\u4eca\u5929\u6ca1\u6709\u68c0\u7d22\u5230 2026-05-27 \u5f53\u65e5\u521a\u53d1\u5e03\u4e14\u8d28\u91cf\u8db3\u591f\u7a33\u5b9a\u7684\u533b\u5b66\u56fe\u50cf\u5206\u5272\u9876\u4f1a\/\u9876\u520a\u8bba\u6587\uff1b\u56e0\u6b64\u6309\u4efb\u52a1\u89c4\u5219\u4ece 2026 \u5e74\u8fd1\u671f\u5df2\u516c\u5f00 PDF \u4e0e\u9876\u4f1a\u63a5\u6536\u8bba\u6587\u4e2d\u56de\u6eaf\u7b5b\u9009\u3002\u6700\u7ec8\u9009\u62e9 2 \u7bc7\u66f4\u503c\u5f97\u7cbe\u8bfb\u7684 CVPR 2026\/arXiv \u8bba\u6587\uff1a\u4e00\u7bc7\u4ee3\u8868\u201c\u533b\u5b66\u5206\u5272 foundation\/VLM + \u4e0d\u786e\u5b9a\u6027\u5efa\u6a21 + \u8de8\u57df\u6cdb\u5316\u201d\uff0c\u53e6\u4e00\u7bc7\u4ee3\u8868\u201c\u534a\u76d1\u7763\u533b\u5b66\u5206\u5272 + \u751f\u6210\u5f0f\u5148\u9a8c\/Stable Diffusion VAE + \u5206\u5e03\u5bf9\u9f50\u201d\u3002\u6574\u4f53\u8d8b\u52bf\u662f\uff1a\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6b63\u5728\u4ece\u5355\u7eaf U-Net\/Transformer \u7ed3\u6784\u6539\u9020\uff0c\u8f6c\u5411\u5229\u7528\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u8868\u5f81\u3001\u6587\u672c\u8bed\u4e49\u3001\u751f\u6210\u6a21\u578b\u5148\u9a8c\u548c\u53ef\u9760\u6027\/\u4f4e\u6807\u6ce8\u9c81\u68d2\u6027\u6765\u63d0\u5347\u6cdb\u5316\u3002<\/p>\n
\u4eca\u5929\u68c0\u7d22\u8303\u56f4\u8986\u76d6 arXiv\u3001CVPR 2026 virtual papers\u3001DuckDuckGo \u7d22\u5f15\u5230\u7684 arXiv\/\u9879\u76ee\u9875\uff0c\u5e76\u91cd\u70b9\u67e5\u8be2 medical image segmentation\u3001polyp segmentation\u3001semi-supervised medical image segmentation\u3001vision-language medical segmentation\u3001Mamba medical segmentation\u3001universal medical segmentation \u7b49\u5173\u952e\u8bcd\uff1barXiv API \u5728\u672c\u6b21\u4efb\u52a1\u4e2d\u591a\u6b21\u8fd4\u56de 429\uff0c\u56e0\u6b64\u540c\u65f6\u4f7f\u7528 arXiv HTML \u9875\u9762\u3001CVPR 2026 \u9875\u9762\u4e0e\u641c\u7d22\u5f15\u64ce\u7ed3\u679c\u4ea4\u53c9\u6838\u9a8c\u3002\u672a\u627e\u5230\u5f53\u5929\u8db3\u591f\u65b0\u7684\u9ad8\u8d28\u91cf\u8bba\u6587\u540e\uff0c\u6309\u8981\u6c42\u56de\u6eaf\u5230 2026 \u5e74 2\u20134 \u6708\u7684 CVPR 2026\/\u9884\u5370\u672c\u8bba\u6587\uff1b\u4e24\u7bc7\u5165\u9009\u8bba\u6587\u5747\u4e3a 