7ed3\u8bba<\/h2>\n<p>\u4eca\u5929\u6ca1\u6709\u68c0\u7d22\u5230\u660e\u786e\u5df2\u63a5\u6536 MICCAI\/CVPR\/ICCV\/ECCV\/NeurIPS\/ICLR\/MedIA\/TMI \u7b49\u9876\u4f1a\/\u9876\u520a\u7684\u5168\u65b0\u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\uff1b\u53bb\u91cd\u540e\uff0c\u6700\u503c\u5f97\u5173\u6ce8\u7684\u662f\u4e24\u7bc7 2026-05-15 arXiv \u65b0\u7a3f\uff1a<strong>Semi-MedRef<\/strong> \u4e0e <strong>Evaluation of Anatomical Shape Priors<\/strong>\u3002\u524d\u8005\u4ee3\u8868\u201c\u533b\u5b66 referring segmentation + \u534a\u76d1\u7763 + \u56fe\u6587\u5bf9\u9f50\u201d\u7684\u65b0\u65b9\u5411\uff0c\u540e\u8005\u662f\u4e00\u7bc7\u6709\u4ef7\u503c\u7684\u8d1f\u7ed3\u679c\u7814\u7a76\uff0c\u63d0\u9192\u6211\u4eec\u5728\u5f3a 3D U-Net \u57fa\u7ebf\u4e0a\u7b80\u5355\u52a0\u5165\u624b\u5de5\u5f62\u72b6\u5148\u9a8c\u672a\u5fc5\u6709\u6548\u3002<\/p>\n<h2>\u68c0\u7d22\u8bf4\u660e<\/h2>\n<p>\u4eca\u65e5\u68c0\u7d22\u8303\u56f4\u8986\u76d6 arXiv \u6700\u65b0\u63d0\u4ea4\u3001\u533b\u5b66\u56fe\u50cf\u5206\u5272\/medical image segmentation\u3001polyp segmentation\u30013D medical image segmentation\u3001U-Net\/nnU-Net\u3001Mamba\u3001SAM\/foundation model\u3001referring medical segmentation \u7b49\u5173\u952e\u8bcd\uff0c\u5e76\u91cd\u70b9\u67e5\u770b 2026-05-15 \u4ee5\u540e\u65b0\u589e\u5019\u9009\uff1b\u672a\u53d1\u73b0\u5f53\u5929\u53ef\u786e\u8ba4\u7684\u9876\u4f1a\/\u9876\u520a\u6b63\u5f0f\u63a5\u6536\u7248\u672c\uff0c\u56e0\u6b64\u4ece\u6700\u65b0\u4e14 PDF \u53ef\u83b7\u53d6\u7684 arXiv preprint \u4e2d\u7b5b\u9009\u3002\u6240\u6709\u5165\u9009\u8bba\u6587\u5747\u4e3a 2025 \u5e74\u53ca\u4ee5\u540e\u3002\u5df2\u68c0\u67e5\u5386\u53f2\u63a8\u8350\u8bb0\u5f55\u5e76\u6392\u9664\u4e86\u91cd\u590d\u8bba\u6587\uff1b\u672c\u6b21\u8df3\u8fc7\u7684\u91cd\u590d\u5019\u9009\u5305\u62ec Med-DisSeg\u3001SpectraFlow\u3001SplitFed-CL\u3001DuetFair\u3001CMFDNet\u3001Topo-VM-UNetV2 \u7b49\u3002<\/p>\n<h2>WordPress \u53d1\u5e03<\/h2>\n<ul>\n<li>WordPress \u6587\u7ae0\u94fe\u63a5\uff1a\u53d1\u5e03\u540e\u89c1\u672c\u6587\u5143\u4fe1\u606f<\/li>\n<li>WordPress Post ID\uff1a\u53d1\u5e03\u540e\u89c1\u672c\u6587\u5143\u4fe1\u606f<\/li>\n<\/ul>\n<hr \/>\n<h2>\u8bba\u6587 1\uff1aSemi-MedRef: Semi-Supervised Medical Referring Image Segmentation with Cross-Modal Alignment<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aSemi-MedRef: Semi-Supervised Medical Referring Image Segmentation with Cross-Modal Alignment<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aYuchen Li, Zhen Zhao, Yi Liu, Luping Zhou \/ Yuchen Li<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-15 arXiv v1<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint\uff0carXiv:2605.15720<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.15720<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1a\/tmp\/medseg_daily_2026-05-19\/semi_medref_2605.15720.pdf\uff1bhttps:\/\/arxiv.org\/pdf\/2605.15720<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u672a\u83b7\u53d6\uff1barXiv \u9875\u9762\u548c PDF \u6b63\u6587\u4e2d\u672a\u786e\u8ba4\u5b98\u65b9 GitHub<\/li>\n<li>\u4efb\u52a1\uff1aMedical Referring Image Segmentation\uff08MRIS\uff09\uff0c\u5373\u7528\u533b\u5b66\u56fe\u50cf + \u6587\u672c\u4f4d\u7f6e\/\u75c5\u7076\u63cf\u8ff0\u5171\u540c\u6307\u5b9a\u5206\u5272\u76ee\u6807\uff1b\u534a\u76d1\u7763\u4f4e\u6807\u6ce8\u8bbe\u7f6e<\/li>\n<li>\u6570\u636e\u96c6\uff1aQaTa-COV19\u3001MosMedData+<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1ateacher-student semi-supervised segmentation\uff1bcross-modal augmentation\uff1bimage-text contrastive learning\uff1bU-Net\/ConvNeXt + CXR-BERT \u591a\u6a21\u6001\u5206\u5272\u6846\u67b6<\/li>\n<\/ul>\n<h3>paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n<h4>1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n<p>Semi-MedRef \u7684\u6838\u5fc3\u4ef7\u503c\u4e0d\u5728\u63d0\u51fa\u4e00\u4e2a\u5168\u65b0 U-Net backbone\uff0c\u800c\u5728\u4e8e\u628a\u534a\u76d1\u7763\u533b\u5b66\u6587\u672c\u5f15\u5bfc\u5206\u5272\u4e2d\u6700\u5bb9\u6613\u51fa\u9519\u7684\u201c\u5f3a\u589e\u5f3a\u7834\u574f\u56fe\u6587\u7a7a\u95f4\u5bf9\u9f50\u201d\u95ee\u9898\u62c6\u6210\u53ef\u64cd\u4f5c\u7684 T-PatchMix\u3001PosAug \u548c ITCL \u4e09\u4e2a\u673a\u5236\uff0c\u9002\u5408\u5173\u6ce8\u5c11\u6807\u6ce8\u3001\u6587\u672c\u63d0\u793a\u548c\u533b\u5b66 foundation\/referring segmentation \u7684\u7814\u7a76\u8005\u7cbe\u8bfb\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76\u7684\u662f <strong>semi-supervised medical referring image segmentation<\/strong>\uff1a\u8f93\u5165\u4e0d\u662f\u5355\u7eaf\u56fe\u50cf\uff0c\u800c\u662f\u56fe\u50cf <code>I<\/code> \u4e0e\u6587\u672c\u63cf\u8ff0 <code>T<\/code>\uff0c\u6a21\u578b\u8f93\u51fa\u4e8c\u503c mask <code>Y<\/code>\u3002\u8fd9\u79cd\u4efb\u52a1\u5728\u80f8\u7247\/CT \u611f\u67d3\u533a\u57df\u3001\u5e26\u4f4d\u7f6e\u63cf\u8ff0\u7684\u75c5\u7076\u5206\u5272\u4e2d\u5f88\u6709\u610f\u4e49\uff0c\u56e0\u4e3a\u7eaf\u89c6\u89c9\u5206\u5272\u6709\u65f6\u4e0d\u77e5\u9053\u201c\u8981\u5206\u54ea\u4e00\u4e2a\u533a\u57df\u201d\uff0c\u800c\u6587\u672c\u53ef\u7ed9\u51fa upper\/lower\u3001left\/right\u3001lesion site \u7b49\u5b9a\u4f4d\u7ebf\u7d22\u3002<\/p>\n<p>\u4f5c\u8005\u628a\u95ee\u9898\u8bbe\u4e3a\u5c11\u91cf\u6709\u6807\u6ce8\u56fe\u6587-mask \u4e09\u5143\u7ec4 <code>Dl<\/code> \u4e0e\u5927\u91cf\u65e0 