{"id":1071,"date":"2026-05-22T10:45:02","date_gmt":"2026-05-22T02:45:02","guid":{"rendered":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/22\/2026-05-22-%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%9augcp-%e4%b8%8e-panguide3d\/"},"modified":"2026-05-22T10:45:03","modified_gmt":"2026-05-22T02:45:03","slug":"2026-05-22-%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%9augcp-%e4%b8%8e-panguide3d","status":"publish","type":"post","link":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/22\/2026-05-22-%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%9augcp-%e4%b8%8e-panguide3d\/","title":{"rendered":"2026-05-22 \u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u7cbe\u8bfb\uff1aUGCP \u4e0e PanGuide3D"},"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\u53d1\u73b0\u6765\u81ea MICCAI\/CVPR\/\u9876\u520a\u5b98\u7f51\u4e14\u6bd4\u6628\u65e5\u66f4\u6210\u719f\u7684\u6b63\u5f0f\u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\uff0c\u56e0\u6b64\u5728 2026 \u5e74 5 \u6708 arXiv \u6700\u65b0\u5019\u9009\u4e2d\u7b5b\u9009\u51fa 2 \u7bc7\u66f4\u503c\u5f97\u8ddf\u8e2a\u7684 preprint\uff1a\u4e00\u7bc7\u662f\u9762\u5411\u8840\u7ba1\u5206\u5272\u7684\u53ef\u63d2\u62d4\u7ed3\u6784\u5316\u63a8\u7406\u6a21\u5757 <strong>UGCP<\/strong>\uff0c\u53e6\u4e00\u7bc7\u662f\u9762\u5411\u80f0\u817a\u80bf\u7624 3D CT \u5206\u5272\u8de8\u961f\u5217\u6cdb\u5316\u7684 <strong>PanGuide3D<\/strong>\u3002\u6574\u4f53\u8d8b\u52bf\u662f\uff1a\u533b\u5b66\u5206\u5272\u65b0\u5de5\u4f5c\u6b63\u5728\u4ece\u201c\u53ea\u6362 backbone\u201d\u8f6c\u5411\u201c\u5728\u9884\u6d4b\u9636\u6bb5\/\u89e3\u7801\u9636\u6bb5\u663e\u5f0f\u52a0\u5165\u7ed3\u6784\u5148\u9a8c\u3001\u5668\u5b98\u5148\u9a8c\u3001\u4e0d\u786e\u5b9a\u6027\u548c\u8de8\u961f\u5217\u53ef\u9760\u6027\u8bc4\u4f30\u201d\u3002<\/p>\n<h2>\u68c0\u7d22\u8bf4\u660e<\/h2>\n<p>\u672c\u6b21\u68c0\u7d22\u8986\u76d6 arXiv \u6700\u65b0\u63d0\u4ea4\u3001\u533b\u5b66\u56fe\u50cf\u5206\u5272\u76f8\u5173\u5173\u952e\u8bcd\uff08medical image segmentation\u3001vessel segmentation\u30013D segmentation\u3001pancreas tumor segmentation\u3001Mamba\u3001foundation model\u3001polyp\/organ segmentation\uff09\u4ee5\u53ca\u5386\u53f2 cron \u8f93\u51fa\u3002\u7531\u4e8e\u5f53\u5929\u672a\u68c0\u7d22\u5230\u66f4\u9ad8\u8d28\u91cf\u4e14\u672a\u91cd\u590d\u7684\u6b63\u5f0f\u9876\u4f1a\/\u9876\u520a\u65b0\u8bba\u6587\uff0c\u4eca\u65e5\u4ece 2026 \u5e74 4\u20135 \u6708 arXiv preprint \u4e2d\u9009\u62e9\u4e24\u7bc7\uff1b\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\u8df3\u8fc7\u7684\u5386\u53f2\u91cd\u590d\u5019\u9009\u5305\u62ec\uff1aBeyond Euclidean Prototypes \/ SGP-Net\u3001Harmonized Feature Conditioning and Frequency-Prompt Personalization\u3001Patch-MoE Mamba\u3001MedCore\u3001USEMA \u7b49\u3002<\/p>\n<h2>WordPress \u53d1\u5e03<\/h2>\n<ul>\n<li>WordPress \u6587\u7ae0\u94fe\u63a5\uff1ahttps:\/\/www.eutaboo.com\/index.php\/2026\/05\/22\/2026-05-22-%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%9augcp-%e4%b8%8e-panguide3d\/<\/li>\n<li>WordPress Post ID\uff1a1071<\/li>\n<\/ul>\n<hr \/>\n<h2>\u8bba\u6587 1\uff1aUncertainty-Guided Conservative Propagation for Structured Inference in Vessel Segmentation<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aUncertainty-Guided Conservative Propagation for Structured Inference in Vessel Segmentation<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aHuan Huang, Michele Esposito, Chen Zhao \/ Huan Huang<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-19 submitted<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint\uff0carXiv:2605.20543<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.20543<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.20543<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/chenzhao2023\/UGCP_PR\uff08\u8bba\u6587\u4e2d\u5199 UGC_PR\uff0c\u8bbf\u95ee\u65f6\u91cd\u5b9a\u5411\u5230 UGCP_PR\uff09<\/li>\n<li>\u4efb\u52a1\uff1a2D\/3D vessel segmentation\uff1b\u7ed3\u6784\u8fde\u7eed\u6027\u3001\u4e2d\u5fc3\u7ebf\u4e00\u81f4\u6027\u3001\u8fb9\u754c\u8bef\u5dee\u4fee\u6b63<\/li>\n<li>\u6570\u636e\u96c6\uff1aFIVES fundus vessel\u3001ICA invasive coronary angiography\u3001ImageCAS CCTA\u3001COSTA TOF-MRA<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1abackbone-agnostic structured inference module\uff1buncertainty-guided logit-space propagation\uff1bevidential uncertainty\uff1bstructure-aware local update<\/li>\n<\/ul>\n<h3>paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n<h4>1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n<p>UGCP \u7684\u6838\u5fc3\u4ef7\u503c\u662f\u628a\u8840\u7ba1\u5206\u5272\u4ece\u201c\u4e00\u6b21\u524d\u5411\u8f93\u51fa mask\u201d\u6539\u5199\u4e3a <strong>\u6709\u9650\u6b65 logit-space \u7ed3\u6784\u5316\u66f4\u65b0<\/strong>\uff1a\u8ba9\u4f4e\u4e0d\u786e\u5b9a\u6027\u533a\u57df\u5411\u90bb\u8fd1\u9ad8\u4e0d\u786e\u5b9a\u6027\u8840\u7ba1\u533a\u57df\u4f20\u64ad\uff0c\u540c\u65f6\u7528\u8fb9\u754c\u8c03\u5236\u548c source anchor \u9632\u6b62\u8fc7\u5ea6\u6269\u6563\uff0c\u56e0\u6b64\u5bf9\u7ec6\u957f\u3001\u5206\u53c9\u3001\u6613\u65ad\u88c2\u7684\u8840\u7ba1\u7ed3\u6784\u6bd4\u666e\u901a\u540e\u5904\u7406\u66f4\u6709\u9488\u5bf9\u6027\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76\u8840\u7ba1\u5206\u5272\uff0c\u5305\u62ec\u773c\u5e95\u8840\u7ba1\u3001\u51a0\u8109\u9020\u5f71\u3001CCTA \u51a0\u72b6\u52a8\u8109\u548c TOF-MRA \u8111\u8840\u7ba1\u3002\u8840\u7ba1\u4e0e\u666e\u901a\u5668\u5b98\/\u75c5\u7076\u5206\u5272\u4e0d\u540c\uff1a\u76ee\u6807\u901a\u5e38\u7ec6\u957f\u3001\u9ad8\u5206\u53c9\u3001\u5c40\u90e8\u5bf9\u6bd4\u4f4e\uff0c\u8bc4\u4ef7\u4e0d\u4ec5\u770b\u533a\u57df Dice\uff0c\u8fd8\u5f3a\u4f9d\u8d56\u8fde\u901a\u6027\u3001\u4e2d\u5fc3\u7ebf\u548c\u62d3\u6251\u7ed3\u6784\u3002\u8bba\u6587\u660e\u786e\u6307\u51fa\uff0c\u8bb8\u591a CNN\/Transformer \u5206\u5272\u5668\u867d\u7136\u80fd\u8f93\u51fa\u50cf\u7d20\u7ea7\u6982\u7387\uff0c\u4f46\u63a8\u7406\u9636\u6bb5\u4ecd\u662f one-shot mapping\uff1b\u5f53\u5c40\u90e8\u56fe\u50cf\u8bc1\u636e\u5f31\u3001\u566a\u58f0\u5f3a\u6216\u8840\u7ba1\u91cd\u53e0\u65f6\uff0c\u6a21\u578b\u5bb9\u6613\u4ea7\u751f\u65ad\u88c2\u5206\u652f\u3001\u4f2a\u8fde\u63a5\u548c\u8fb9\u754c\u6f02\u79fb\u3002<\/p>\n<p><strong>\u5185\u90e8 