2025 \u5e74\u53ca\u4ee5\u540e\u8bba\u6587\u3002\u5df2\u68c0\u67e5\u5386\u53f2\u63a8\u8350\u8bb0\u5f55\u5e76\u6392\u9664\u4e86\u91cd\u590d\u8bba\u6587\uff1b\u5386\u53f2\u5df2\u63a8\u8350\u5e76\u8df3\u8fc7\u7684\u91cd\u590d\u5019\u9009\u5305\u62ec Patch-MoE Mamba\u3001DepthPolyp\u3002\u53e6\u6709\u5019\u9009\u5982 MedCRP-CL\u3001TCSeg\/overconfidence semi-supervised 3D segmentation\u3001VesMamba\u3001OSA\u3001Med-DisSeg \u7b49\uff0c\u672c\u6b21\u56e0\u4e0e\u7528\u6237\u8bfe\u9898\u5339\u914d\u5ea6\u3001PDF\/\u6b63\u6587\u53ef\u83b7\u53d6\u6027\u3001\u65b9\u6cd5\u53ef\u8fc1\u79fb\u6027\u6216\u7814\u7a76\u4ef7\u503c\u6392\u5e8f\u672a\u8fdb\u5165\u6700\u7ec8 2 \u7bc7\u3002<\/p>\n
MedCLIPSeg \u6700\u91cd\u8981\u7684\u4ef7\u503c\u5728\u4e8e\u628a CLIP \u7c7b\u533b\u5b66\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u4ece\u201c\u5168\u5c40\u56fe\u6587\u5bf9\u9f50\u201d\u63a8\u8fdb\u5230\u201cpatch-level\u3001\u53cc\u5411\u3001\u6982\u7387\u5f0f\u56fe\u6587\u878d\u5408\u201d\u7684\u5206\u5272\u6846\u67b6\uff0c\u540c\u65f6\u628a\u5206\u5272\u7cbe\u5ea6\u3001\u4f4e\u6807\u6ce8\u6548\u7387\u3001OOD \u6cdb\u5316\u548c\u50cf\u7d20\u7ea7\u4e0d\u786e\u5b9a\u6027\u653e\u5728\u540c\u4e00\u4e2a\u5b9e\u9a8c\u4f53\u7cfb\u91cc\u9a8c\u8bc1\u3002<\/p>\n
\u8bba\u6587\u7814\u7a76\u7684\u95ee\u9898\u662f\uff1a\u5728\u533b\u5b66\u56fe\u50cf\u5206\u5272\u4e2d\uff0c\u5982\u4f55\u5229\u7528\u6587\u672c\u8bed\u4e49\u548c\u89c6\u89c9\u8bed\u8a00\u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u5728\u6807\u6ce8\u4e0d\u8db3\u3001\u8fb9\u754c\u6a21\u7cca\u3001\u8de8\u8bbe\u5907\/\u8de8\u4e2d\u5fc3\u57df\u504f\u79fb\u660e\u663e\u7684\u60c5\u51b5\u4e0b\u83b7\u5f97\u66f4\u53ef\u9760\u7684\u5206\u5272\u7ed3\u679c\u3002\u4f5c\u8005\u660e\u786e\u6307\u51fa\u4f20\u7edf U-Net\u3001nnU-Net\u3001Transformer\/ViT \u7c7b\u5206\u5272\u7f51\u7edc\u901a\u5e38\u4f9d\u8d56\u5927\u91cf\u50cf\u7d20\u7ea7 mask\uff0c\u4e14\u8f93\u51fa\u504f\u786e\u5b9a\u6027\uff1b\u5728 OOD \u6570\u636e\u548c\u6a21\u7cca\u8fb9\u754c\u5904\u5bb9\u6613\u8fc7\u5ea6\u81ea\u4fe1\u3002CLIP\/BiomedCLIP\/UniMedCLIP \u7b49 VLM \u867d\u7136\u5177\u5907\u56fe\u6587\u5bf9\u9f50\u80fd\u529b\uff0c\u4f46\u539f\u59cb