mask \u7684\u56fe\u6587\u5bf9 <code>Du<\/code>\u3002\u6a21\u578b <code>f\u03b8(I,T)<\/code> \u7ecf\u8fc7 sigmoid \u5f97\u5230\u50cf\u7d20\u6982\u7387\u56fe <code>P \u2208 [0,1]^{H\u00d7W}<\/code>\u3002\u96be\u70b9\u662f\uff1a\u534a\u76d1\u7763\u5b66\u4e60\u4f9d\u8d56 teacher-student\u3001\u4f2a\u6807\u7b7e\u548c\u5f3a\u589e\u5f3a\uff1b\u4f46 referring segmentation \u4e2d\u56fe\u50cf\u589e\u5f3a\u4e0e\u6587\u672c\u589e\u5f3a\u82e5\u4e0d\u540c\u6b65\uff0c\u4f1a\u7834\u574f\u201c\u6587\u672c\u6307\u4ee3\u533a\u57df\u201d\u548c\u201c\u56fe\u50cf\u4e2d\u5b9e\u9645\u533a\u57df\u201d\u7684\u5bf9\u5e94\u5173\u7cfb\uff0c\u5c24\u5176\u533b\u5b66\u6587\u672c\u5e38\u4f9d\u8d56 laterality \u4e0e\u89e3\u5256\u4f4d\u7f6e\uff0c\u4f8b\u5982 left\/right\u3001upper\/middle\/lower\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u6307\u51fa\u73b0\u6709\u65b9\u6cd5\u4e3b\u8981\u6709\u4e09\u7c7b\u4e0d\u8db3\uff1a<\/p>\n<ol>\n<li>\u81ea\u7136\u56fe\u50cf RES \u534a\u76d1\u7763\u65b9\u6cd5\uff08\u5982 RESMatch\/SemiRES\uff09\u53ef\u505a weak-to-strong consistency\uff0c\u4f46\u901a\u5e38\u53ea\u5904\u7406\u7b80\u5355 flip \u4e0b left\/right \u66ff\u6362\uff0c\u5bf9\u533b\u5b66\u4f4d\u7f6e\u8bed\u8a00\u4e0d\u8db3\u3002<\/li>\n<li>\u533b\u5b66\u6587\u672c\u63d0\u793a\u5206\u5272\u65b9\u6cd5\uff08\u5982 LViT\u3001Textmatch\u3001GuideDecoder\u3001MMI-UNet\uff09\u5f3a\u8c03\u89c6\u89c9-\u6587\u672c\u878d\u5408\uff0c\u4f46\u5728\u5f3a\u589e\u5f3a\u3001\u4f4e\u6807\u6ce8\u3001\u65e0\u6807\u7b7e\u4f2a\u76d1\u7763\u4e0b\u5982\u4f55\u4fdd\u6301 cross-modal alignment \u4ecd\u4e0d\u5145\u5206\u3002<\/li>\n<li>CutMix\/PatchMix \u5728\u5355\u6a21\u6001 SSL \u4e2d\u6709\u6548\uff0c\u4f46\u76f4\u63a5\u7528\u4e8e\u56fe\u6587\u533b\u5b66\u5206\u5272\u65f6\u53ef\u80fd\u628a\u56fe\u50cf patch \u6362\u6389\u5374\u4e0d\u66f4\u65b0\u6587\u672c\uff0c\u5bfc\u81f4\u201c\u6587\u672c\u8bf4\u5de6\u4e0b\u80ba\uff0c\u56fe\u50cf patch \u6765\u81ea\u53f3\u4e0a\u533a\u57df\u201d\u7684\u4f2a\u76d1\u7763\u6c61\u67d3\u3002<\/li>\n<\/ol>\n<p>\u56e0\u6b64\uff0c\u8bba\u6587\u771f\u6b63\u8981\u89e3\u51b3\u7684\u4e0d\u662f segmentation decoder \u672c\u8eab\uff0c\u800c\u662f <strong>\u534a\u76d1\u7763 MRIS \u4e2d\u5f3a\u6270\u52a8\u4e0e\u56fe\u6587\u7a7a\u95f4\u8bed\u4e49\u4e00\u81f4\u6027\u7684\u51b2\u7a81<\/strong>\u3002<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>Semi-MedRef \u662f\u4e00\u4e2a\u4e24\u9636\u6bb5 teacher-student \u6846\u67b6\uff1a<\/p>\n<ul>\n<li><strong>Burn-in \u9636\u6bb5<\/strong>\uff1a\u53ea\u7528\u6709\u6807\u6ce8\u6570\u636e\u8bad\u7ec3 student\uff0c\u76d1\u7763\u635f\u5931\u4e3a Dice + BCE\uff0c\u5373 <code>Lsup = LDiceCE(\u03c3(f\u03b8S(I,T)), Y)<\/code>\u3002<\/li>\n<li><strong>Semi-supervised \u9636\u6bb5<\/strong>\uff1ateacher \u7528 EMA \u66f4\u65b0\uff0c<code>\u03b8T \u2190 m \u03b8T + (1-m)\u03b8S<\/code>\uff0c\u5176\u4e2d <code>m=0.999<\/code>\u3002teacher \u5728\u5f31\u589e\u5f3a\u89c6\u56fe\u4e0a\u751f\u6210\u6982\u7387\u56fe\u5e76\u9608\u503c\u5316\u5f97\u5230\u786c\u4f2a mask\uff1bstudent \u5728\u5f3a\u589e\u5f3a\u56fe\u6587\u89c6\u56fe\u4e0a\u5b66\u4e60\u5339\u914d\u4f2a mask\u3002<\/li>\n<li><strong>\u603b\u4f53\u635f\u5931<\/strong>\uff1a<code>L = Lsup + \u03bbu Lunsup + \u03bbsup_itcl Lsup_itcl + \u03bbunsup_itcl Lunsup_itcl<\/code>\u3002<\/li>\n<\/ul>\n<p>\u6846\u67b6\u4e2d\u7684\u4e09\u4e2a\u5173\u952e\u90e8\u4ef6\u662f\uff1a<\/p>\n<ol>\n<li><strong>PosAug<\/strong>\uff1a\u5bf9\u6587\u672c\u4e2d\u7684\u4f4d\u7f6e\u77ed\u8bed\u505a dropout \u6216 fuzzing\uff0c\u907f\u514d\u6a21\u578b\u8fc7\u62df\u5408\u7cbe\u786e\u4f4d\u7f6e\u8bcd\u3002<\/li>\n<li><strong>T-PatchMix<\/strong>\uff1a\u628a PatchMix\/CutMix \u6539\u9020\u6210\u56fe\u6587\u540c\u6b65\u589e\u5f3a\uff0c\u53ea\u5728\u4f4d\u7f6e\u517c\u5bb9\u6216 teacher \u9ad8\u7f6e\u4fe1\u75c5\u7076\u533a\u57df\u6df7\u5408 patch\uff0c\u5e76\u540c\u6b65\u66f4\u65b0\u6587\u672c\u4f4d\u7f6e span\u3002<\/li>\n<li><strong>ITCL<\/strong>\uff1a\u57fa\u4e8e coarse position pseudo-label \u6784\u9020 soft positives \u7684\u56fe\u6587\u5bf9\u6bd4\u5b66\u4e60\uff0c\u4e0d\u518d\u50cf CLIP \u90a3\u6837\u628a batch \u4e2d\u6240\u6709\u5176\u4ed6\u6587\u672c\u90fd\u5f53\u786c\u8d1f\u6837\u672c\u3002<\/li>\n<\/ol>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 A\uff1aPosAug\uff08Position-aware Text Augmentation\uff09<\/strong><br \/>\n\u8f93\u5165\u662f student \u5206\u652f\u6587\u672c\u4e2d\u7684\u4f4d\u7f6e\u77ed\u8bed\uff0c\u4f8b\u5982 \u201cupper left lung\u201d\u3002\u8f93\u51fa\u662f\u88ab\u6270\u52a8\u540e\u7684\u6587\u672c\uff1a\u4e00\u79cd\u662f\u7528 <code>[UNK_POS]<\/code> \u66ff\u6362\u4f4d\u7f6e\u77ed\u8bed\uff0c\u53e6\u4e00\u79cd\u662f\u628a\u5177\u4f53\u4f4d\u7f6e fuzz \u6210 \u201ca region of the lung\u201d \u8fd9\u7c7b\u5f31\u4f4d\u7f6e\u8868\u8ff0\u3002\u5176\u4f5c\u7528\u662f\u8ba9\u6a21\u578b\u5728\u6587\u672c\u4f4d\u7f6e\u4e0d\u5b8c\u6574\u6216\u4e0d\u7cbe\u786e\u65f6\u4ecd\u4f9d\u8d56\u56fe\u50cf\u8bc1\u636e\u3002\u5b83\u7684\u521b\u65b0\u6027\u4e2d\u7b49\uff0c\u4f46\u5bf9\u533b\u5b66 referring segmentation \u5f88\u5b9e\u7528\uff1b\u8fc1\u79fb\u5230 polyp segmentation \u65f6\u4ef7\u503c\u53d6\u51b3\u4e8e\u662f\u5426\u6709\u6587\u672c\u6307\u4ee3\uff0c\u5982\u679c\u53ea\u662f\u6807\u51c6\u5168\u81ea\u52a8\u606f\u8089\u5206\u5272\u5219\u7528\u5904\u6709\u9650\u3002<\/p>\n<p><strong>\u6a21\u5757 