paper map\uff1a<\/strong> \u672c\u6587\u7814\u7a76\u8840\u7ba1\u5206\u5272\u4e2d one-shot prediction \u96be\u4ee5\u4fee\u590d\u5c40\u90e8\u4e0d\u786e\u5b9a\u4e0e\u7ed3\u6784\u4e0d\u8fde\u7eed\u7684\u95ee\u9898\uff0c\u8bbe\u5b9a\u662f 2D\/3D \u4e8c\u503c\u8840\u7ba1\u5206\u5272\u3002\u4e3b\u62db\u662f\u5728 segmentation logit head \u540e\u63d2\u5165 UGCP\uff0c\u628a\u521d\u59cb logits \u89c6\u4e3a\u72b6\u6001 <code>s(0)<\/code>\uff0c\u901a\u8fc7 uncertainty-guided flux\u3001structure-aware edge modulation \u548c source stabilization \u505a <code>T=2<\/code> \u6b65\u5c40\u90e8\u66f4\u65b0\u3002\u5b83\u58f0\u79f0\u53ef\u4f5c\u4e3a U-Net \/ SwinUNETR \u7684\u901a\u7528\u63d2\u4ef6\uff0c\u5728\u56db\u4e2a\u516c\u5f00\u6570\u636e\u96c6\u4e0a\u63d0\u5347 DSC\u3001clDice\u3001HD95\uff1b\u8bc1\u636e\u4e3b\u8981\u662f\u4e94\u6298\u4ea4\u53c9\u9a8c\u8bc1\u4e3b\u8868\u3001\u7ec4\u4ef6\u6d88\u878d\u3001\u6b65\u6570\/\u6b65\u957f\u6d88\u878d\u3001\u4e0e CRF \u540e\u5904\u7406\u6bd4\u8f83\u3001\u590d\u6742\u5ea6\u8868\u548c\u4e0d\u786e\u5b9a\u6027\u53ef\u89c6\u5316\u3002\u771f\u6b63\u8d1f\u8f7d\u5728\u201c\u4e0d\u786e\u5b9a\u6027\u662f\u5426\u80fd\u53ef\u9760\u6307\u793a\u53ef\u4f20\u64ad\u533a\u57df\u201d\u4ee5\u53ca\u201cflux\/source \u8bbe\u8ba1\u662f\u5426\u786e\u5b9e\u6bd4\u5e73\u6ed1\/CRF \u66f4\u7a33\u201d\u3002\u4e3b\u8981\u98ce\u9669\u662f\u76ee\u524d\u51e0\u4e4e\u53ea\u9a8c\u8bc1\u4e8c\u503c\u8840\u7ba1\u4efb\u52a1\uff0c\u4e14\u4e25\u91cd\u6f0f\u68c0\u533a\u57df\u4ecd\u4f9d\u8d56\u521d\u59cb backbone\uff0c\u4e0d\u4e00\u5b9a\u80fd\u51ed\u4f20\u64ad\u6062\u590d\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u628a\u5df2\u6709\u65b9\u6cd5\u4e0d\u8db3\u62c6\u6210\u4e09\u7c7b\uff1a<\/p>\n<ol>\n<li><strong>\u7279\u5f81\u589e\u5f3a\u4e0d\u80fd\u4fdd\u8bc1\u9884\u6d4b\u7ea7\u7ed3\u6784\u6b63\u786e\u3002<\/strong> U-Net\u3001attention U-Net\u3001Transformer vessel segmentation \u53ef\u4ee5\u6269\u5927\u611f\u53d7\u91ce\u6216\u6539\u5584\u591a\u5c3a\u5ea6\u7279\u5f81\uff0c\u4f46\u6700\u7ec8\u4ecd\u76f4\u63a5\u6620\u5c04\u5230 logits\/probabilities\uff1b\u5c40\u90e8\u8bc1\u636e\u4e0d\u8db3\u65f6\uff0c\u9884\u6d4b\u9519\u8bef\u4f1a\u4fdd\u7559\u5230\u6700\u7ec8 mask\u3002<\/li>\n<li><strong>\u7ed3\u6784\u5148\u9a8c\u5e38\u505c\u7559\u5728\u8bad\u7ec3\u635f\u5931\u6216\u540e\u5904\u7406\u3002<\/strong> \u62d3\u6251\u635f\u5931\u3001centerline loss\u3001CRF\u3001\u8fde\u901a\u6027\u540e\u5904\u7406\u80fd\u5e2e\u52a9\u7ed3\u6784\u4e00\u81f4\u6027\uff0c\u4f46\u5f88\u591a\u4e0d\u662f\u6a21\u578b\u5185\u90e8\u63a8\u7406\u673a\u5236\uff0c\u65e0\u6cd5\u5728\u63a8\u7406\u65f6\u6839\u636e\u5f53\u524d\u4e0d\u786e\u5b9a\u533a\u57df\u81ea\u9002\u5e94\u66f4\u65b0\u3002<\/li>\n<li><strong>\u4e0d\u786e\u5b9a\u6027\u4f30\u8ba1\u6ca1\u6709\u8fdb\u5165\u66f4\u65b0\u8fc7\u7a0b\u3002<\/strong> \u8bb8\u591a uncertainty-aware \u65b9\u6cd5\u53ea\u628a\u4e0d\u786e\u5b9a\u6027\u7528\u4e8e\u8bad\u7ec3\u6b63\u5219\u3001\u7f6e\u4fe1\u5ea6\u8bc4\u4f30\u6216\u534a\u76d1\u7763\u7b5b\u9009\uff0c\u800c\u4e0d\u662f\u8ba9\u53ef\u9760\u4f4d\u7f6e\u5b9e\u9645\u53bb\u201c\u652f\u63f4\u201d\u90bb\u8fd1\u6a21\u7cca\u4f4d\u7f6e\u3002<\/li>\n<\/ol>\n<p>\u56e0\u6b64\uff0c\u8bba\u6587\u8981\u89e3\u51b3\u7684\u662f\uff1a\u5982\u4f55\u5728\u4e0d\u6362 backbone \u7684\u524d\u63d0\u4e0b\uff0c\u628a\u5c40\u90e8\u53ef\u9760\u6027\u3001\u8fb9\u754c\u7ed3\u6784\u548c\u9884\u6d4b\u7a33\u5b9a\u6027\u6574\u5408\u6210\u4e00\u4e2a\u53ef\u8bad\u7ec3\u3001\u53ef\u5fae\u3001\u63a8\u7406\u65f6\u5de5\u4f5c\u7684\u7ed3\u6784\u5316\u66f4\u65b0\u6a21\u5757\u3002<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>UGCP \u63d2\u5728 segmentation head \u4e4b\u540e\u3002backbone <code>B(x)<\/code> \u5148\u63d0\u53d6 feature <code>h<\/code>\uff0c\u5206\u7c7b\u5934 <code>W_s<\/code> \u5f97\u5230\u521d\u59cb logit state\uff1a<\/p>\n<p><code>s(0) = W_s(h)<\/code><\/p>\n<p>\u4f20\u7edf\u65b9\u6cd5\u5728\u8fd9\u91cc\u76f4\u63a5 softmax\/sigmoid \u8f93\u51fa\uff1bUGCP \u5219\u6267\u884c\u6709\u9650\u6b65\u66f4\u65b0\uff1a<\/p>\n<p><code>s(t+1) = s(t) + \u03b8 U(s(t)),  t = 0,...,T-1<\/code><\/p>\n<p>\u5176\u4e2d <code>U<\/code> \u662f\u5c40\u90e8\u90bb\u57df\u66f4\u65b0\u7b97\u5b50\u3002\u8bba\u6587\u628a\u5b83\u89e3\u91ca\u6210 conservation-inspired flux-balance\uff1a\u5f53\u524d\u4f4d\u7f6e\u7684\u66f4\u65b0\u6765\u81ea\u90bb\u5c45\u6d41\u5165\u3001\u6d41\u51fa\u5dee\u5f02\uff0c\u4ee5\u53ca\u4e00\u4e2a\u62c9\u56de\u521d\u59cb\u9884\u6d4b\u7684 source term\u3002<\/p>\n<p>\u5173\u952e\u6b65\u9aa4\u5982\u4e0b\uff1a<\/p>\n<ol>\n<li><strong>Evidential uncertainty\uff1a<\/strong> \u5c06 logits \u6620\u5c04\u5230 Dirichlet concentration <code>\u03b1 = softplus(s)+1<\/code>\uff0c\u671f\u671b\u6982\u7387\u4e3a <code>\u03c0_k = \u03b1_k \/ \u03a3_j \u03b1_j<\/code>\uff0c\u4e0d\u786e\u5b9a\u6027\u4e3a <code>u = K \/ (\u03a3_j \u03b1_j + \u03b5)<\/code>\uff1b\u8bc1\u636e\u8d8a\u5c11\uff0c\u4e0d\u786e\u5b9a\u6027\u8d8a\u9ad8\u3002<\/li>\n<li><strong>Uncertainty-guided directional gate\uff1a<\/strong> \u82e5\u4f4d\u7f6e <code>p<\/code> \u6bd4\u90bb\u5c45 <code>q<\/code> \u66f4\u4e0d\u786e\u5b9a\uff0c\u5219 <code>q\u2192p<\/code> \u7684\u4f20\u64ad\u95e8\u66f4\u5927\uff0c\u8ba9\u53ef\u9760\u90bb\u5c45\u5411\u6a21\u7cca\u4f4d\u7f6e\u4f20\u64ad\u3002<\/li>\n<li><strong>Structure-aware edge modulation\uff1a<\/strong> \u4ece backbone feature \u5f97\u5230 decision-aligned feature <code>f = W_f(h)<\/code>\uff0c\u7528\u76f8\u90bb\u7279\u5f81\u5dee\u5f02\u4ea7\u751f <code>\u03c6_{p,q}<\/code>\uff0c\u51cf\u5c11\u8de8\u8fb9\u754c\u3001\u8de8\u7ed3\u6784\u7684\u9519\u8bef\u4f20\u64ad\u3002<\/li>\n<li><strong>Source stabilization\uff1a<\/strong> \u7528 <code>R_p = r_p(s_p(0)-s_p(t))<\/code> \u628a\u66f4\u65b0\u72b6\u6001\u951a\u5b9a\u56de\u521d\u59cb logits\uff0c\u907f\u514d\u591a\u6b65\u8fed\u4ee3\u9020\u6210 excessive drift\u3002<\/li>\n<li><strong>\u6700\u7ec8\u8f93\u51fa\uff1a<\/strong> <code>T<\/code> \u6b65\u540e\u7531 Dirichlet expectation \u5f97\u5230\u524d\u666f\u6982\u7387\u3002<\/li>\n<\/ol>\n<p>\u8bad\u7ec3\u76ee\u6807\u4e3a Dice loss + BCE loss + evidential uncertainty regularization\u3002\u5b9e\u9a8c\u9ed8\u8ba4 <code>T=2, \u03b8=1<\/code>\uff0c2D \u4f7f\u7528 4-neighborhood\uff0c3D \u4f7f\u7528 6-neighborhood\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 A\uff1aEvidential uncertainty from logits<\/strong><br \/>\n- \u8f93\u5165\uff1a\u5f53\u524d logit state <code>s(t)<\/code>\u3002<br \/>\n- \u8f93\u51fa\uff1aDirichlet \u53c2\u6570 <code>\u03b1(t)<\/code>\u3001\u671f\u671b\u6982\u7387 <code>\u03c0(t)<\/code> \u548c uncertainty map <code>u(t)<\/code>\u3002<br \/>\n- \u4f5c\u7528\uff1a\u907f\u514d\u53ea\u7528 softmax entropy\uff1bentropy \u53ea\u770b\u5f52\u4e00\u5316\u5206\u5e03\uff0c\u800c Dirichlet evidence \u8fd8\u80fd\u8868\u8fbe\u201c\u8bc1\u636e\u603b\u91cf\u4e0d\u8db3\u201d\u3002<br \/>\n- \u521b\u65b0\u6027\u5224\u65ad\uff1aevidential learning \u672c\u8eab\u4e0d\u662f\u65b0\u6982\u5ff5\uff0c\u4f46\u5c06\u5176\u4f5c\u4e3a\u5c40\u90e8\u4f20\u64ad\u65b9\u5411\u548c source gate \u7684\u63a7\u5236\u4fe1\u53f7\uff0c\u7528\u4e8e vessel structured inference\uff0c\u8bbe\u8ba1\u8f83\u6e05\u695a\u3002<br \/>\n- \u53ef\u8fc1\u79fb\u6027\uff1a\u9002\u5408\u8fc1\u79fb\u5230 polyp \u8fb9\u754c\u6a21\u7cca\u30013D \u5c0f\u75c5\u7076\u5206\u5272\u548c\u4e0d\u786e\u5b9a\u6027\u9a71\u52a8 refinement\uff1b\u4f46\u5982\u679c\u4efb\u52a1\u4e0d\u662f\u7ec6\u957f\u7ed3\u6784\uff0c\u4f20\u64ad\u90bb\u57df\u548c source \u5f3a\u5ea6\u9700\u91cd\u8c03\u3002<\/p>\n<p><strong>\u6a21\u5757 B\uff1aUncertainty-guided flux gate <code>\u03b3<\/code><\/strong><br \/>\n- \u8f93\u5165\uff1a\u76f8\u90bb\u4f4d\u7f6e <code>p,q<\/code> \u7684\u4e0d\u786e\u5b9a\u6027 <code>u_p,u_q<\/code>\u3002<br \/>\n- \u64cd\u4f5c\uff1a\u901a\u8fc7 sigmoid \u6e29\u5ea6\u95e8\u63a7\uff0c\u4f7f\u4f4e\u4e0d\u786e\u5b9a\u6027\u4f4d\u7f6e\u5411\u9ad8\u4e0d\u786e\u5b9a\u6027\u4f4d\u7f6e\u4f20\u64ad\u3002<br \/>\n- \u89e3\u51b3\u95ee\u9898\uff1a\u8ba9\u53ef\u9760\u8840\u7ba1\u54cd\u5e94\u8865\u5168\u90bb\u8fd1\u5f31\u54cd\u5e94\u533a\u57df\uff0c\u5c24\u5176\u662f\u65ad\u88c2\u8840\u7ba1\u6bb5\u3002<br \/>\n- \u5c40\u9650\uff1a\u5982\u679c\u521d\u59cb\u53ef\u9760\u533a\u57df\u672c\u8eab\u662f false positive\uff0c\u4f20\u64ad\u4e5f\u53ef\u80fd\u6269\u6563\u9519\u8bef\uff1b\u56e0\u6b64\u9700\u8981 edge modulation \u548c source anchor\u3002<\/p>\n<p><strong>\u6a21\u5757 C\uff1aStructure-aware edge modulation <code>\u03c6<\/code><\/strong><br \/>\n- \u8f93\u5165\uff1adecision-aligned feature <code>f_p,f_q<\/code>\u3002<br \/>\n- \u64cd\u4f5c\uff1a\u7528 learnable projection \u8861\u91cf\u76f8\u90bb feature discrepancy\uff0c\u5bf9 incoming flux \u505a\u6709\u7b26\u53f7\u8c03\u5236\u3002<br \/>\n- \u89e3\u51b3\u95ee\u9898\uff1a\u51cf\u5c11\u8de8\u80cc\u666f\/\u8840\u7ba1\u8fb9\u754c\u7684 contamination\uff0c\u907f\u514d\u628a\u8840\u7ba1\u54cd\u5e94\u6269\u6563\u5230\u4e0d\u76f8\u5bb9\u533a\u57df\u3002<br \/>\n- \u4e0e attention\/CRF \u7684\u5173\u7cfb\uff1a\u50cf\u4e00\u4e2a\u8f7b\u91cf\u3001\u5c40\u90e8\u3001\u53ef\u5fae\u7684\u7ed3\u6784\u95e8\u63a7\uff1b\u76f8\u6bd4 CRF\uff0c\u5b83\u5d4c\u5165\u8bad\u7ec3\u76ee\u6807\u5e76\u4f5c\u7528\u4e8e logits\uff0c\u800c\u4e0d\u662f\u6982\u7387\u540e\u5904\u7406\u3002<\/p>\n<p><strong>\u6a21\u5757 D\uff1aSource term <code>R<\/code><\/strong><br \/>\n- \u8f93\u5165\uff1a\u521d\u59cb logits <code>s(0)<\/code> \u4e0e\u5f53\u524d logits <code>s(t)<\/code>\u3002<br \/>\n- \u64cd\u4f5c\uff1a\u6839\u636e\u4e0d\u786e\u5b9a\u6027\u95e8\u63a7\u5c06\u72b6\u6001\u62c9\u56de\u521d\u59cb\u9884\u6d4b\u3002<br \/>\n- \u4f5c\u7528\uff1a\u9632\u6b62\u8fed\u4ee3\u4f20\u64ad\u8d8a\u8d70\u8d8a\u8fdc\uff0c\u6291\u5236\u8fc7\u5e73\u6ed1\u548c\u6f02\u79fb\u3002<br \/>\n- \u91cd\u8981\u6027\uff1a\u6d88\u878d\u663e\u793a\u9010\u6b65\u52a0\u5165 <code>\u03b3<\/code>\u3001<code>\u03c6<\/code>\u3001<code>R<\/code> \u540e ICA \u548c ImageCAS \u6307\u6807\u6301\u7eed\u6539\u5584\uff0c\u8bf4\u660e gain \u4e0d\u662f\u7b80\u5355\u90bb\u57df\u5e73\u6ed1\u5e26\u6765\u7684\u3002<\/p>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u6570\u636e\u96c6\u8986\u76d6 2D \u548c 3D\uff1a<\/p>\n<ul>\n<li><strong>FIVES<\/strong>\uff1a\u773c\u5e95\u8840\u7ba1\uff0c800 \u5f20\u56fe\u50cf\uff0c\u8bad\u7ec3 crop \u4e3a 512\u00d7512\u3002<\/li>\n<li><strong>ICA<\/strong>\uff1ainvasive coronary angiography\uff0c616 \u5f20 2D \u51a0\u8109\u9020\u5f71\u56fe\u50cf\u3002<\/li>\n<li><strong>ImageCAS<\/strong>\uff1aCCTA\uff0c1,000 \u4e2a 3D case\uff0c\u8bad\u7ec3 patch 96\u00d796\u00d796\u3002<\/li>\n<li><strong>COSTA<\/strong>\uff1aTOF-MRA\uff0c355 \u4e2a 3D volume\uff0c\u8bad\u7ec3 patch 64\u00d764\u00d764\u3002<\/li>\n<\/ul>\n<p>backbone \u5305\u62ec 2D\/3D U-Net \u548c 2D\/3D SwinUNETR\u3002\u6240\u6709\u6570\u636e\u96c6\u4f7f\u7528 five-fold cross-validation\uff0c\u6307\u6807\u4e3a DSC\u3001clDice\u3001HD95\u3002<\/p>\n<p>\u4e3b\u7ed3\u679c\uff08Table 2\uff09\u663e\u793a UGCP \u5bf9\u6240\u6709 backbone\/dataset \u5747\u6709\u6b63\u5411\u63d0\u5347\uff1a<\/p>\n<ul>\n<li>COSTA + U-Net\uff1aDSC 0.8422\u21920.8542\uff0cclDice 0.8878\u21920.9030\uff0cHD95 5.07\u21923.34\u3002<\/li>\n<li>COSTA + SwinUNETR\uff1aDSC 0.8560\u21920.8790\uff0cclDice 0.8902\u21920.9155\uff0cHD95 5.94\u21922.09\u3002<\/li>\n<li>ImageCAS + U-Net\uff1aDSC 0.8147\u21920.8188\uff0cclDice 0.8553\u21920.8585\uff0cHD95 8.89\u21928.76\u3002<\/li>\n<li>FIVES + U-Net\uff1aDSC 0.8780\u21920.8831\uff0cHD95 62.73\u219250.70 pixels\u3002<\/li>\n<li>ICA + U-Net\uff1aDSC 0.8932\u21920.9045\uff0cHD95 13.57\u219210.87 pixels\u3002<\/li>\n<\/ul>\n<p>\u5173\u952e\u6d88\u878d\uff1a<\/p>\n<ul>\n<li><strong>\u4ec5\u52a0 evidential loss \u4e0d\u591f\u3002<\/strong> Table 3 \u663e\u793a\u4ece <code>Lbase<\/code> \u6362\u5230 <code>LUGCP<\/code> \u4f46\u4e0d\u505a UGCP \u4f20\u64ad\uff0cICA\/ImageCAS \u53cd\u800c\u7565\u964d\uff1b\u771f\u6b63\u589e\u76ca\u6765\u81ea uncertainty-regulated structured update\u3002<\/li>\n<li><strong>\u7ec4\u4ef6\u9010\u6b65\u6709\u6548\u3002<\/strong> Table 4 \u4e2d\u4ece naive propagation \u5230\u52a0 <code>\u03b3<\/code>\u3001\u518d\u52a0 <code>\u03c6<\/code>\u3001\u518d\u52a0 <code>R<\/code>\uff0cICA \u548c ImageCAS \u7684 DSC\/clDice\/HD95 \u9010\u6b65\u6539\u5584\u3002<\/li>\n<li><strong>\u66f4\u65b0\u6b65\u6570\u4e0d\u80fd\u592a\u591a\u3002<\/strong> Table 