CLIP \u4e3b\u8981\u5bf9\u9f50\u5168\u5c40 [CLS]\/[EOS] \u8868\u5f81\uff0c\u76f4\u63a5\u7528\u4e8e dense localization \u65f6\u7a7a\u95f4\u5b9a\u4f4d\u80fd\u529b\u4e0d\u8db3\u3002\u56e0\u6b64\u6838\u5fc3\u95ee\u9898\u4e0d\u662f\u201c\u518d\u9020\u4e00\u4e2a U-Net\u201d\uff0c\u800c\u662f\u5982\u4f55\u8ba9 CLIP patch token \u4e0e\u533b\u5b66\u6587\u672c token \u8fdb\u884c\u53ef\u9760\u7684\u5c40\u90e8\u8bed\u4e49\u4ea4\u4e92\uff0c\u5e76\u663e\u5f0f\u8868\u8fbe\u9884\u6d4b\u4e0d\u786e\u5b9a\u6027\u3002<\/p>\n
Paper map\uff1a\u8bba\u6587\u7814\u7a76\u6587\u672c\u9a71\u52a8\u533b\u5b66\u56fe\u50cf\u5206\u5272\uff0c\u8bbe\u7f6e\u4e3a\u591a\u5668\u5b98\u3001\u591a\u6a21\u6001\u3001\u4f4e\u6807\u6ce8\u4e0e\u8de8\u57df\u6cdb\u5316\u3002\u4e3b\u62db\u662f\u51bb\u7ed3\/\u4fdd\u7559 CLIP \u7c7b\u7f16\u7801\u5668\uff0c\u5728\u591a\u4e2a\u6df1\u5c42\u52a0\u5165 Probabilistic Vision-Language Adapter\uff08PVL Adapter\uff09\uff0c\u7528\u6982\u7387 Key\/Value \u7684 cross-modal attention\u3001\u6b8b\u5dee\u95e8\u63a7\u3001\u53cc\u5411\u89c6\u89c9-\u6587\u672c\u4ea4\u4e92\u548c soft patch-level contrastive loss \u6765\u505a dense segmentation\u3002\u5b83\u58f0\u79f0\u80fd\u63d0\u5347 accuracy\u3001data efficiency\u3001OOD generalization \u4e0e calibration\uff0c\u8bc1\u636e\u4e3b\u8981\u6765\u81ea 16 \u4e2a\u6570\u636e\u96c6\u3001\u6570\u636e\u6bd4\u4f8b\u5b9e\u9a8c\u3001\u8de8\u57df\u6d4b\u8bd5\u3001\u6d88\u878d\u3001Brier score \u4e0e uncertainty-error correlation\u3002\u5173\u952e\u6280\u672f\u5bf9\u8c61\u662f PVL Adapter\u3001AttnPVL\u3001probabilistic K\/V\u3001confidence-weighted attention\u3001MC value sampling\u3001pixel-text similarity head\u3001soft contrastive loss\u3002\u771f\u6b63\u7684\u77e5\u8bc6\u8d1f\u8f7d\u5728\u201c\u6982\u7387\u6ce8\u610f\u529b\u662f\u5426\u771f\u7684\u6539\u5584 OOD\/\u6821\u51c6\u201d\u4ee5\u53ca\u201c\u6587\u672c prompt \u4e0e CLIP backbone \u662f\u5426\u516c\u5e73\u4e14\u53ef\u590d\u73b0\u201d\u3002\u4e3b\u8981\u5931\u8d25\u98ce\u9669\u662f\u63d0\u793a\u8bcd\u751f\u6210\u3001\u6570\u636e\u5207\u5206\u548c CLIP-based baseline \u9002\u914d\u7ec6\u8282\u53ef\u80fd\u5f71\u54cd\u7ed3\u679c\uff0c\u4e14 30 \u6b21 MC sampling \u7684\u5b9e\u9645\u63a8\u7406\u6210\u672c\u6ca1\u6709\u5728\u4e3b\u7ed3\u679c\u4e2d\u5b8c\u5168\u4f53\u73b0\u3002<\/p>\n