B\uff1aT-PatchMix\uff08Image-Text Alignment-preserving Patch Mixing\uff09<\/strong><br \/>\n\u8f93\u5165\u4e3a\u4e24\u4e2a\u65e0\u6807\u7b7e\u56fe\u6587\u5bf9 <code>(Ii,Ti)<\/code>\u3001<code>(Ij,Tj)<\/code> \u548c block-wise mask <code>Mi<\/code>\u3002\u56fe\u50cf\u6df7\u5408\u4e3a <code>Imix=(1-Mi)\u2299Ii + Mi\u2299Ij<\/code>\u3002\u4f5c\u8005\u63d0\u4f9b\u4e24\u79cd\u91c7\u6837\u89c4\u5219\uff1a\u4f4d\u7f6e\u7ea6\u675f\u6df7\u5408\u4f1a\u628a\u6587\u672c\u4f4d\u7f6e\u6620\u5c04\u5230 left\/right \u00d7 upper\/middle\/lower \u7684\u7c97\u533a\u57df\uff0c\u53ea\u5728\u517c\u5bb9\u533a\u57df\u6df7\u5408\uff1b\u6982\u7387\u9a71\u52a8\u6df7\u5408\u4ece teacher \u9ad8\u7f6e\u4fe1\u6982\u7387\u56fe\u4e2d\u9009 patch\uff0c\u5e76\u7528 lesion-gating ratio <code>ri = sum(Pj^T\u2299Mi)\/(sum Mi + \u03b5)<\/code> \u8fc7\u6ee4\u975e\u75c5\u7076 patch\u3002\u4f2a\u6807\u7b7e\u4e5f\u540c\u6b65\u6df7\u5408 teacher probability\uff0c\u5e76\u5728\u6df7\u5408\u6709\u6548\u65f6\u66ff\u6362\u6216\u5408\u5e76\u6587\u672c\u4f4d\u7f6e span\u3002<\/p>\n<p>\u8fd9\u4e2a\u6a21\u5757\u662f\u771f\u6b63\u7684\u673a\u5236\u6838\u5fc3\uff1a\u5b83\u628a\u5355\u6a21\u6001 CutMix \u8f6c\u6362\u6210\u201c\u56fe\u50cf patch\u3001\u4f2a mask\u3001\u4f4d\u7f6e\u6587\u672c\u201d\u4e09\u8005\u540c\u6b65\u7684\u589e\u5f3a\u3002\u5bf9 3D medical segmentation \u53ef\u8fc1\u79fb\u4e3a anatomy-aware crop\/mix\uff0c\u4f46\u9700\u8981 3D \u4f4d\u7f6e\u8bcd\u6216\u5668\u5b98 atlas\uff1b\u5bf9 DAMamba\/U-Net \u4e3b\u5e72\u6539\u9020\u672c\u8eab\u4e0d\u662f backbone \u6a21\u5757\uff0c\u800c\u66f4\u50cf\u8bad\u7ec3\u7b56\u7565\u3002<\/p>\n<p><strong>\u6a21\u5757 C\uff1aITCL\uff08Position-guided Image-Text Contrastive Learning\uff09<\/strong><br \/>\n\u6a21\u578b\u4ece\u878d\u5408\u7279\u5f81\u56fe <code>F<\/code> \u5f97\u5230\u56fe\u50cf embedding <code>v=ProjI(GAP(F))<\/code>\uff0c\u4ece\u6587\u672c\u5f97\u5230 embedding <code>u=ProjT(T)<\/code>\uff0c\u76f8\u4f3c\u5ea6\u4e3a <code>Sij=v_i^T u_j\/\u03c4<\/code>\u3002\u5173\u952e\u662f\u4f5c\u8005\u7528\u4f4d\u7f6e pseudo-label <code>q_i\u2208{0,1}^6<\/code> \u8868\u793a left\/right \u00d7 upper\/middle\/lower\uff0c\u4ee5 Jaccard affinity <code>Aij=|qi\u2229qj|\/|qi\u222aqj|<\/code> \u6784\u9020 soft positive \u6743\u91cd\uff0c\u518d\u505a\u53cc\u5411 weighted contrastive loss\u3002\u8fd9\u6837\uff0c\u540c\u4e00 batch \u4e2d\u63cf\u8ff0\u76f8\u8fd1\u89e3\u5256\u4f4d\u7f6e\u7684\u6837\u672c\u4e0d\u4f1a\u88ab\u7b80\u5355\u5f53\u6210\u786c\u8d1f\u6837\u672c\u3002<\/p>\n<p>\u8fd9\u4e2a\u8bbe\u8ba1\u6bd4\u76f4\u63a5\u5957 CLIP loss \u66f4\u5408\u7406\uff1b\u6d88\u878d\u4e2d CLIP \u7248\u672c Dice \u4f4e\u4e8e ITCL \u7248\u672c\uff0c\u652f\u6301\u4f5c\u8005\u7684\u673a\u5236\u5224\u65ad\u3002<\/p>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u5b9e\u9a8c\u4f7f\u7528\u4e24\u4e2a MRIS \u6570\u636e\u96c6\uff1a<\/p>\n<ul>\n<li><strong>QaTa-COV19<\/strong>\uff1a5716 train\u30011429 validation\u30012113 test\uff1b<\/li>\n<li><strong>MosMedData+<\/strong>\uff1a2183 train\u3001273 validation\u3001273 test\u3002<\/li>\n<\/ul>\n<p>\u534a\u76d1\u7763\u6bd4\u4f8b\u8bbe\u7f6e\u4e3a 1%\u30012%\u30015%\u300115% \u6709\u6807\u6ce8\uff0c\u5176\u4f59\u8bad\u7ec3\u6837\u672c\u4f5c\u4e3a\u65e0\u6807\u6ce8\uff1b\u4e3b\u8981\u6307\u6807\u662f Dice \u548c mIoU\u3002\u5b9e\u73b0\u4f7f\u7528 PyTorch Lightning \u4e0e MONAI\uff0c\u56fe\u50cf resized \u5230 224\u00d7224\uff0cbatch size 32\uff0cAdamW \u521d\u59cb\u5b66\u4e60\u7387 <code>3e-4<\/code>\uff0c\u5355\u5f20 NVIDIA RTX A6000\u3002\u9ed8\u8ba4 backbone \u662f MMI-UNet\uff0c\u89c6\u89c9\u7f16\u7801\u5668 ConvNeXt\uff0c\u6587\u672c\u7f16\u7801\u5668 CXR-BERT\u3002<\/p>\n<p>\u4e3b\u8981\u7ed3\u679c\uff1a<\/p>\n<ul>\n<li>\u5728 <strong>QaTa-COV19<\/strong> \u4e0a\uff0cOurs(MMI-UNet) \u5728 2%\/5%\/15%\/100% label ratio \u4e0b Dice \u5206\u522b\u4e3a 87.25\u300188.31\u300189.84\u300191.39\uff0c\u9ad8\u4e8e MMI-UNet baseline \u7684 84.63\u300187.35\u300188.72\u300190.88\u3002<\/li>\n<li>\u5728 <strong>MosMedData+<\/strong> \u4e0a\uff0cOurs(MMI-UNet) \u5728 5%\/15%\/100% \u4e0b Dice \u4e3a 74.51\u300176.42\u300178.70\uff0c\u9ad8\u4e8e MMI-UNet \u7684 72.26\u300174.00\u300178.42\uff1b2% \u60c5\u51b5\u4e0b Ours(GuideDecoder) \u7684 Dice 72.87 \u9ad8\u4e8e GuideDecoder 66.34\uff0cOurs(MMI-UNet) 71.16 \u9ad8\u4e8e MMI-UNet 68.23\u3002<\/li>\n<li>\u6d88\u878d\u663e\u793a\uff0c\u5728 QaTa-COV19 1%\/2% \u4e0b\uff0c\u4ece baseline \u5230 teacher-student EMA\u3001PosAug\u3001T-PatchMix\u3001ITCL \u9010\u6b65\u63d0\u9ad8\uff1b1% Dice \u4ece 81.02 \u63d0\u5347\u5230 86.39\uff0c2% Dice \u4ece 84.63 \u63d0\u5347\u5230 87.25\u3002<\/li>\n<li>\u589e\u5f3a\u6d88\u878d\u663e\u793a naive CutMix \u4f7f Dice \u964d\u5230 83.09\uff0c\u4f4e\u4e8e baseline 84.58\uff1bT-PatchMix \u7248\u672c\u4e3a 85.59\uff1b\u52a0\u5165 ITCL \u7684 Ours-All \u8fbe 86.39\uff0c\u8bf4\u660e\u201c\u5bf9\u9f50\u4fdd\u6301\u201d\u4e0d\u662f\u88c5\u9970\uff0c\u800c\u662f\u5fc5\u8981\u6761\u4ef6\u3002<\/li>\n<\/ul>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u8bc1\u636e\u6574\u4f53\u8f83\u5f3a\uff0c\u4f46\u8fb9\u754c\u6e05\u6670\uff1a<\/p>\n<ul>\n<li>\u53ef\u4fe1\u70b9\uff1a\u6709\u4e24\u4e2a\u533b\u5b66 MRIS \u6570\u636e\u96c6\uff1b\u591a\u4e2a label ratio\uff1b\u5bf9\u6bd4\u81ea\u7136\u56fe\u50cf