5 \u663e\u793a <code>T=2<\/code> \u6700\u597d\uff1b<code>T=3\/4<\/code> \u53ef\u80fd\u8fc7\u4f20\u64ad\u3002<\/li>\n<li><strong>\u4f18\u4e8e CRF \u540e\u5904\u7406\u3002<\/strong> COSTA \u4e0a U-Net baseline + CRF \u5f97 DSC 0.8435\u3001clDice 0.8992\u3001HD95 5.01\uff1bbaseline + UGCP \u5f97 DSC 0.8542\u3001clDice 0.9030\u3001HD95 3.34\u3002<\/li>\n<li><strong>\u8ba1\u7b97\u5f00\u9500\uff1a<\/strong> 2D U-Net \u53c2\u6570\u57fa\u672c\u4e0d\u53d8\uff0cGFLOPs 9.520\u21929.646\uff0c\u65f6\u95f4 2.395ms\u21923.893ms\uff1b3D U-Net GFLOPs 22.991\u219224.768\uff0c\u65f6\u95f4 3.201ms\u21926.843ms\u3002\u53c2\u6570\u5f00\u9500\u5f88\u5c0f\uff0c\u4f46 3D runtime \u7ea6\u7ffb\u500d\uff0c\u9700\u8981\u6ce8\u610f\u3002<\/li>\n<\/ul>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u70b9\uff1a<\/p>\n<ul>\n<li>\u8986\u76d6 2D\/3D\u3001\u773c\u5e95\/\u51a0\u8109\/\u8111\u8840\u7ba1\/CCTA\uff0c\u591a\u6a21\u6001\u9a8c\u8bc1\u6bd4\u5355\u4e00 retinal vessel \u66f4\u5f3a\u3002<\/li>\n<li>\u540c\u65f6\u7528 U-Net \u548c SwinUNETR \u8bc1\u660e plug-in \u7279\u6027\uff0c\u4e0d\u53ea\u662f\u7ed1\u5b9a\u67d0\u4e2a backbone\u3002<\/li>\n<li>\u62a5\u544a DSC\u3001clDice\u3001HD95\uff0c\u5176\u4e2d clDice \u5bf9\u8840\u7ba1\u62d3\u6251\u66f4\u76f8\u5173\u3002<\/li>\n<li>\u6709\u4e94\u6298\u4ea4\u53c9\u9a8c\u8bc1\u5747\u503c\u00b1\u6807\u51c6\u5dee\u3001\u7ec4\u4ef6\u6d88\u878d\u3001\u8d85\u53c2\u6570\u6d88\u878d\u3001CRF \u5bf9\u6bd4\u548c\u590d\u6742\u5ea6\u8868\u3002<\/li>\n<\/ul>\n<p>\u9700\u8981\u8c28\u614e\u7684\u70b9\uff1a<\/p>\n<ul>\n<li>\u76ee\u524d\u96c6\u4e2d\u5728 <strong>binary vessel segmentation<\/strong>\uff1b\u4f5c\u8005\u4e5f\u627f\u8ba4\u5c1a\u672a\u9a8c\u8bc1 artery-vein separation\u3001branch-level labeling \u7b49\u591a\u7c7b\u522b\u8840\u7ba1\u4efb\u52a1\u3002<\/li>\n<li>\u5bf9 ImageCAS \u7684 DSC \u63d0\u5347\u5f88\u5c0f\uff08U-Net +0.0041\uff0cSwinUNETR +0.0029\uff09\uff0c\u8bf4\u660e\u5728\u67d0\u4e9b 3D CCTA \u573a\u666f\u589e\u76ca\u4e3b\u8981\u4f53\u73b0\u5728 clDice\/HD95\uff0c\u800c\u4e0d\u662f\u533a\u57df Dice\u3002<\/li>\n<li>UGCP \u4f9d\u8d56\u521d\u59cb backbone\u3002\u5982\u679c\u67d0\u6bb5\u8840\u7ba1\u5b8c\u5168\u6f0f\u68c0\u4e14\u6ca1\u6709\u90bb\u8fd1\u53ef\u9760 evidence\uff0c\u5c40\u90e8\u4f20\u64ad\u65e0\u6cd5\u51ed\u7a7a\u6062\u590d\u3002<\/li>\n<li>\u8bba\u6587\u662f arXiv preprint\uff0c\u867d\u7136\u63d0\u4f9b GitHub \u94fe\u63a5\uff0c\u4f46\u4ecd\u9700\u5b9e\u9645\u8fd0\u884c\u786e\u8ba4\u4ee3\u7801\u5b8c\u6574\u6027\u548c\u590d\u73b0\u6210\u672c\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>\uff1aUGCP \u4e0d\u662f\u66ff\u4ee3 U-Net\uff0c\u800c\u662f\u653e\u5728 logit head \u540e\u7684\u7ed3\u6784\u5316\u63a8\u7406\u63d2\u4ef6\uff1b\u539f\u5219\u4e0a\u53ef\u63a5\u5230 nnU-Net \u8f93\u51fa\u7aef\uff0c\u4f46\u9700\u8981\u9002\u914d nnU-Net \u7684\u591a\u7c7b\u522b\/\u591a\u6807\u7b7e\u8bbe\u7f6e\u3002<\/li>\n<li>\u4e0e <strong>MedNeXt \/ ConvNeXt-like segmentation<\/strong>\uff1a\u53ef\u4f5c\u4e3a decoder\/head \u540e\u5904\u7406\u5f0f\u53ef\u8bad\u7ec3\u6a21\u5757\uff0c\u4e0e backbone \u65e0\u51b2\u7a81\u3002<\/li>\n<li>\u4e0e <strong>UNetR \/ SwinUNETR \/ Transformer-based segmentation<\/strong>\uff1a\u8bba\u6587\u5df2\u7ecf\u5728 SwinUNETR \u4e0a\u9a8c\u8bc1\uff1b\u5b83\u8865\u7684\u662f Transformer \u8f93\u51fa\u540e\u7684\u5c40\u90e8\u7ed3\u6784\u4e00\u81f4\u6027\uff0c\u800c\u4e0d\u662f\u5168\u5c40\u5efa\u6a21\u80fd\u529b\u672c\u8eab\u3002<\/li>\n<li>\u4e0e <strong>Mamba \/ VMamba \/ SegMamba \/ DAMamba<\/strong>\uff1aUGCP \u7684\u5c40\u90e8\u4fdd\u5b88\u4f20\u64ad\u4e0e Mamba \u7684\u957f\u7a0b selective scan \u4e92\u8865\uff1bMamba \u8d1f\u8d23\u5168\u5c40\/\u5e8f\u5217\u4f9d\u8d56\uff0cUGCP \u53ef\u4f5c\u4e3a\u672b\u7aef uncertainty-guided refinement\u3002<\/li>\n<li>\u4e0e <strong>foundation model \/ MedSAM<\/strong>\uff1a\u53ef\u7528\u4e8e SAM\/MedSAM mask logits \u540e\u7684\u7ed3\u6784\u5316 refinement\uff0c\u7279\u522b\u9002\u5408\u8840\u7ba1\u3001\u606f\u8089\u8fb9\u754c\u548c\u8584\u7ed3\u6784\uff0c\u4f46\u9700\u8981\u9a8c\u8bc1\u662f\u5426\u4f1a\u7834\u574f prompt-specific segmentation\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5bf9 polyp segmentation\u3001DAMamba \u6216\u901a\u7528\u533b\u5b66\u5206\u5272\u6846\u67b6\u6709\u4e09\u7c7b\u4ef7\u503c\uff1a<\/p>\n<ol>\n<li><strong>\u4f5c\u4e3a\u8fb9\u754c\/\u8fde\u901a\u6027 refinement \u63d2\u4ef6\u3002<\/strong> \u606f\u8089\u867d\u7136\u4e0d\u662f\u8840\u7ba1\uff0c\u4f46\u4e5f\u5e38\u6709\u8fb9\u754c\u6a21\u7cca\u3001\u5c40\u90e8\u4f4e\u5bf9\u6bd4\u548c\u5047\u9633\u6027\u533a\u57df\uff1b\u53ef\u628a UGCP \u7684 uncertainty-guided logit update \u6539\u6210\u8fb9\u754c\u4e0d\u786e\u5b9a\u533a\u57df refinement\u3002<\/li>\n<li><strong>\u4f5c\u4e3a DAMamba \u6539\u9020\u601d\u8def\u3002<\/strong> DAMamba\/Mamba \u5206\u652f\u8f93\u51fa\u540e\uff0c\u53ef\u52a0\u5165\u4e00\u4e2a\u8f7b\u91cf UGCP head\uff1a\u7528 Mamba feature \u4f30\u8ba1 <code>\u03c6<\/code>\uff0c\u7528 logits evidence \u4f30\u8ba1 <code>u<\/code>\uff0c\u4ec5\u505a 1\u20132 \u6b65\u66f4\u65b0\uff0c\u91cd\u70b9\u89c2\u5bdf boundary Dice\u3001HD95\u3001clDice \u6216 connected components\u3002<\/li>\n<li><strong>\u4f5c\u4e3a\u5b9e\u9a8c\u8bc4\u4ef7\u53c2\u8003\u3002<\/strong> \u5982\u679c\u8bba\u6587\u58f0\u79f0\u6539\u5584\u7ed3\u6784\u4e00\u81f4\u6027\uff0c\u4e0d\u5e94\u53ea\u62a5 Dice\uff1bUGCP \u7684 clDice\u3001HD95\u3001CRF \u5bf9\u6bd4\u3001\u6b65\u6570\u6d88\u878d\u503c\u5f97\u501f\u9274\u3002<\/li>\n<\/ol>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5efa\u8bae\u7cbe\u8bfb\u3002<\/strong> \u5982\u679c\u4f60\u7684\u7814\u7a76\u5173\u6ce8 Mamba\/Transformer-based medical segmentation \u7684\u540e\u7aef refinement\u3001\u8fb9\u754c\u4fee\u6b63\u6216\u4e0d\u786e\u5b9a\u6027\u5efa\u6a21\uff0cUGCP \u7684\u65b9\u6cd5\u90e8\u5206\u548c\u6d88\u878d\u5f88\u503c\u5f97\u8bfb\u3002\u4f18\u5148\u8bfb Section 3\u3001Table 2\u20138\u3001Limitations\uff1bRelated Work \u53ef\u7565\u8bfb\u3002\u5b83\u4e0d\u662f\u901a\u7528 SOTA backbone\uff0c\u4f46\u4f5c\u4e3a\u201cprediction-level structured inference module\u201d\u5f88\u9002\u5408\u8fc1\u79fb\u548c\u6539\u9020\u3002<\/p>\n<hr \/>\n<h2>\u8bba\u6587 2\uff1aPanGuide3D: Cohort-Robust Pancreas Tumor Segmentation via Probabilistic Pancreas Conditioning and a Transformer Bottleneck<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aPanGuide3D: Cohort-Robust Pancreas Tumor Segmentation via Probabilistic Pancreas Conditioning and a Transformer Bottleneck<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aSunny Joy Ma, Xiang Ma \/ Sunny Joy Ma<\/li>\n<li>\u65f6\u95f4\uff1a2026-04-22 submitted<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint\uff0carXiv:2604.20981<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2604.20981<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2604.20981<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u672a\u516c\u5f00\uff1b\u8bba\u6587\u5199\u660e code available from corresponding author upon reasonable request\uff0c\u672a\u6765\u53ef\u80fd\u516c\u5f00<\/li>\n<li>\u4efb\u52a1\uff1a3D contrast-enhanced CT pancreas and pancreatic tumor segmentation\uff1bcross-cohort generalization<\/li>\n<li>\u6570\u636e\u96c6\uff1aPanTS 2025 pancreatic tumor segmentation dataset\uff1bMSD Task07 Pancreas<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1annU-Net-style 3D U-Net\uff1bdual decoder\uff1bprobabilistic pancreas conditioning\uff1bmulti-scale soft gating\uff1blightweight Transformer bottleneck<\/li>\n<\/ul>\n<h3>paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n<h4>1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n<p>PanGuide3D \u7684\u4ef7\u503c\u4e0d\u5728\u4e8e\u63d0\u51fa\u590d\u6742\u65b0 backbone\uff0c\u800c\u5728\u4e8e\u7528 <strong>\u6982\u7387\u80f0\u817a\u56fe\u4f5c\u4e3a\u591a\u5c3a\u5ea6\u8f6f\u89e3\u5256\u5148\u9a8c<\/strong> \u7ea6\u675f\u80bf\u7624 decoder\uff0c\u5e76\u628a\u8bc4\u4f30\u91cd\u70b9\u653e\u5728 PanTS\u2192MSD \u8de8\u961f\u5217\u6cdb\u5316\u3001\u5c0f\u75c5\u7076\u654f\u611f\u6027\u548c false positive \u6291\u5236\u4e0a\uff0c\u8fd9\u6bd4\u53ea\u5728\u5355\u6570\u636e\u96c6\u5237 Dice \u66f4\u63a5\u8fd1\u4e34\u5e8a 3D \u80bf\u7624\u5206\u5272\u7684\u771f\u5b9e\u75db\u70b9\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76 contrast-enhanced CT \u4e2d\u80f0\u817a\u4e0e\u80f0\u817a\u80bf\u7624\u5206\u5272\u3002\u80f0\u817a\u80bf\u7624\u5206\u5272\u96be\u70b9\u5305\u62ec\uff1a\u80bf\u7624\u76f8\u5bf9\u5668\u5b98\u5f88\u5c0f\u3001\u589e\u5f3a\u6a21\u5f0f\u5f02\u8d28\u3001\u4e0e\u5468\u56f4\u8f6f\u7ec4\u7ec7\/\u8840\u7ba1\u76f8\u4f3c\u3001\u6b63\u8d1f\u6837\u672c\u6781\u4e0d\u5e73\u8861\uff0c\u5e76\u4e14\u4e0d\u540c\u961f\u5217\u4e4b\u95f4\u5b58\u5728 scanner\u3001contrast timing\u3001slice thickness\u3001\u60a3\u8005\u7fa4\u4f53\u3001\u80bf\u7624\u5927\u5c0f\u5206\u5e03\u548c\u6807\u6ce8\u4e60\u60ef\u5dee\u5f02\u3002\u4f5c\u8005\u5f3a\u8c03\uff0c\u771f\u5b9e\u90e8\u7f72\u4e2d\u6a21\u578b\u4e0d\u53ea\u8981\u5728\u8bad\u7ec3\u961f\u5217\u5185\u8868\u73b0\u597d\uff0c\u8fd8\u8981\u5728\u5916\u90e8\u961f\u5217\u4e0a\u907f\u514d\u6f0f\u68c0\u5c0f\u80bf\u7624\u548c\u4ea7\u751f\u89e3\u5256\u4e0a\u4e0d\u5408\u7406\u7684\u5047\u9633\u6027\u3002<\/p>\n<p><strong>\u5185\u90e8 paper map\uff1a<\/strong> \u672c\u6587\u7814\u7a76 3D \u80f0\u817a\u80bf\u7624 CT \u5206\u5272\u5728 cohort shift \u4e0b\u7684\u53ef\u9760\u6027\u95ee\u9898\uff0c\u8bbe\u5b9a\u662f PanTS \u8bad\u7ec3\/\u6d4b\u8bd5\u4e0e\u5916\u90e8 MSD Task07 \u6d4b\u8bd5\u3002\u4e3b\u62db\u662f nnU-Net-style shared 3D encoder + pancreas decoder + pancreas-conditioned tumor decoder\uff0c\u5e76\u5728 bottleneck \u52a0\u8f7b\u91cf Transformer\u3002\u5b83\u58f0\u79f0\u6982\u7387\u6027 organ conditioning \u4e0e global context \u4e92\u8865\uff0c\u53ef\u63d0\u5347 tumor Dice\u3001tumor sensitivity\u3001patient sensitivity \u5e76\u964d\u4f4e false-positive volume\uff1b\u8bc1\u636e\u4e3b\u8981\u662f PanTS\/MSD \u4e3b\u7ed3\u679c\u8868\u30012Decoder\/TransBNeck \u6d88\u878d\u3001\u5c3a\u5bf8\/\u90e8\u4f4d\u5206\u5c42\u5206\u6790\u3001\u5b9a\u6027\u56fe\u548c calibration plot\u3002\u771f\u6b63\u8d1f\u8f7d\u5728\u201c\u80f0\u817a\u6982\u7387\u56fe\u662f\u5426\u80fd\u4f5c\u4e3a\u8f6f\u5148\u9a8c\u7a33\u5b9a\u80bf\u7624\u9884\u6d4b\u201d\u4ee5\u53ca\u201c\u8de8\u961f\u5217\u8bc4\u4f30\u662f\u5426\u516c\u5e73\u201d\u3002\u4e3b\u8981\u98ce\u9669\u662f\u4ee3\u7801\u672a\u516c\u5f00\u3001\u5b9e\u9a8c\u53ea\u6709\u4e24\u4e2a\u516c\u5f00\u961f\u5217\uff0c\u4e14 nnU-Net baseline \u8868\u73b0\u5f02\u5e38\u4f4e\uff0c\u9700\u786e\u8ba4\u8bad\u7ec3\/\u9608\u503c\/\u4efb\u52a1\u5b9a\u4e49\u662f\u5426\u5b8c\u5168\u516c\u5e73\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u6307\u51fa\u51e0\u7c7b\u4e0d\u8db3\uff1a<\/p>\n<ol>\n<li><strong>\u786c\u7ea7\u8054\/ROI \u65b9\u6cd5\u5bb9\u6613\u9519\u8bef\u4f20\u64ad\u3002<\/strong> \u5148\u5206\u5272\u80f0\u817a\u518d\u88c1\u526a ROI \u6216\u7b2c\u4e8c\u9636\u6bb5\u5206\u5272\u80bf\u7624\uff0c\u53ef\u4ee5\u51cf\u5c11\u641c\u7d22\u7a7a\u95f4\uff0c\u4f46\u5982\u679c\u80f0\u817a mask \u9519\u8bef\u6216 ROI \u592a\u786c\uff0c\u80bf\u7624 decoder \u4f1a\u88ab\u4e0a\u6e38\u9519\u8bef\u9650\u5236\u3002<\/li>\n<li><strong>\u666e\u901a multi-head U-Net \u6ca1\u6709\u663e\u5f0f\u6ce8\u5165\u5668\u5b98\u4e0d\u786e\u5b9a\u6027\u3002<\/strong> \u591a\u4efb\u52a1\u5171\u4eab encoder \u53ef\u4ee5\u540c\u65f6\u5b66 pancreas\/tumor\uff0c\u4f46 tumor head \u4e0d\u4e00\u5b9a\u77e5\u9053\u201c\u54ea\u91cc\u662f\u8f83\u53ef\u80fd\u7684\u80f0\u817a\u533a\u57df\u201d\u3002<\/li>\n<li><strong>\u7eaf CNN \u5c40\u90e8\u6027\u4e0d\u8db3\u4ee5\u5904\u7406 cohort shift\u3002<\/strong> \u80f0\u817a\u80bf\u7624\u9519\u8bef\u5e38\u6d89\u53ca\u5168\u5c40\u89e3\u5256\u4e0a\u4e0b\u6587\uff0c\u5982\u628a\u8fdc\u79bb\u80f0\u817a\u7684 bowel enhancement \u5f53\u6210\u80bf\u7624\uff1b\u4ec5\u4f9d\u8d56\u5c40\u90e8\u7eb9\u7406\u5bb9\u6613\u8de8\u961f\u5217\u5931\u6548\u3002<\/li>\n<li><strong>\u5f88\u591a\u5de5\u4f5c\u53ea\u62a5 in-distribution Dice\u3002<\/strong> \u4f5c\u8005\u8ba4\u4e3a\u5fc5\u987b\u663e\u5f0f\u6d4b\u8bd5 PanTS\u2192MSD \u7684 cohort transfer\uff0c\u5e76\u62a5\u544a detection\u3001false positive\u3001size\/location stratification \u548c calibration\u3002<\/li>\n<\/ol>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>PanGuide3D \u662f\u4e00\u4e2a\u7b80\u5355\u7684 3D