Route record\uff1aPrimary adapter = method-algorithm\uff1bSecondary adapter = benchmark-evaluation\uff1bEvidence packs = general\u3001experimental-eval\u3001robustness-and-ood\u3001reproducibility-and-compute\uff1bRoute confidence = \u9ad8\u3002\u9009\u62e9\u8be5\u8def\u7ebf\u662f\u56e0\u4e3a\u8bba\u6587\u8d21\u732e\u4e3b\u8981\u662f\u65b9\u6cd5\u6a21\u5757\uff0c\u4f46\u5176\u53ef\u4fe1\u5ea6\u9ad8\u5ea6\u4f9d\u8d56\u8de8\u57df\u3001\u591a\u6570\u636e\u96c6\u548c\u6d88\u878d\u5b9e\u9a8c\u3002<\/p>\n
\u4f5c\u8005\u5bf9\u5df2\u6709\u65b9\u6cd5\u7684\u6279\u8bc4\u5206\u4e09\u5c42\u3002\u7b2c\u4e00\uff0cU-Net\u3001UNet++\u3001Attention U-Net\u3001nnU-Net\u3001DeepLab\u3001TransUNet\u3001Swin-UNet\u3001UNETR \u7b49 vision-only \u5206\u5272\u65b9\u6cd5\u867d\u7136\u5728 supervised benchmark \u4e0a\u5f3a\uff0c\u4f46\u4f9d\u8d56 pixel-wise mask\uff0c\u4e14\u5bf9 scanner\u3001protocol\u3001patient population shift \u654f\u611f\u3002\u7b2c\u4e8c\uff0cCLIPSeg\u3001CRIS\u3001DenseCLIP\u3001ZegCLIP\u3001CAT-Seg\u3001MaPLe \u7b49\u901a\u7528\u5f00\u653e\u8bcd\u8868\/\u6587\u672c\u9a71\u52a8\u5206\u5272\u65b9\u6cd5\u5e76\u975e\u4e3a\u533b\u5b66\u56fe\u50cf\u7684\u7ec6\u7c92\u5ea6\u8fb9\u754c\u3001\u4f4e\u5bf9\u6bd4\u548c\u9ad8\u7c7b\u95f4\u76f8\u4f3c\u6027\u8bbe\u8ba1\uff0c\u76f4\u63a5\u8fc1\u79fb\u5230\u533b\u5b66\u573a\u666f\u65f6 spatial grounding \u4e0d\u7a33\u3002\u7b2c\u4e09\uff0c\u5df2\u6709\u533b\u5b66 VLM adaptation \u591a\u96c6\u4e2d\u5728 prompt tuning\u3001decoder tuning \u6216\u5355\u5411 text-to-vision modulation\uff0c\u7f3a\u5c11\u6df1\u5c42\u3001\u53cc\u5411\u3001\u6982\u7387\u5f0f\u7684 vision-language fusion\uff0c\u4e5f\u5f88\u5c11\u628a\u4e0d\u786e\u5b9a\u6027\u56fe\u4f5c\u4e3a\u53ef\u9760\u6027\u8f93\u51fa\u3002<\/p>\n