RES\u3001\u533b\u5b66\u6587\u672c\u5f15\u5bfc\u5206\u5272\u548c\u534a\u76d1\u7763\u65b9\u6cd5\uff1b\u6709\u7ec4\u4ef6\u7ea7\u6d88\u878d\uff0c\u5e76\u7279\u522b\u9a8c\u8bc1 naive CutMix \u4f1a\u635f\u5bb3\u6027\u80fd\uff0c\u76f4\u63a5\u652f\u6491\u8bba\u6587\u4e3b\u5f20\u3002<\/li>\n<li>\u5f31\u70b9\uff1a\u4efb\u52a1\u96c6\u4e2d\u5728 COVID \u80f8\u90e8\u5f71\u50cf\u4e0e\u4f4d\u7f6e\u6587\u672c\uff0c\u5c1a\u4e0d\u80fd\u8bc1\u660e\u5bf9\u591a\u5668\u5b98 CT\u3001MRI\u3001\u5185\u955c\u606f\u8089\u6216 3D \u4f53\u6570\u636e\u666e\u9002\uff1bTextmatch \u90e8\u5206\u7ed3\u679c\u5f15\u7528\u539f\u8bba\u6587\u4e14\u975e\u6240\u6709 label ratio \u90fd\u6709\uff1b\u672a\u770b\u5230\u7edf\u8ba1\u663e\u8457\u6027\u68c0\u9a8c\uff1b\u4ee3\u7801\u672a\u786e\u8ba4\u516c\u5f00\uff1b\u8f93\u5165 resize \u5230 224\u00d7224\uff0c\u4e0d\u80fd\u4ee3\u8868\u9ad8\u5206\u8fa8\u7387 3D \u533b\u5b66\u5206\u5272\u5de5\u7a0b\u3002<\/li>\n<li>\u7ed3\u8bba\u5f3a\u5ea6\uff1a\u53ef\u4ee5\u76f8\u4fe1\u5176\u201c\u5728 MRIS \u4f4e\u6807\u6ce8\u573a\u666f\u4e0b\uff0calignment-aware augmentation \u4f18\u4e8e naive multimodal augmentation\u201d\u7684\u4e3b\u5f20\uff1b\u4f46\u4e0d\u5e94\u6269\u5927\u4e3a\u201c\u901a\u7528\u533b\u5b66\u56fe\u50cf\u5206\u5272\u65b0 SOTA backbone\u201d\u3002<\/li>\n<\/ul>\n<h4>8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n<ul>\n<li>\u4e0e <strong>U-Net\/nnU-Net<\/strong>\uff1aSemi-MedRef \u4e0d\u662f\u66ff\u4ee3 U-Net \u7684\u7ed3\u6784\uff0c\u800c\u662f\u53ef\u53e0\u52a0\u5728 U-Net-like \u591a\u6a21\u6001\u5206\u5272\u5668\u4e0a\u7684\u8bad\u7ec3\/\u589e\u5f3a\u6846\u67b6\u3002\u5176\u9ed8\u8ba4\u6a21\u578b MMI-UNet \u4ecd\u4fdd\u7559 U-Net \u5f0f\u5206\u5272\u601d\u60f3\u3002<\/li>\n<li>\u4e0e <strong>MedNeXt \/ CNN-based segmentation<\/strong>\uff1a\u53ef\u628a ConvNeXt\/MedNeXt \u7c7b encoder \u5f53\u89c6\u89c9\u5206\u652f\uff0c\u4f46\u672c\u6587\u8d21\u732e\u4e0d\u662f\u5377\u79ef\u6a21\u5757\u3002<\/li>\n<li>\u4e0e <strong>TransUNet \/ Swin-UNet \/ LViT<\/strong>\uff1a\u66f4\u63a5\u8fd1 LViT \u7684 language-guided segmentation \u65b9\u5411\uff1b\u5982\u679c\u7528 Transformer \u878d\u5408\u56fe\u6587\uff0cPosAug\/T-PatchMix\/ITCL \u4ecd\u53ef\u4f5c\u4e3a\u8bad\u7ec3\u7b56\u7565\u3002<\/li>\n<li>\u4e0e <strong>Mamba \/ VMamba \/ SegMamba \/ DAMamba<\/strong>\uff1a\u6ca1\u6709\u4f7f\u7528 SSM\/Mamba\uff1b\u4f46\u5982\u679c DAMamba \u505a\u6587\u672c\u63d0\u793a\u6216\u534a\u76d1\u7763\u533b\u5b66 referring segmentation\uff0cT-PatchMix \u4e0e ITCL \u53ef\u4f5c\u4e3a\u8bad\u7ec3\u5c42\u9762\u7684\u8865\u5145\u3002<\/li>\n<li>\u4e0e <strong>foundation medical segmentation \/ SAM\/MedSAM<\/strong>\uff1a\u4e0d\u50cf MedSAM \u90a3\u6837\u505a promptable universal segmentation\uff1b\u5b83\u66f4\u5173\u6ce8\u6587\u672c\u4f4d\u7f6e\u63cf\u8ff0\u4e0e\u533b\u5b66\u56fe\u50cf\u4e4b\u95f4\u7684\u5f31\u76d1\u7763\u5bf9\u9f50\u3002\u53ef\u4e0e SAM-generated pseudo label \u6216 MedSAM encoder \u7ed3\u5408\uff0c\u4f46\u8bba\u6587\u672a\u9a8c\u8bc1\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5982\u679c\u4f60\u7684\u6838\u5fc3\u8bfe\u9898\u662f polyp segmentation \u6216 DAMamba backbone\uff0cSemi-MedRef \u4e0d\u662f\u76f4\u63a5 baseline \u66ff\u4ee3\uff1b\u4f46\u5b83\u6709\u4e09\u70b9\u53ef\u501f\u9274\uff1a<\/p>\n<ol>\n<li><strong>\u8bad\u7ec3\u7b56\u7565\u501f\u9274<\/strong>\uff1a\u628a CutMix\/Copy-Paste \u6539\u9020\u6210 anatomy-aware \u6216 lesion-aware \u7248\u672c\uff0c\u907f\u514d\u7834\u574f\u4f2a\u6807\u7b7e\u8bed\u4e49\u3002<\/li>\n<li><strong>\u5f31\u6807\u6ce8\/\u6587\u672c\u63d0\u793a\u65b9\u5411<\/strong>\uff1a\u5982\u679c\u672a\u6765\u505a\u62a5\u544a\u6587\u672c\u3001\u4f4d\u7f6e\u63d0\u793a\u3001\u533b\u751f\u63cf\u8ff0\u5f15\u5bfc\u7684\u5206\u5272\uff0c\u8fd9\u7bc7\u53ef\u4ee5\u4f5c\u4e3a related work \u4e2d MRIS + SSL \u7684\u65b0\u8fd1\u53c2\u8003\u3002<\/li>\n<li><strong>\u6d88\u878d\u8bbe\u8ba1\u501f\u9274<\/strong>\uff1a\u5b83\u660e\u786e\u6bd4\u8f83 naive CutMix\u3001generic TextStrong\u3001CLIP loss \u548c ITCL\uff0c\u5bf9\u8bc1\u660e\u201c\u533b\u5b66\u7279\u5b9a\u5bf9\u9f50\u673a\u5236\u5fc5\u8981\u201d\u5f88\u6709\u5e2e\u52a9\u3002<\/li>\n<\/ol>\n<p>\u5bf9\u7eaf\u5168\u81ea\u52a8\u606f\u8089\u5206\u5272\uff0c\u5efa\u8bae\u53ea\u5438\u6536 lesion-gated mixing \u4e0e alignment-aware augmentation \u7684\u601d\u60f3\uff0c\u4e0d\u5fc5\u7167\u642c\u6587\u672c\u5206\u652f\u3002<\/p>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5efa\u8bae\u7cbe\u8bfb<\/strong>\u3002\u5982\u679c\u4f60\u6b63\u5728\u5199\u533b\u5b66\u56fe\u50cf\u5206\u5272 foundation\/referring\/weakly supervised \u76f8\u5173\u8bba\u6587\uff0c\u503c\u5f97\u5168\u6587\u8bfb\uff1b\u5982\u679c\u53ea\u505a CNN\/Mamba backbone\uff0c\u4f18\u5148\u8bfb\u65b9\u6cd5\u548c\u6d88\u878d\u90e8\u5206\u5373\u53ef\u3002\u91cd\u70b9\u770b T-PatchMix \u7684\u4f2a\u6807\u7b7e\u540c\u6b65\u903b\u8f91\u548c ITCL \u7684 soft positive \u6784\u9020\u3002<\/p>\n<hr \/>\n<h2>\u8bba\u6587 2\uff1aEvaluation of Anatomical Shape Priors in Deep Learning-Based Cardiac Multi-Compartment Segmentation<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aEvaluation of Anatomical Shape Priors in Deep Learning-Based Cardiac Multi-Compartment Segmentation<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aMichael Hudler, Franz Thaler, Martin Urschler \/ Michael Hudler<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-15 