encoder-decoder \u6846\u67b6\uff1a<\/p>\n<ol>\n<li><strong>Shared 3D encoder\uff1a<\/strong> nnU-Net-style \u591a\u5206\u8fa8\u7387 3D Conv\u2013Norm\u2013Activation blocks + strided downsampling\u3002<\/li>\n<li><strong>Lightweight Transformer bottleneck\uff1a<\/strong> \u5728\u6700\u6df1\u5c42 feature \u4e0a\u505a\u5168\u5c40\u4e0a\u4e0b\u6587\u805a\u5408\uff0c\u8865\u8db3 CNN \u5bf9\u8fdc\u7a0b\u7ed3\u6784\u5173\u7cfb\u7684\u5efa\u6a21\u4e0d\u8db3\u3002<\/li>\n<li><strong>Pancreas decoder\uff1a<\/strong> \u8f93\u51fa\u6982\u7387\u6027 pancreas map\uff0c\u800c\u4e0d\u662f\u4e8c\u503c\u786c mask\u3002<\/li>\n<li><strong>Tumor decoder with probabilistic conditioning\uff1a<\/strong> tumor decoder \u5728\u591a\u4e2a\u5c3a\u5ea6\u63a5\u6536 pancreas probability map\uff0c\u901a\u8fc7 differentiable soft gating \/ feature modulation \u8ba9\u80bf\u7624\u9884\u6d4b\u66f4\u5173\u6ce8\u80f0\u817a\u4e00\u81f4\u533a\u57df\uff0c\u540c\u65f6\u4fdd\u7559\u4ece\u4e0d\u5b8c\u7f8e pancreas prediction \u4e2d\u6062\u590d\u7684\u5f39\u6027\u3002<\/li>\n<li><strong>Final output\uff1a<\/strong> concatenate pancreas and tumor logits\uff0c\u8054\u5408\u8bad\u7ec3\u4e24\u4e2a channel\u3002<\/li>\n<\/ol>\n<p>\u8bad\u7ec3\u76ee\u6807\u662f pancreas\/tumor \u4e24\u4e2a\u901a\u9053 Dice loss + BCE loss\u3002\u8bad\u7ec3\u91c7\u7528 3D patch\uff08\u9ed8\u8ba4 160\u00d7160\u00d796\uff09\u3001tumor-centered sampling\u3001tumor-positive\/negative weighted sampler\u3001Gaussian noise\/blur\/brightness\/gamma\/low-resolution simulation \u7b49 3D \u533b\u5b66\u589e\u5f3a\u3002\u5b9e\u9a8c\u4f7f\u7528 PyTorch\/MONAI\uff0c8\u00d7A100 80GB\uff0c\u6df7\u5408\u7cbe\u5ea6\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 A\uff1aProbabilistic pancreas conditioning<\/strong><br \/>\n- \u8f93\u5165\uff1apancreas decoder \u4ea7\u751f\u7684\u591a\u5c3a\u5ea6 pancreas probability map\uff0c\u4ee5\u53ca tumor decoder \u5bf9\u5e94\u5c3a\u5ea6 feature\u3002<br \/>\n- \u8f93\u51fa\uff1a\u7ecf\u8fc7 soft anatomical gating\/modulation \u7684 tumor feature\u3002<br \/>\n- \u89e3\u51b3\u95ee\u9898\uff1a\u628a\u201c\u80bf\u7624\u5e94\u4e0e\u80f0\u817a\u89e3\u5256\u4f4d\u7f6e\u4e00\u81f4\u201d\u4f5c\u4e3a\u8f6f\u5148\u9a8c\u6ce8\u5165\uff0c\u800c\u4e0d\u662f\u7528\u786c ROI \u88c1\u526a\u3002<br \/>\n- \u521b\u65b0\u6027\u5224\u65ad\uff1aorgan-guided tumor segmentation \u4e0d\u662f\u65b0\u65b9\u5411\uff0c\u4f46\u672c\u6587\u628a pancreas probability \u4f5c\u4e3a\u591a\u5c3a\u5ea6\u8fde\u7eed\u6761\u4ef6\uff0c\u5e76\u7528\u8de8\u961f\u5217\u8bc4\u4f30\u8bc1\u660e\u5176\u4ef7\u503c\uff0c\u8d21\u732e\u6bd4\u8f83\u5b9e\u7528\u3002<br \/>\n- \u53ef\u8fc1\u79fb\u6027\uff1a\u975e\u5e38\u9002\u5408\u8fc1\u79fb\u5230\u201c\u5668\u5b98+\u75c5\u7076\u201d\u4efb\u52a1\uff0c\u5982\u809d\u810f\/\u80bf\u7624\u3001\u80be\u810f\/\u80bf\u7624\u3001\u606f\u8089\/\u80a0\u8154\u533a\u57df\u5148\u9a8c\uff1b\u5bf9\u7eaf\u606f\u8089\u5206\u5272\u53ef\u8003\u8651\u7528 lumen\/colon wall probability \u4f5c\u8f6f\u5148\u9a8c\u3002<\/p>\n<p><strong>\u6a21\u5757 B\uff1aLightweight Transformer bottleneck<\/strong><br \/>\n- \u8f93\u5165\uff1ashared encoder \u6700\u6df1\u5c42 3D feature\u3002<br \/>\n- \u8f93\u51fa\uff1a\u5e26\u957f\u7a0b\u4e0a\u4e0b\u6587\u7684 bottleneck feature\u3002<br \/>\n- \u89e3\u51b3\u95ee\u9898\uff1a\u907f\u514d\u4ec5\u9760\u5c40\u90e8\u7eb9\u7406\uff0c\u628a\u8fdc\u79bb\u80f0\u817a\u7684\u589e\u5f3a\u7ed3\u6784\u8bef\u5224\u4e3a\u80bf\u7624\uff1b\u63d0\u5347 cohort shift \u4e0b\u7684\u5168\u5c40\u4e00\u81f4\u6027\u3002<br \/>\n- \u521b\u65b0\u6027\u5224\u65ad\uff1aTransformer bottleneck \u662f\u6210\u719f\u7ec4\u4ef6\uff0c\u5355\u72ec\u4e0d\u65b0\uff1b\u4ef7\u503c\u5728\u4e8e\u4e0e soft organ conditioning \u7ec4\u5408\uff0c\u5e76\u901a\u8fc7 TransBNeck ablation \u9a8c\u8bc1 global context \u7684\u72ec\u7acb\u8d21\u732e\u3002<\/p>\n<p><strong>\u6a21\u5757 C\uff1aDual decoder \/ two-task prediction<\/strong><br \/>\n- \u8f93\u5165\uff1a\u5171\u4eab encoder feature\u3002<br \/>\n- \u8f93\u51fa\uff1apancreas logits \u4e0e tumor logits\u3002<br \/>\n- \u4f5c\u7528\uff1apancreas \u662f\u76f8\u5bf9\u5bb9\u6613\u7684\u5927\u7ed3\u6784\uff0ctumor \u662f\u56f0\u96be\u5c0f\u76ee\u6807\uff1b\u5171\u4eab\u4f4e\u5c42\u7279\u5f81\u540c\u65f6\u8ba9 tumor head \u63a5\u6536 organ localization signal\u3002<br \/>\n- \u98ce\u9669\uff1a\u82e5 pancreas prediction \u5728\u5916\u90e8\u961f\u5217\u5931\u8d25\uff0ctumor conditioning \u4e5f\u4f1a\u53d7\u5f71\u54cd\uff1b\u4e0d\u8fc7\u8f6f\u6982\u7387\u6bd4\u786c\u88c1\u526a\u66f4\u6297\u9519\u3002<\/p>\n<p><strong>\u6a21\u5757 D\uff1a\u8de8\u961f\u5217\u8bc4\u4f30\u8bbe\u8ba1<\/strong><br \/>\n- \u8f93\u5165\uff1aPanTS \u548c MSD \u7edf\u4e00 resampling \u5230 1.5mm isotropic\uff0c\u5e76\u6807\u51c6\u5316\u4e3a\u540c\u4e00\u5b58\u50a8\/\u8bc4\u4f30\u683c\u5f0f\u3002<br \/>\n- \u4f5c\u7528\uff1a\u5c06\u6a21\u578b\u5dee\u5f02\u4e0e preprocessing \u5dee\u5f02\u5c3d\u91cf\u89e3\u8026\uff0c\u7a81\u51fa cohort shift \u672c\u8eab\u3002<br \/>\n- \u91cd\u8981\u6027\uff1a\u8fd9\u7bc7\u8bba\u6587\u7684\u5b9e\u9645\u8d21\u732e\u5f88\u5927\u4e00\u90e8\u5206\u5728 evaluation framing\uff0c\u800c\u4e0d\u662f\u67d0\u4e2a\u590d\u6742\u6a21\u5757\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>PanTS<\/strong>\uff1a\u4f7f\u7528 released raw contrast-enhanced CT\uff1b9,000 training cases \u4e2d 7,200 train \/ 1,800 val\uff0cofficial 901 test held out\u3002PanTS train \u4e2d tumor-positive \u4e3a 684\/7200\uff08\u7ea6 9.5%\uff09\uff0ctest \u4e3a 151\/901\uff08\u7ea6 16.8%\uff09\uff0c\u5e73\u5747\u80bf\u7624\u4f53\u79ef\u7ea6 2.6 cm\u00b3\u3002<\/li>\n<li><strong>MSD Task07 Pancreas<\/strong>\uff1a57\/57 tumor-positive\uff0c\u5e73\u5747\u80bf\u7624\u4f53\u79ef\u7ea6 7.12 cm\u00b3\u3002\u7528\u4e8e\u5916\u90e8\u961f\u5217\u6d4b\u8bd5\u3002<\/li>\n<li>\u7edf\u4e00\u9884\u5904\u7406\uff1a1.5mm isotropic resampling\uff0c\u7edf\u4e00 .npz \u8868\u793a\u3002<\/li>\n<li>Baselines\uff1annU-Net\u3001SwinUNETR\u3001RADFM foundation model baseline\uff1b\u6d88\u878d\u4e3a 2Decoder\uff08\u53bb\u6389 Transformer bottleneck\uff0c\u4fdd\u7559 pancreas-conditioned tumor decoder\uff09\u548c TransBNeck\uff08\u4fdd\u7559 