MedCLIPSeg \u4ee5 UniMedCLIP ViT-B\/16 \u4f5c\u4e3a\u89c6\u89c9 backbone\uff0c\u4ee5 PubMedBERT \u4f5c\u4e3a\u6587\u672c\u7f16\u7801\u5668\u3002\u8f93\u5165\u56fe\u50cf\u88ab\u5207\u6210 patch\uff0c\u89c6\u89c9\u7f16\u7801\u5668\u8f93\u51fa visual tokens\uff1b\u6587\u672c prompt \u7ecf\u6587\u672c\u7f16\u7801\u5668\u8f93\u51fa text tokens\u3002\u4e0e\u666e\u901a CLIP \u53ea\u7528\u5168\u5c40\u76f8\u4f3c\u5ea6\u4e0d\u540c\uff0cMedCLIPSeg \u5728 CLIP \u591a\u4e2a\u6df1\u5c42\u63d2\u5165 PVL Adapter\uff0c\u8ba9\u89c6\u89c9 token \u548c\u6587\u672c token \u5728\u4f4e\u7ef4\u5171\u4eab\u7a7a\u95f4\u4e2d\u4ea4\u4e92\u3002<\/p>\n
\u6574\u4f53\u6d41\u7a0b\u53ef\u4ee5\u62c6\u4e3a\u4e94\u6b65\uff1a PVL Adapter<\/strong>\uff1a\u8f93\u5165\u4e3a\u67d0\u4e00\u5c42\u89c6\u89c9 tokens (V^{(n)}) \u4e0e\u6587\u672c tokens (T^{(n)})\uff0c\u8f93\u51fa\u4e3a\u589e\u5f3a\u540e\u7684 (\\hat V^{(n)})\u3001(\\hat T^{(n)})\u3002\u5b83\u89e3\u51b3\u7684\u662f CLIP \u5168\u5c40\u56fe\u6587\u5bf9\u9f50\u4e0d\u8db3\u4ee5\u652f\u6491\u533b\u5b66 dense prediction \u7684\u95ee\u9898\u3002\u8be5\u6a21\u5757\u4e0d\u662f\u7b80\u5355 concat\uff0c\u4e5f\u4e0d\u662f\u5355\u5411 cross-attention\uff0c\u800c\u662f\u901a\u8fc7\u53cc\u5411 token \u4ea4\u4e92\u8ba9\u6587\u672c\u7ec6\u5316\u89c6\u89c9\u533a\u57df\u3001\u89c6\u89c9\u53cd\u8fc7\u6765\u7ea6\u675f\u6587\u672c\u8868\u793a\u3002<\/p>\n AttnPVL \/ \u6982\u7387 Key-Value \u6ce8\u610f\u529b<\/strong>\uff1a\u6807\u51c6 attention \u7684 (Q,K,V) \u4e2d\uff0cQ \u4ecd\u4e3a\u786e\u5b9a query\uff0c\u4f46 K \u548c V \u88ab\u5efa\u6a21\u4e3a\u9ad8\u65af\u5206\u5e03\uff1a([K_\\mu,K_{\\log\\sigma^2}]=ZW_K)\uff0c([V_\\mu,V_{\\log\\sigma^2}]=ZW_V)\uff0c\u65b9\u5dee\u7528 softplus \u4fdd\u8bc1\u6570\u503c\u7a33\u5b9a\u3002\u6ce8\u610f\u529b\u5206\u6570\u7531\u5747\u503c\u76f8\u4f3c\u5ea6 (S_\\mu=QK_\\mu^\\top\/\\sqrt{D_a}) \u548c\u65b9\u5dee\u60e9\u7f5a (S^2_\\sigma=Q^{\\circ 2}(K^2_\\sigma)^\\top\/D_a) \u5171\u540c\u51b3\u5b9a\uff0c\u6700\u7ec8\u8fd1\u4f3c\u4e3a (\\mathrm{softmax}(S_\\mu-\\beta S_\\sigma))\u3002\u76f4\u89c2\u4e0a\uff0c\u4e0d\u786e\u5b9a\u7684 key token \u5373\u4fbf\u5747\u503c\u76f8\u4f3c\uff0c\u4e5f\u4f1a\u88ab\u4e0b\u8c03\u6743\u91cd\u3002\u8fd9\u662f\u8bba\u6587\u6700\u6838\u5fc3\u7684\u673a\u5236\u521b\u65b0\uff0c\u9002\u5408\u8fc1\u79fb\u5230\u5176\u4ed6\u533b\u5b66\u5206\u5272\u6846\u67b6\u4e2d\u7684 cross-attention \/ skip fusion \/ decoder fusion \u6a21\u5757\u3002<\/p>\n