arXiv v1\uff1b\u8bba\u6587\u9996\u9875\u6807\u6ce8\u53d1\u8868\u4e8e Proceedings of the Third Austrian Symposium on AI, Robotics, and Vision (AIRoV 2026)<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint \/ AIRoV 2026 proceedings<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.15707<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1a\/tmp\/medseg_daily_2026-05-19\/shape_priors_2605.15707.pdf\uff1bhttps:\/\/arxiv.org\/pdf\/2605.15707<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u672a\u83b7\u53d6\uff1bPDF \u4e0e arXiv \u9875\u9762\u672a\u786e\u8ba4\u5b98\u65b9\u4ee3\u7801<\/li>\n<li>\u4efb\u52a1\uff1awhole-heart multi-compartment CT segmentation\uff0c7 \u4e2a\u5fc3\u810f\/\u5927\u8840\u7ba1\u524d\u666f\u7c7b\u522b + \u80cc\u666f<\/li>\n<li>\u6570\u636e\u96c6\uff1aMM-WHS CT\u3001WHS++ CT<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1a\u8d1f\u7ed3\u679c\/\u8bc4\u4f30\u7814\u7a76\uff1b3D U-Net baseline\uff1bshape-aware losses\uff1bpopulation heatmap-guided U-Net variants\uff1banatomical prior evaluation<\/li>\n<\/ul>\n<h3>paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n<h4>1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n<p>\u8fd9\u7bc7\u8bba\u6587\u6700\u91cd\u8981\u7684\u4ef7\u503c\u662f\u7ed9\u51fa\u4e00\u4e2a\u6e05\u6670\u7684\u8d1f\u7ed3\u679c\uff1a\u5728 MM-WHS\/WHS++ whole-heart CT \u5206\u5272\u4e2d\uff0c\u5f3a 3D U-Net \u5df2\u7ecf\u5b66\u5230\u76f8\u5f53\u591a\u9690\u5f0f\u89e3\u5256\u89c4\u5f8b\uff0c\u7b80\u5355\u624b\u5de5\u4f53\u79ef\u3001\u77e9\u3001\u8d28\u5fc3\u5173\u7cfb\u6216\u5e73\u5747 heatmap \u5148\u9a8c\u5e76\u4e0d\u80fd\u7a33\u5b9a\u8d85\u8fc7 baseline\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76 whole-heart multi-compartment CT segmentation\uff0c\u76ee\u6807\u662f\u5206\u5272\u5de6\/\u53f3\u5fc3\u5ba4\u3001\u5de6\/\u53f3\u5fc3\u623f\u3001\u5fc3\u808c\u3001\u5347\u4e3b\u52a8\u8109\u3001\u80ba\u52a8\u8109\u7b49\u591a\u7ed3\u6784\u3002\u8fd9\u4e2a\u4efb\u52a1\u5bf9\u5fc3\u529f\u80fd\u5b9a\u91cf\u3001\u6cbb\u7597\u8ba1\u5212\u3001\u4eff\u771f\u5efa\u6a21\u548c\u56fe\u50cf\u5f15\u5bfc\u5e72\u9884\u5f88\u91cd\u8981\u3002<\/p>\n<p>\u533b\u5b66\u56fe\u50cf\u5206\u5272\u4e2d\uff0c\u4e00\u4e2a\u957f\u671f\u95ee\u9898\u662f\uff1aCNN\/U-Net \u4e3b\u8981\u4f9d\u8d56 appearance-driven learning\uff0c\u7f3a\u5c11\u663e\u5f0f\u89e3\u5256\u5f62\u72b6\u7ea6\u675f\uff1b\u800c\u4f20\u7edf\u7edf\u8ba1\u5f62\u72b6\u6a21\u578b\u5f3a\u8c03 anatomical plausibility\u3002\u4f5c\u8005\u7684\u95ee\u9898\u662f\uff1a\u5728\u73b0\u4ee3 3D U-Net \u5df2\u7ecf\u5f88\u5f3a\u7684\u60c5\u51b5\u4e0b\uff0c\u8f7b\u91cf\u624b\u5de5 shape prior \u662f\u5426\u8fd8\u80fd\u5e26\u6765\u53ef\u6d4b\u91cf\u589e\u76ca\uff1f<\/p>\n<p>\u8fd9\u662f\u4e00\u7bc7\u504f evaluation\/negative result \u7684\u8bba\u6587\uff0c\u4e0d\u662f\u63d0\u51fa\u4e00\u4e2a\u590d\u6742\u65b0\u7f51\u7edc\u3002\u5b83\u7684 paper map \u53ef\u4ee5\u6982\u62ec\u4e3a\uff1a\u7814\u7a76 cardiac CT multi-compartment segmentation\uff1b\u4e3b\u52a8\u4f5c\u662f\u628a\u591a\u7c7b\u5f62\u72b6\u5148\u9a8c\u4f5c\u4e3a loss \u6216 heatmap architecture \u52a0\u5230 3D U-Net \u4e0a\uff1b\u4e3b\u5f20\u662f\u8fd9\u4e9b\u7b80\u5355\u663e\u5f0f\u5148\u9a8c\u6ca1\u6709\u7a33\u5b9a\u63d0\u9ad8\u6027\u80fd\uff1b\u8bc1\u636e\u6765\u81ea MM-WHS \u4e0e WHS++ \u7684 Dice\/Jaccard\/HD\/ASSD \u5bf9\u6bd4\uff1b\u4e3b\u8981\u5931\u8d25\u98ce\u9669\u662f\u5b9e\u9a8c\u89c4\u6a21\u8f83\u5c0f\u3001\u5b9e\u73b0\u7ec6\u8282\u4f9d\u8d56 thesis\/\u8865\u5145\u6750\u6599\uff0c\u7ed3\u8bba\u53ef\u80fd\u53ea\u9002\u7528\u4e8e\u7279\u5b9a\u6570\u636e\u96c6\u548c\u5148\u9a8c\u5f62\u5f0f\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u8ba4\u4e3a\u6807\u51c6 3D U-Net \u7684\u4e0d\u8db3\u662f\u6ca1\u6709\u663e\u5f0f\u7f16\u7801\u89e3\u5256\u53ef\u884c\u6027\uff0c\u4f8b\u5982\u5668\u5b98\u4f53\u79ef\u8303\u56f4\u3001\u8d28\u5fc3\u5173\u7cfb\u3001\u7a7a\u95f4\u5e03\u5c40\u3001\u5e73\u5747\u5f62\u72b6\u5206\u5e03\u7b49\u3002\u56e0\u6b64\uff0c\u7406\u8bba\u4e0a\u5b83\u53ef\u80fd\u4ea7\u751f\u89e3\u5256\u4e0d\u5408\u7406 mask\u3002\u4f20\u7edf statistical shape models \u53ef\u4ee5\u7f16\u7801\u5f62\u72b6\uff0c\u4f46\u901a\u5e38\u5de5\u7a0b\u590d\u6742\uff0c\u96be\u4ee5\u76f4\u63a5\u878d\u5165\u6df1\u5ea6\u7f51\u7edc\u8bad\u7ec3\u3002<\/p>\n<p>\u4e0d\u8fc7\uff0c\u8bba\u6587\u7684\u6838\u5fc3\u53d1\u73b0\u53cd\u8fc7\u6765\u6311\u6218\u4e86\u8fd9\u4e2a\u52a8\u673a\uff1a\u5728\u5f53\u524d whole-heart \u6570\u636e\u96c6\u4e0a\uff0cbaseline 3D U-Net \u901a\u8fc7\u56fe\u50cf\u3001\u6570\u636e\u589e\u5f3a\u3001\u591a\u5c3a\u5ea6 encoder-decoder \u548c skip connection \u5df2\u7ecf\u9690\u5f0f\u5b66\u5230\u5168\u5c40\u7ed3\u6784\uff0c\u5269\u4e0b\u9519\u8bef\u66f4\u591a\u96c6\u4e2d\u5728\u8fb9\u754c\u548c\u8584\u7ed3\u6784\uff1b\u7c97\u7c92\u5ea6 shape prior \u5bf9\u8fd9\u4e9b\u5c40\u90e8\u8bef\u5dee\u5e2e\u52a9\u6709\u9650\u3002<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>\u8bba\u6587\u4ee5\u6807\u51c6 3D U-Net \u4e3a\u53c2\u7167\uff0c\u8bc4\u4f30\u4e24\u7c7b\u663e\u5f0f\u5f62\u72b6\u5148\u9a8c\uff1a<\/p>\n<ol>\n<li><strong>Shape-aware losses<\/strong>\uff1a\u4e0d\u6539 backbone\uff0c\u53ea\u5728 baseline Generalized Dice + Cross-Entropy \u5916\u52a0\u5165\u5f62\u72b6\u6b63\u5219\u3002<\/li>\n<li><strong>Architectural