Transformer\uff0c\u53bb\u6389\u72ec\u7acb pancreas head \/ conditioning\uff09\u3002\u4e24\u9636\u6bb5 U-Net \u56e0\u65e0\u8bad\u7ec3\u4ee3\u7801\u672a\u590d\u73b0\u3002<\/li>\n<\/ul>\n<p>\u4e3b\u7ed3\u679c\uff08Table 2\uff09\uff1a<\/p>\n<p><strong>MSD \u5916\u90e8\u961f\u5217\uff1a<\/strong><br \/>\n- nnU-Net\uff1aTumor Dice 0.066\uff0cPatient sensitivity 0.068\uff0cFP volume 19.646 cm\u00b3\u3002<br \/>\n- RADFM\uff1aTumor Dice 0.154\uff0cPatient sensitivity 0.333\uff0cFP volume 8.666 cm\u00b3\u3002<br \/>\n- SwinUNETR\uff1aTumor Dice 0.274\uff0cPatient sensitivity 0.602\uff0cFP volume 8.550 cm\u00b3\u3002<br \/>\n- 2Decoder\uff1aTumor Dice 0.410\uff0cPatient sensitivity 0.737\uff0cFP volume 4.781 cm\u00b3\u3002<br \/>\n- TransBNeck\uff1aTumor Dice 0.421\uff0cPatient sensitivity 0.772\uff0cFP volume 4.291 cm\u00b3\u3002<br \/>\n- <strong>PanGuide3D\uff1aTumor Dice 0.475\uff0cTumor sensitivity 0.842\uff0cPatient sensitivity 0.842\uff0cFP volume 2.004 cm\u00b3\u3002<\/strong><\/p>\n<p><strong>PanTS in-cohort test\uff1a<\/strong><br \/>\n- nnU-Net\uff1aTumor Dice 0.202\uff0cPatient sensitivity 0.394\uff0cFP volume 16.970 cm\u00b3\u3002<br \/>\n- SwinUNETR\uff1aTumor Dice 0.336\uff0cPatient sensitivity 0.699\uff0cFP volume 9.883 cm\u00b3\u3002<br \/>\n- 2Decoder\uff1aTumor Dice 0.367\uff0cPatient sensitivity 0.714\u3002<br \/>\n- TransBNeck\uff1aTumor Dice 0.412\uff0cPatient sensitivity 0.731\u3002<br \/>\n- <strong>PanGuide3D\uff1aTumor Dice 0.460\uff0cTumor sensitivity 0.787\uff0cPatient sensitivity 0.819\uff0cFP volume 3.070 cm\u00b3\u3002<\/strong><\/p>\n<p>\u5176\u4ed6\u8bc1\u636e\uff1a<\/p>\n<ul>\n<li>\u80bf\u7624\u5927\u5c0f\u5206\u6790\u663e\u793a\u5c0f\u75c5\u7076\u662f\u4e3b\u8981\u5931\u8d25\u6765\u6e90\uff1bPanGuide3D \u5728\u4f4e\u4f53\u79ef\u533a\u57df\u76f8\u6bd4 nnU-Net \u6709\u66f4\u5c11 near-zero Dice collapse\u3002<\/li>\n<li>\u6309 pancreas head\/body\/tail \u548c small\/medium\/large \u5206\u5c42\u7684 heatmap \u663e\u793a\uff0cPanGuide3D \u5728\u591a\u6570\u56f0\u96be strata \u4e2d\u66f4\u7a33\u3002<\/li>\n<li>\u5b9a\u6027\u56fe\u663e\u793a nnU-Net \u5728\u80f0\u817a\u5916\u4ea7\u751f off-site false positives\uff0c\u800c PanGuide3D \u7684\u80bf\u7624\u9884\u6d4b\u66f4\u8d34\u8fd1\u80f0\u817a\u89e3\u5256\u3002<\/li>\n<li>calibration plot \u663e\u793a PanGuide3D \u6bd4 nnU-Net \u66f4\u63a5\u8fd1\u53ef\u9760\u6027\u5bf9\u89d2\u7ebf\uff0c\u4f46\u6700\u9ad8\u7f6e\u4fe1 bin \u4ecd\u8fc7\u5ea6\u81ea\u4fe1\u3002<\/li>\n<\/ul>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u70b9\uff1a<\/p>\n<ul>\n<li>\u8bba\u6587\u6ca1\u6709\u53ea\u5237\u5355\u6570\u636e\u96c6 Dice\uff0c\u800c\u662f\u628a PanTS\u2192MSD cohort transfer \u4f5c\u4e3a\u6838\u5fc3\u8bc4\u4f30\u3002<\/li>\n<li>\u62a5\u544a tumor Dice\u3001pancreas Dice\u3001false-positive volume\u3001tumor sensitivity\u3001patient sensitivity\uff0c\u6bd4\u5355 Dice \u66f4\u8d34\u8fd1\u4e34\u5e8a\u6f0f\u68c0\/\u8bef\u68c0\u98ce\u9669\u3002<\/li>\n<li>\u6709 2Decoder \u548c TransBNeck \u4e24\u4e2a\u9488\u5bf9\u6838\u5fc3\u8bbe\u8ba1\u7684\u6d88\u878d\uff0c\u8bf4\u660e organ conditioning \u548c Transformer context \u5404\u6709\u8d21\u732e\uff0cfull model \u6700\u5f3a\u3002<\/li>\n<li>\u8ba8\u8bba\u4e86 calibration \u548c\u5c0f\u75c5\u7076\u5206\u5c42\uff0c\u8fd9\u662f 3D \u80bf\u7624\u5206\u5272\u4e2d\u975e\u5e38\u91cd\u8981\u7684\u53ef\u9760\u6027\u7ef4\u5ea6\u3002<\/li>\n<\/ul>\n<p>\u9700\u8981\u8c28\u614e\u7684\u70b9\uff1a<\/p>\n<ul>\n<li><strong>\u4ee3\u7801\u672a\u516c\u5f00<\/strong>\uff0c\u53ea\u80fd\u5411\u4f5c\u8005\u5408\u7406\u8bf7\u6c42\uff1b\u590d\u73b0\u95e8\u69db\u9ad8\u3002<\/li>\n<li>\u4f7f\u7528 8\u00d7A100 80GB\uff0c\u8bad\u7ec3\u8d44\u6e90\u8f83\u91cd\uff0c\u4e0d\u9002\u5408\u4f5c\u4e3a\u8f7b\u91cf\u590d\u73b0\u9879\u76ee\u3002<\/li>\n<li>nnU-Net baseline \u5728 PanTS\/MSD \u4e0a\u8868\u73b0\u975e\u5e38\u4f4e\uff0c\u5c24\u5176 MSD Tumor Dice 0.066\uff0c\u9700\u8981\u8b66\u60d5\u662f\u5426\u4e0e\u9608\u503c\u3001\u91c7\u6837\u3001\u8bad\u7ec3\u914d\u7f6e\u3001class imbalance \u6216\u4efb\u52a1\u5b9a\u4e49\u6709\u5173\uff1b\u867d\u7136\u4f5c\u8005\u8bf4\u4f7f\u7528 matched protocol\uff0c\u4f46\u4ecd\u9700\u7b2c\u4e09\u65b9\u590d\u73b0\u786e\u8ba4\u3002<\/li>\n<li>\u5916\u90e8\u6d4b\u8bd5\u53ea\u6709 MSD \u4e00\u4e2a\u961f\u5217\uff1b\u201ccohort-robust\u201d\u7ed3\u8bba\u6bd4\u5355\u961f\u5217\u5f3a\uff0c\u4f46\u8ddd\u79bb\u591a\u4e2d\u5fc3\u771f\u5b9e\u6cdb\u5316\u4ecd\u6709\u9650\u3002<\/li>\n<li>\u6700\u9ad8\u7f6e\u4fe1 bin \u4ecd over-confident\uff0c\u4e0d\u80fd\u76f4\u63a5\u628a\u6982\u7387\u5f53\u4e34\u5e8a\u53ef\u9760\u6027\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>\uff1aPanGuide3D \u662f nnU-Net-style 3D U-Net \u7684\u4efb\u52a1\u5316\u6539\u9020\uff0c\u4e0d\u662f\u901a\u7528\u81ea\u52a8\u914d\u7f6e\u6846\u67b6\uff1b\u5b83\u5728 decoder \u8bbe\u8ba1\u4e0a\u6bd4\u6807\u51c6 nnU-Net \u66f4\u663e\u5f0f\u5229\u7528 organ prior\u3002<\/li>\n<li>\u4e0e <strong>MedNeXt<\/strong>\uff1a\u6ca1\u6709\u76f4\u63a5\u5bf9\u6bd4\uff1bMedNeXt \u66f4\u5173\u6ce8 ConvNeXt-like backbone\uff0cPanGuide3D \u66f4\u5173\u6ce8 organ-conditioned decoding\u3002\u4e24\u8005\u53ef\u7ec4\u5408\uff1aMedNeXt encoder + pancreas-conditioned tumor decoder\u3002<\/li>\n<li>\u4e0e <strong>UNetR \/ SwinUNETR \/ Transformer segmentation<\/strong>\uff1a\u8bba\u6587\u5bf9\u6bd4 SwinUNETR\uff1bPanGuide3D \u53ea\u5728 bottleneck \u4f7f\u7528\u8f7b\u91cf Transformer\uff0c\u4e0d\u662f\u5168 Transformer \u67b6\u6784\uff0c\u5f3a\u8c03 CNN \u7a33\u5b9a\u6027 + global context\u3002<\/li>\n<li>\u4e0e <strong>TransUNet \/ TransFuse<\/strong>\uff1a\u601d\u60f3\u4e0a\u7c7b\u4f3c hybrid CNN-Transformer\uff0c\u4f46\u9762\u5411 3D CT \u548c organ-tumor multi-task\uff0c\u5e76\u52a0\u5165\u6982\u7387\u5668\u5b98\u5148\u9a8c\u3002<\/li>\n<li>\u4e0e <strong>Mamba \/ VMamba \/ SegMamba \/ DAMamba<\/strong>\uff1a\u6ca1\u6709\u4f7f\u7528 Mamba\uff1b\u4f46\u53ef\u5c06 Transformer bottleneck \u66ff\u6362\u4e3a 3D Mamba \/ VMamba block\uff0c\u7528\u9009\u62e9\u6027\u626b\u63cf\u5efa\u6a21\u957f\u7a0b\u4e0a\u4e0b\u6587\uff0c\u518d\u4fdd\u7559 pancreas probability conditioning\u3002<\/li>\n<li>\u4e0e <strong>foundation model \/ RADFM \/ MedSAM<\/strong>\uff1aRADFM \u4f5c\u4e3a baseline \u8868\u73b0\u4e0d\u5982\u4efb\u52a1\u5316\u6a21\u578b\uff0c\u63d0\u793a\u5728\u5c0f\u75c5\u7076 3D \u80bf\u7624\u5206\u5272\u4e2d\uff0c\u901a\u7528 foundation representation \u4e0d\u4e00\u5b9a\u66ff\u4ee3\u4efb\u52a1\u5148\u9a8c\u548c\u7cbe\u7ec6\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5982\u679c\u4f60\u7684\u65b9\u5411\u662f polyp segmentation\u3001DAMamba \u6216\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u8bbe\u8ba1\uff0c\u8fd9\u7bc7\u8bba\u6587\u6709\u4ee5\u4e0b\u4ef7\u503c\uff1a<\/p>\n<ol>\n<li><strong>\u5668\u5b98\/\u4e0a\u4e0b\u6587\u5148\u9a8c\u7684 decoder \u8bbe\u8ba1\u3002<\/strong> \u5bf9\u606f\u8089\u4efb\u52a1\uff0c\u53ef\u7c7b\u6bd4\u201c\u80a0\u8154\/\u9ecf\u819c\u533a\u57df\u6982\u7387\u56fe\u2192\u606f\u8089 decoder soft gating\u201d\uff1b\u5bf9 3D \u80bf\u7624\u4efb\u52a1\uff0c\u53ef\u7c7b\u6bd4\u201corgan probability\u2192lesion decoder\u201d\u3002<\/li>\n<li><strong>DAMamba \u6539\u9020\u65b9\u5411\u3002<\/strong> \u53ef\u4ee5\u628a PanGuide3D \u7684 Transformer bottleneck \u6362\u6210 Mamba bottleneck\uff0c\u5e76\u4fdd\u6301 dual decoder + soft organ conditioning\uff0c\u6784\u6210\u4e00\u4e2a\u6e05\u6670\u3001\u53ef\u89e3\u91ca\u7684 3D hybrid CNN-Mamba segmentation framework\u3002<\/li>\n<li><strong>\u5b9e\u9a8c\u5199\u4f5c\u4ef7\u503c\u3002<\/strong> \u8bba\u6587\u7684 cross-cohort\u3001small-lesion stratification\u3001false-positive volume\u3001patient sensitivity \u548c calibration \u5206\u6790\u5f88\u9002\u5408\u501f\u9274\u5230 introduction\/experiments\uff0c\u5c24\u5176\u5f53\u4f60\u7684\u8bba\u6587\u60f3\u5f3a\u8c03 robustness \u800c\u4e0d\u662f\u5355 Dice\u3002<\/li>\n<li><strong>\u4f5c\u4e3a baseline \u7684\u8c28\u614e\u6027\u3002<\/strong> \u7531\u4e8e\u4ee3\u7801\u672a\u516c\u5f00\u4e14\u8d44\u6e90\u8981\u6c42\u9ad8\uff0c\u77ed\u671f\u4e0d\u9002\u5408\u4f5c\u4e3a\u5fc5\u987b\u590d\u73b0 baseline\uff1b\u66f4\u9002\u5408\u4f5c\u4e3a related work \u548c\u7ed3\u6784\u8bbe\u8ba1\u53c2\u8003\u3002<\/li>\n<\/ol>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5efa\u8bae\u7cbe\u8bfb\uff0c\u4f46\u4ee5\u201c\u8bbe\u8ba1\u4e0e\u8bc4\u4f30\u601d\u8def\u201d\u4e3a\u4e3b\u3002<\/strong> \u5982\u679c\u4f60\u505a 3D medical image segmentation\u3001organ-tumor \u8054\u5408\u5206\u5272\u6216\u8de8\u961f\u5217\u9c81\u68d2\u6027\uff0c\u8fd9\u7bc7\u503c\u5f97\u8bfb Introduction\u3001Figure 1\u3001Table 2\u3001Methods \u4e2d\u8bad\u7ec3\/\u8bc4\u4f30\u8bbe\u7f6e\u548c Discussion\/Limitations\u3002\u82e5\u4f60\u7684\u77ed\u671f\u76ee\u6807\u662f 2D polyp SOTA\uff0c\u53ef\u7565\u8bfb\u65b9\u6cd5\uff0c\u628a probabilistic conditioning \u548c\u8de8\u57df\u8bc4\u4f30\u4f5c\u4e3a\u53ef\u8fc1\u79fb\u601d\u60f3\uff0c\u800c\u4e0d\u5fc5\u590d\u73b0\u5168\u5957 3D pipeline\u3002<\/p>\n<hr \/>\n<h2>\u4eca\u65e5\u63a8\u8350\u4f18\u5148\u7ea7<\/h2>\n<ol>\n<li><strong>Uncertainty-Guided Conservative Propagation \/ UGCP<\/strong>\uff1a\u66f4\u9002\u5408\u4f18\u5148\u8bfb\u3002\u5b83\u662f\u53ef\u63d2\u62d4\u3001\u6a21\u5757\u8fb9\u754c\u6e05\u695a\u7684 refinement \u65b9\u6cd5\uff0c\u5bf9 U-Net\u3001Transformer\u3001Mamba\u3001DAMamba \u540e\u7aef\u90fd\u8f83\u5bb9\u6613\u8fc1\u79fb\uff1b\u800c\u4e14\u6709\u4ee3\u7801\u94fe\u63a5\u548c\u8f83\u5b8c\u6574\u6d88\u878d\u3002<\/li>\n<li><strong>PanGuide3D<\/strong>\uff1a\u66f4\u9002\u5408\u4f5c\u4e3a 3D organ-tumor \u5206\u5272\u4e0e cross-cohort evaluation \u7684\u8bbe\u8ba1\u53c2\u8003\u3002\u65b9\u6cd5\u6982\u5ff5\u6e05\u695a\uff0c\u4f46\u4ee3\u7801\u672a\u516c\u5f00\u3001\u8d44\u6e90\u8981\u6c42\u9ad8\uff0c\u77ed\u671f\u590d\u73b0\u96be\u5ea6\u66f4\u5927\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 PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.20543<\/li>\n<li>\u8bba\u6587 2\uff1a\u5df2\u9644 PDF\uff1b\u672c\u5730\u6587\u4ef6 PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2604.20981<\/li>\n<\/ul>\n<h2>\u4eca\u65e5\u53ef\u6267\u884c\u5efa\u8bae<\/h2>\n<ol>\n<li>\u4f18\u5148\u7cbe\u8bfb\u5e76\u5c1d\u8bd5\u590d\u73b0 <strong>UGCP \u7684 logit-space uncertainty-guided propagation<\/strong>\uff1a\u5148\u63a5\u5230\u73b0\u6709 U-Net\/DAMamba \u8f93\u51fa logits \u540e\uff0c\u4ec5\u505a <code>T=1\/2<\/code> \u6b65\u66f4\u65b0\uff0c\u89c2\u5bdf Dice\u3001HD95\u3001\u8fb9\u754c F-score \u6216 connected components \u662f\u5426\u6539\u5584\u3002<\/li>\n<li>\u82e5\u505a 3D \u533b\u5b66\u5206\u5272\u6846\u67b6\uff0c\u53ef\u4ee5\u628a <strong>PanGuide3D \u7684 organ-conditioned tumor decoder<\/strong> \u4f5c\u4e3a\u53ef\u89e3\u91ca\u8bbe\u8ba1\uff1a<code>organ probability map + lesion decoder soft gating + Mamba\/Transformer bottleneck<\/code> \u662f\u4e00\u4e2a\u6e05\u6670\u7684\u6539\u9020\u7ec4\u5408\u3002<\/li>\n<li>\u5199\u8bba\u6587\u5b9e\u9a8c\u65f6\u5efa\u8bae\u501f\u9274\u4e24\u7bc7\u7684\u8bc4\u4ef7\u7ef4\u5ea6\uff1aUGCP \u7684 clDice\/HD95\/CRF \u5bf9\u6bd4\uff0cPanGuide3D \u7684 false-positive volume\u3001patient sensitivity\u3001small-lesion stratification\u3001calibration\uff1b\u8fd9\u4e9b\u6bd4\u5355\u7eaf Dice \u66f4\u80fd\u652f\u6491\u201c\u7ed3\u6784\u53ef\u9760\u6027\/\u4e34\u5e8a\u53ef\u7528\u6027\u201d\u7684 claim\u3002<\/li>\n<\/ol>\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\u53d1\u73b0\u6765\u81ea MICCAI\/CVPR\/\u9876\u520a\u5b98\u7f51\u4e14\u6bd4\u6628\u65e5\u66f4\u6210\u719f\u7684\u6b63\u5f0f\u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587 &#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-1071","post","type-post","status-publish","format-standard","hentry","category-85"],"views":16,"_links":{"self":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1071","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=1071"}],"version-history":[{"count":1,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1071\/revisions"}],"predecessor-version":[{"id":1072,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1071\/revisions\/1072"}],"wp:attachment":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/media?parent=1071"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/categories?post=1071"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/tags?post=1071"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}