priors<\/strong>\uff1a\u4ece\u8bad\u7ec3\u6807\u7b7e\u914d\u51c6\u5f97\u5230 population-level multi-class probability heatmaps\uff0c\u5e76\u628a\u8fd9\u4e9b heatmap \u7528\u8f85\u52a9\u5934\u3001\u53cc decoder\u3001\u53cc encoder \u6216 cascade \u7684\u65b9\u5f0f\u878d\u5165 U-Net\u3002<\/li>\n<\/ol>\n<p>\u6570\u636e\u9884\u5904\u7406\u5305\u62ec\uff1a\u628a CT \u548c\u6807\u7b7e\u91cd\u5b9a\u5411\u5230\u5171\u540c\u89e3\u5256\u7ea6\u5b9a\u3001\u7b49\u4f53\u7d20\u91cd\u91c7\u6837\u3001\u7528 label centroid \u5c45\u4e2d\u3001\u5d4c\u5165\u6807\u51c6 field of view\uff0c\u5e76\u7528 Procrustes alignment \u4ece\u8bad\u7ec3\u6807\u7b7e\u6784\u9020 population heatmaps\u3002<\/p>\n<p>Baseline \u662f\u6807\u51c6 3D U-Net\uff1a\u5355\u901a\u9053 CT \u8f93\u5165\uff0c8 \u4e2a\u8f93\u51fa\u7c7b\uff08\u542b\u80cc\u666f\uff09\uff0cencoder-decoder + skip connection\uff0cbase channels=64\uff0c\u4e0b\u91c7\u6837\u540e\u901a\u9053\u7ffb\u500d\uff0cLeakyReLU\uff0c\u4f7f\u7528 Generalized Dice + Cross-Entropy loss\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 A\uff1aVolume regularization<\/strong><br \/>\n\u8f93\u5165\u662f soft prediction \u7684\u5404\u7c7b\u522b\u4f53\u79ef\uff0c\u7ea6\u675f\u5176\u4e0d\u8981\u504f\u79bb\u8bad\u7ec3\u96c6\u4f30\u8ba1\u7684\u7c7b\u522b\u4f53\u79ef\u5747\u503c\u548c\u6807\u51c6\u5dee\u3002\u5b83\u89e3\u51b3\u7684\u662f\u201c\u5668\u5b98\u592a\u5927\/\u592a\u5c0f\u201d\u7684\u5168\u5c40\u95ee\u9898\u3002\u521b\u65b0\u6027\u8f83\u4f4e\uff0c\u4f46\u89e3\u91ca\u6027\u5f3a\u3002\u9002\u5408\u5c0f\u6837\u672c\u6216\u6781\u7aef\u5f02\u5e38\u68c0\u6d4b\uff1b\u4f46\u5728\u6b63\u5e38 whole-heart CT \u4e2d\uff0c\u4f53\u79ef\u53d8\u5f02\u8f83\u590d\u6742\uff0c\u5355\u4e00\u5747\u503c\/\u65b9\u5dee\u7ea6\u675f\u53ef\u80fd\u8fc7\u7c97\u3002<\/p>\n<p><strong>\u6a21\u5757 B\uff1aMoment-based shape regularization<\/strong><br \/>\n\u6bd4\u8f83\u9884\u6d4b mask \u7684 soft \u4e00\u9636\/\u4e8c\u9636\u7a7a\u95f4\u77e9\u4e0e\u8bad\u7ec3\u96c6\u53c2\u8003\u77e9\uff0c\u4f8b\u5982 centroid\u3001ellipsoid-like shape moments\u3002\u5b83\u8bd5\u56fe\u7ea6\u675f\u7ed3\u6784\u4f4d\u7f6e\u548c\u7c97\u7565\u5f62\u72b6\u3002\u5b9e\u9a8c\u4e2d\u5b83\u5728 WHS++ Dice \u4e0a\u7565\u9ad8\u4e8e baseline\uff0c\u4f46\u5e45\u5ea6\u5f88\u5c0f\uff0cMM-WHS \u4e0a\u57fa\u672c\u6301\u5e73\u3002<\/p>\n<p><strong>\u6a21\u5757 C\uff1aAnatomical relation loss<\/strong><br \/>\n\u7ea6\u675f\u7c7b\u522b\u8d28\u5fc3\u4e4b\u95f4\u7684 pairwise distances \u548c angular relations\u3002\u5b83\u7684\u76ee\u6807\u662f\u4fdd\u6301\u5fc3\u810f\u591a\u7ed3\u6784\u4e4b\u95f4\u7684\u7a7a\u95f4\u5173\u7cfb\uff0c\u4f46 MM-WHS \u4e0a Dice \u4ece baseline 90.85 \u964d\u5230 88.98\uff0c\u8bf4\u660e\u8fc7\u786c\u6216\u8fc7\u7c97\u7684\u5173\u7cfb\u7ea6\u675f\u53ef\u80fd\u548c\u771f\u5b9e\u53d8\u5f02\/\u914d\u51c6\u8bef\u5dee\u51b2\u7a81\u3002<\/p>\n<p><strong>\u6a21\u5757 D\uff1aPopulation heatmap-guided architectures<\/strong><br \/>\n\u4f5c\u8005\u6d4b\u8bd5\u4e86\u591a\u4e2a\u7ed3\u6784\uff1a\u6700\u540e decoder \u5c42\u8f85\u52a9 heatmap prediction head\u3001multi-layer deep supervision heatmap\u30012-Decoder\uff08\u4e00\u4e2a segmentation decoder\uff0c\u4e00\u4e2a heatmap decoder\uff09\u30012-Encoder\uff08\u56fe\u50cf\u548c heatmap \u53cc\u8f93\u5165\uff09\u3001Cascaded \u4e09\u4e2a U-Net \u7c97\u5230\u7ec6 refinement\u3002\u5b83\u4eec\u628a\u5e73\u5747\u89e3\u5256\u4f4d\u7f6e\u4f5c\u4e3a\u7f51\u7edc\u989d\u5916\u4efb\u52a1\u6216\u8f93\u5165\uff0c\u4f46\u591a\u6570\u6ca1\u6709\u8d85\u8fc7 baseline\u3002<\/p>\n<p>\u8fd9\u4e9b\u6a21\u5757\u5bf9\u5176\u4ed6\u533b\u5b66\u5206\u5272\u6846\u67b6\u7684\u8fc1\u79fb\u4ef7\u503c\u5728\u4e8e\uff1a\u5b83\u4eec\u662f\u5bb9\u6613\u5b9e\u73b0\u7684 \u201cshape prior baseline\u201d\uff0c\u53ef\u4f5c\u4e3a\u4f60\u8bc4\u4f30\u66f4\u590d\u6742 topology\/diffusion\/flow prior \u662f\u5426\u771f\u6b63\u6709\u7528\u7684\u5bf9\u7167\u7ec4\uff1b\u4f46\u4e0d\u5efa\u8bae\u76f4\u63a5\u4f5c\u4e3a polyp segmentation \u6216 DAMamba \u7684\u6838\u5fc3\u521b\u65b0\uff0c\u56e0\u4e3a\u7ed3\u679c\u663e\u793a\u5176\u6536\u76ca\u4e0d\u7a33\u5b9a\u3002<\/p>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u6570\u636e\u4e0e\u8bbe\u7f6e\uff1a<\/p>\n<ul>\n<li><strong>MM-WHS CT<\/strong>\uff1a20 \u4e2a annotated CT \u8bad\u7ec3\uff0c40 \u4e2a hidden-label CT \u6d4b\u8bd5\uff1b7 \u4e2a\u524d\u666f\u7c7b\u522b\u3002<\/li>\n<li><strong>WHS++ CT<\/strong>\uff1a\u4f7f\u7528\u5176\u8bad\u7ec3\u96c6\u540e\u534a\u90e8\u5206 20 \u4e2a CT case \u4f5c\u4e3a\u53ef\u8bbf\u95ee ground truth \u7684\u8865\u5145\u8bc4\u4f30\u3002<\/li>\n<li>\u8f93\u5165 resolution \u6d4b\u8bd5\u4e86 <code>64^3<\/code> \u4e0e <code>128^3<\/code> crop\uff1b\u6240\u6709\u7ed3\u6784\u4f7f\u7528\u76f8\u540c\u51e0\u4f55\u4e0e\u5f3a\u5ea6\u589e\u5f3a\u3002<\/li>\n<li>\u6307\u6807\u5305\u62ec Dice\u3001Jaccard\u3001Hausdorff Distance\uff08HD\uff09\u3001Average Symmetric Surface Distance\uff08ASSD\uff09\uff1bWHS++ \u4e3b\u8981\u6c47\u603b Dice \u7b49\u53ef\u7528\u6307\u6807\u3002<\/li>\n<\/ul>\n<p>\u5173\u952e\u7ed3\u679c\uff1a<\/p>\n<ul>\n<li><strong>MM-WHS CT, 64\u00b3, shape-aware losses<\/strong>\uff1abaseline Dice 90.85\uff0cvolume regularization 90.85\uff0cmoment regularization 90.84\uff0canatomical relation 88.98\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u4f53\u79ef\/\u77e9\u57fa\u672c\u6301\u5e73\uff0c\u5173\u7cfb loss \u660e\u663e\u53d8\u5dee\u3002<\/li>\n<li><strong>MM-WHS CT, 64\u00b3, architecture priors<\/strong>\uff1abaseline Dice 90.85\uff1bHM multilayer 90.60\uff1b2-Decoder 90.73\uff1bCascaded 90.32\u3002\u6ca1\u6709\u8d85\u8fc7 baseline\uff0c\u5c3d\u7ba1 Cascaded \u7684 HD \u7565\u4f18\uff087.55 vs 7.64 mm\uff09\u3002<\/li>\n<li><strong>MM-WHS CT, 128\u00b3<\/strong>\uff1abaseline Dice 92.05\uff1b2-Encoder 92.02\uff1bCascaded 92.04\uff0c\u51e0\u4e4e\u8ffd\u5e73\u4f46\u672a\u8d85\u8fc7 baseline\u3002<\/li>\n<li><strong>WHS++ CT, 64\u00b3<\/strong>\uff1abaseline Dice 88.93\uff1bvolume 89.09\uff1bmoment 89.16\uff1banatomical relation 88.66\u3002loss prior \u6709\u6781\u5c0f\u63d0\u5347\uff0c\u4f46\u4e0d\u7a33\u5b9a\uff1barchitecture prior \u4e2d HM multilayer 88.72\u30012-Encoder 87.75\u3001Cascaded 86.13\uff0c\u5747\u4e0d\u5982 baseline\u3002<\/li>\n<li>\u5b9a\u6027\u7ed3\u679c\u663e\u793a baseline\u3001shape loss\u30012-Decoder \u90fd\u80fd\u6062\u590d\u5168\u5c40\u5fc3\u810f\u7ed3\u6784\uff0c\u5dee\u5f02\u4e3b\u8981\u5728\u8fb9\u754c\u548c\u5c0f\u7ed3\u6784\u3002<\/li>\n<\/ul>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u8fd9\u7bc7\u7684\u8bc1\u636e\u4ef7\u503c\u5728\u4e8e\u201c\u8d1f\u7ed3\u679c\u6e05\u695a\u3001baseline \u5f3a\u3001\u6bd4\u8f83\u76f4\u63a5\u201d\u3002\u5b83\u6ca1\u6709\u5938\u5927\u65b0\u65b9\u6cd5\uff0c\u800c\u662f\u5b9e\u8bc1\u8bf4\u660e\u7b80\u5355 shape prior \u4e0d\u7a33\u5b9a\u3002<\/p>\n<p>\u53ef\u4fe1\u70b9\uff1a<\/p>\n<ul>\n<li>\u4f7f\u7528\u516c\u5f00\u4e14\u7ecf\u5178\u7684 whole-heart segmentation \u6570\u636e\u96c6\uff1b<\/li>\n<li>baseline \u662f\u5f3a 3D U-Net\uff0c\u4e14\u7ed3\u679c\u4e0e MM-WHS\/WHS++ \u76f8\u5173\u6311\u6218\u4e2d\u7684\u5f3a\u65b9\u6cd5\u76f8\u8fd1\uff1b<\/li>\n<li>\u540c\u65f6\u6d4b\u8bd5 loss prior \u4e0e architecture prior\uff0c\u907f\u514d\u53ea\u5426\u5b9a\u4e00\u79cd\u5b9e\u73b0\uff1b<\/li>\n<li>\u540c\u65f6\u62a5\u544a overlap \u548c boundary metric\uff0c\u6ce8\u610f\u5230\u67d0\u4e9b\u65b9\u6cd5\u53ef\u80fd\u6539\u5584 HD \u4f46\u4e0d\u6539\u5584 Dice\u3002<\/li>\n<\/ul>\n<p>\u5c40\u9650\uff1a<\/p>\n<ul>\n<li>\u8bba\u6587\u6b63\u6587\u5f88\u77ed\uff0c\u5f88\u591a\u5b9e\u73b0\u7ec6\u8282\u6307\u5411 thesis\uff1b<\/li>\n<li>\u6837\u672c\u91cf\u8f83\u5c0f\uff0cMM-WHS\/WHS++ \u672c\u8eab\u5e76\u4e0d\u8986\u76d6\u6240\u6709\u5fc3\u810f\u53d8\u5f02\uff1b<\/li>\n<li>\u624b\u5de5\u5148\u9a8c\u8f83\u7c97\uff0c\u4e0d\u80fd\u4ee3\u8868 learned anatomical prior\u3001diffusion prior\u3001flow matching prior \u6216 topology-aware prior\uff1b<\/li>\n<li>\u6ca1\u6709\u4ee3\u7801\uff0c\u96be\u4ee5\u786e\u8ba4\u8bad\u7ec3 recipe\u3001\u516c\u5e73\u6027\u548c\u8c03\u53c2\u8303\u56f4\uff1b<\/li>\n<li>\u8d1f\u7ed3\u679c\u4e0d\u5e94\u63a8\u5e7f\u5230\u6240\u6709\u533b\u5b66\u5206\u5272\uff0c\u5c24\u5176\u4e0d\u9002\u7528\u4e8e\u62d3\u6251\u6781\u5176\u5173\u952e\u7684\u8840\u7ba1\u3001\u6c14\u9053\u3001\u795e\u7ecf\u7ea4\u7ef4\u3001\u606f\u8089\u8fb9\u754c\u6216\u80bf\u7624\u4e0d\u89c4\u5219\u5f62\u6001\u573a\u666f\u3002<\/li>\n<\/ul>\n<h4>8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n<ul>\n<li>\u4e0e <strong>U-Net\/3D U-Net\/nnU-Net<\/strong>\uff1a\u5b83\u5f3a\u5316\u4e86\u4e00\u4e2a\u91cd\u8981\u5224\u65ad\uff1a\u5f3a 3D U-Net baseline \u5f88\u96be\u88ab\u7b80\u5355\u5148\u9a8c\u6a21\u5757\u51fb\u8d25\u3002\u505a\u65b0\u6846\u67b6\u65f6\u5fc5\u987b\u4e25\u8083\u5bf9\u6bd4 U-Net\/nnU-Net\uff0c\u800c\u4e0d\u80fd\u53ea\u6bd4\u8f83\u5f31 baseline\u3002<\/li>\n<li>\u4e0e <strong>MedNeXt<\/strong>\uff1aMedNeXt \u7b49\u73b0\u4ee3 CNN backbone \u53ef\u80fd\u66f4\u5f3a\uff0c\u56e0\u6b64\u7b80\u5355 shape prior \u7684\u8fb9\u9645\u6536\u76ca\u53ef\u80fd\u66f4\u5c0f\uff1b\u4f46 learned prior \u4e0e boundary-aware decoder \u4ecd\u53ef\u80fd\u6709\u4ef7\u503c\u3002<\/li>\n<li>\u4e0e <strong>UNETR\/Swin-UNet\/Transformer<\/strong>\uff1a\u8bba\u6587\u6ca1\u6709\u9a8c\u8bc1 Transformer\uff0c\u4f46\u5176\u7ed3\u8bba\u63d0\u9192\uff1a\u957f\u7a0b\u4f9d\u8d56\u6216\u5168\u5c40 prior \u4e0d\u7b49\u4e8e\u81ea\u52a8\u63d0\u5347\uff0c\u5fc5\u987b\u8bc1\u660e\u5bf9\u5c40\u90e8\u8fb9\u754c\/\u5c0f\u7ed3\u6784\u6709\u5e2e\u52a9\u3002<\/li>\n<li>\u4e0e <strong>Mamba\/VMamba\/SegMamba\/DAMamba<\/strong>\uff1a\u540c\u6837\u6ca1\u6709\u9a8c\u8bc1 Mamba\uff1b\u5bf9 DAMamba \u7684\u542f\u53d1\u662f\uff0c\u4e0d\u8981\u53ea\u58f0\u79f0 Mamba \u6355\u83b7\u5168\u5c40\u5f62\u72b6\uff0c\u8981\u7528\u5f3a 3D U-Net \u6216 nnU-Net \u6bd4\u8f83\uff0c\u5e76\u52a0\u5165\u8fb9\u754c\/HD\/ASSD \u8bc1\u636e\u3002<\/li>\n<li>\u4e0e <strong>foundation medical segmentation<\/strong>\uff1a\u5b83\u4e0d\u5c5e\u4e8e SAM\/MedSAM \u65b9\u5411\uff1b\u4f46\u53ef\u4f5c\u4e3a foundation model adaptation \u4e2d\u201c\u663e\u5f0f\u5148\u9a8c\u662f\u5426\u771f\u7684\u6709\u7528\u201d\u7684\u53cd\u9762\u8bc1\u636e\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u8fd9\u7bc7\u5bf9\u4f60\u7684\u7814\u7a76\u5f88\u6709\u65b9\u6cd5\u8bba\u4ef7\u503c\uff1a<\/p>\n<ol>\n<li><strong>\u4f5c\u4e3a related work\/\u5b9e\u9a8c\u8bbe\u8ba1\u8b66\u793a<\/strong>\uff1a\u5982\u679c\u4f60\u8981\u5728 DAMamba \u6216 U-Net \u91cc\u52a0\u5165 topology\/shape\/anatomical prior\uff0c\u5fc5\u987b\u8bc1\u660e\u5b83\u8d85\u8fc7\u5f3a baseline\uff0c\u800c\u4e0d\u662f\u53ea\u8bb2\u5148\u9a8c\u5408\u7406\u3002<\/li>\n<li><strong>\u4f5c\u4e3a baseline \u8bbe\u8ba1\u53c2\u8003<\/strong>\uff1avolume\u3001moment\u3001centroid relation\u3001heatmap prior \u53ef\u4ee5\u4f5c\u4e3a\u7b80\u5355 prior baselines\uff0c\u7528\u6765\u886c\u6258\u66f4\u5f3a\u7684 learned prior \u6216 boundary-aware module\u3002<\/li>\n<li><strong>\u5bf9 polyp segmentation \u7684\u95f4\u63a5\u542f\u53d1<\/strong>\uff1a\u606f\u8089\u5f62\u6001\u53d8\u5316\u5927\uff0c\u7c97\u5747\u503c\u5f62\u72b6\u5148\u9a8c\u53ef\u80fd\u66f4\u4e0d\u9002\u5408\uff1b\u66f4\u6709\u5e0c\u671b\u7684\u662f boundary-aware\u3001uncertainty-aware\u3001texture\/illumination-aware \u6216 foundation distillation\uff0c\u800c\u4e0d\u662f\u5f3a\u884c\u52a0\u5165\u56fa\u5b9a shape prior\u3002<\/li>\n<\/ol>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5efa\u8bae\u7cbe\u8bfb\uff0c\u4f46\u4ee5\u201c\u8d1f\u7ed3\u679c\u548c\u5b9e\u9a8c\u8bbe\u8ba1\u201d\u89d2\u5ea6\u8bfb<\/strong>\u3002\u5982\u679c\u4f60\u6b63\u5728\u5199 introduction\/related work\uff0c\u53ef\u4ee5\u5f15\u7528\u5b83\u8bf4\u660e\u663e\u5f0f\u5148\u9a8c\u5e76\u975e\u514d\u8d39\u6536\u76ca\uff1b\u5982\u679c\u4f60\u6b63\u5728\u8bbe\u8ba1 DAMamba\/\u62d3\u6251\u7ea6\u675f\u6a21\u5757\uff0c\u5e94\u91cd\u70b9\u8bfb Methods \u548c Results\uff0c\u501f\u9274\u5176 baseline-first \u7684\u9a8c\u8bc1\u6001\u5ea6\u3002<\/p>\n<hr \/>\n<h2>\u4eca\u65e5\u63a8\u8350\u4f18\u5148\u7ea7<\/h2>\n<ol>\n<li><strong>Semi-MedRef<\/strong>\uff1a\u66f4\u503c\u5f97\u4f18\u5148\u8bfb\u3002\u5b83\u6709\u660e\u786e\u673a\u5236\u3001\u8f83\u5b8c\u6574\u6d88\u878d\u3001\u4e0e\u4f4e\u6807\u6ce8\/\u6587\u672c\u63d0\u793a\/foundation medical segmentation \u65b9\u5411\u66f4\u63a5\u8fd1\uff0c\u9002\u5408\u6269\u5c55\u5230\u533b\u5b66 referring segmentation \u6216\u534a\u76d1\u7763\u8bad\u7ec3\u7b56\u7565\u3002<\/li>\n<li><strong>Evaluation of Anatomical Shape Priors<\/strong>\uff1a\u66f4\u9002\u5408\u4f5c\u4e3a\u8d1f\u7ed3\u679c\u4e0e\u5b9e\u9a8c\u8bbe\u8ba1\u53c2\u8003\u3002\u5b83\u5bf9\u201c\u5f3a U-Net baseline + \u5f62\u72b6\u5148\u9a8c\u662f\u5426\u6709\u5fc5\u8981\u201d\u7ed9\u51fa\u6e05\u6670\u8bc1\u636e\uff0c\u9002\u5408\u5199 related work \u548c\u8bbe\u8ba1 ablation\u3002<\/li>\n<\/ol>\n<h2>\u4eca\u65e5 PDF \u83b7\u53d6\u60c5\u51b5<\/h2>\n<ul>\n<li>\u8bba\u6587 1\uff1a\u5df2\u9644 PDF\uff1b\u672c\u5730\u6587\u4ef6 <code>\/tmp\/medseg_daily_2026-05-19\/semi_medref_2605.15720.pdf<\/code>\uff0c\u53ef\u8bbf\u95ee PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.15720<\/li>\n<li>\u8bba\u6587 2\uff1a\u5df2\u9644 PDF\uff1b\u672c\u5730\u6587\u4ef6 <code>\/tmp\/medseg_daily_2026-05-19\/shape_priors_2605.15707.pdf<\/code>\uff0c\u53ef\u8bbf\u95ee PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.15707<\/li>\n<\/ul>\n<h2>\u4eca\u65e5\u53ef\u6267\u884c\u5efa\u8bae<\/h2>\n<ol>\n<li><strong>\u5148\u8bfb Semi-MedRef \u7684\u65b9\u6cd5\u4e0e\u6d88\u878d<\/strong>\uff1a\u5c24\u5176 T-PatchMix \u5982\u4f55\u540c\u6b65\u56fe\u50cf patch\u3001\u4f2a mask \u548c\u6587\u672c\u4f4d\u7f6e span\uff1b\u8fd9\u4e2a\u601d\u60f3\u53ef\u8fc1\u79fb\u4e3a lesion-aware Copy-Paste \u6216 anatomy-aware augmentation\u3002<\/li>\n<li><strong>\u628a shape prior \u8bba\u6587\u4f5c\u4e3a baseline \u8b66\u793a\u52a0\u5165\u7b14\u8bb0<\/strong>\uff1a\u5982\u679c\u540e\u7eed\u505a DAMamba + topology\/shape prior\uff0c\u5b9e\u9a8c\u4e2d\u5e94\u52a0\u5165 strong 3D U-Net\/nnU-Net\uff0c\u5e76\u62a5\u544a Dice\u3001HD\u3001ASSD\uff0c\u800c\u4e0d\u662f\u53ea\u62a5\u544a\u5355\u4e00 Dice\u3002<\/li>\n<li><strong>\u4e0d\u8981\u628a\u4e24\u7bc7\u90fd\u5f53\u4f5c\u76f4\u63a5 polyp backbone<\/strong>\uff1aSemi-MedRef \u66f4\u504f\u6587\u672c\/\u534a\u76d1\u7763\u8bad\u7ec3\u7b56\u7565\uff0cshape prior \u66f4\u504f\u5fc3\u810f 3D \u8d1f\u7ed3\u679c\uff1b\u5bf9\u606f\u8089\u5206\u5272\u53ef\u501f\u9274\u7684\u662f\u8fb9\u754c\/\u4f2a\u6807\u7b7e\/\u589e\u5f3a\u9a8c\u8bc1\u903b\u8f91\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u590d\u73b0\u5b8c\u6574\u6846\u67b6\u3002<\/li>\n<\/ol>\n<h2>\u53c2\u8003\u94fe\u63a5<\/h2>\n<ul>\n<li>Semi-MedRef arXiv\uff1ahttps:\/\/arxiv.org\/abs\/2605.15720<\/li>\n<li>Semi-MedRef PDF\uff1ahttps:\/\/arxiv.org\/pdf\/2605.15720<\/li>\n<li>Evaluation of Anatomical Shape Priors arXiv\uff1ahttps:\/\/arxiv.org\/abs\/2605.15707<\/li>\n<li>Evaluation of Anatomical Shape Priors PDF\uff1ahttps:\/\/arxiv.org\/pdf\/2605.15707<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u4eca\u65e5\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6700\u65b0\u8bba\u6587\u7cbe\u8bfb\u8ffd\u8e2a \u4eca\u65e5\u7ed3\u8bba \u4eca\u5929\u6ca1\u6709\u68c0\u7d22\u5230\u660e\u786e\u5df2\u63a5\u6536 MICCAI\/CVPR\/ICCV\/ECCV\/NeurIPS\/ &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[85],"tags":[],"class_list":["post-1068","post","type-post","status-publish","format-standard","hentry","category-85"],"views":4,"_links":{"self":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1068","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/comments?post=1068"}],"version-history":[{"count":0,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1068\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/media?parent=1068"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/categories?post=1068"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/tags?post=1068"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}