Value sampling \u4e0e uncertainty map<\/strong>\uff1a\u8bad\u7ec3\u65f6\u5bf9 Value \u5206\u5e03\u91c7\u6837\u4e00\u6b21\uff0c\u6d4b\u8bd5\u65f6\u591a\u6b21 stochastic forward\uff08\u4f5c\u8005\u7ecf\u9a8c\u4e0a 30 \u6b21\uff09\u5f97\u5230 mask \u6837\u672c\u5206\u5e03\uff0c\u7528 predictive entropy \u5f62\u6210\u50cf\u7d20\u7ea7 uncertainty map\u3002\u8fd9\u4e2a\u8bbe\u8ba1\u9002\u5408\u8fb9\u754c\u6a21\u7cca\u7684 lesion\u3001polyp\u3001tumor \u5206\u5272\uff1b\u4f46\u5982\u679c\u90e8\u7f72\u5728\u5b9e\u65f6\u5185\u955c\u6216 3D \u4f53\u6570\u636e\u4e0a\uff0c30 \u6b21\u91c7\u6837\u6210\u672c\u8f83\u9ad8\u3002<\/p>\n Residual gating<\/strong>\uff1aPVL \u8f93\u51fa\u4e0d\u662f\u76f4\u63a5\u66ff\u6362\u539f token\uff0c\u800c\u662f (Y=g\\odot O_{proj}+(1-g)\\odot X)\u3002\u5b83\u7684\u4f5c\u7528\u662f\u907f\u514d\u8bad\u7ec3\u65e9\u671f cross-modal attention \u566a\u58f0\u7834\u574f CLIP \u8868\u5f81\u3002\u8fd9\u4e2a\u95e8\u63a7\u601d\u60f3\u5bf9 U-Net\/Transformer\/Mamba hybrid segmentation \u7684\u8de8\u5c42\u878d\u5408\u4e5f\u6709\u8fc1\u79fb\u4ef7\u503c\u3002<\/p>\n Soft patch-level contrastive loss<\/strong>\uff1a\u4f5c\u8005\u628a patch embedding \u5e73\u5747\u4e3a\u533a\u57df\u8868\u5f81\uff0c\u4e0e\u6587\u672c embedding \u5bf9\u9f50\uff1b\u76ee\u6807\u4e0d\u662f one-hot\uff0c\u800c\u662f\u6839\u636e\u6587\u672c\u95f4\u76f8\u4f3c\u5ea6\u6784\u9020 soft target\u3002\u8fd9\u4e2a\u6a21\u5757\u7684\u521b\u65b0\u6027\u4e2d\u7b49\uff0c\u4f46\u5b9e\u7528\u6027\u9ad8\uff0c\u5c24\u5176\u9002\u5408\u533b\u5b66\u6587\u672c prompt \u4e4b\u95f4\u8bed\u4e49\u76f8\u8fd1\u7684\u60c5\u51b5\u3002<\/p>\n \u5bf9 polyp segmentation\uff1a\u8bba\u6587\u5305\u542b Kvasir-SEG \u8bad\u7ec3\/\u6d4b\u8bd5\u548c CVC-ColonDB\u3001CVC-ClinicDB\u3001CVC-300\u3001BKAI \u7b49 OOD polyp\/endoscopy \u6570\u636e\uff0c\u76f8\u5173\u6027\u5f88\u5f3a\u3002\u5bf9 3D medical segmentation\uff1a\u672c\u6587\u65b9\u6cd5\u4e3b\u8981\u662f 2D\/RGB \u6216 2D \u5316\u8f93\u5165\uff0c\u4e0d\u80fd\u76f4\u63a5\u89e3\u51b3 3D \u4f53\u6570\u636e\u957f\u7a0b\u5efa\u6a21\uff1b\u4f46\u6982\u7387 cross-modal adapter \u53ef\u4f5c\u4e3a 3D encoder \u7684\u8bed\u4e49\u6761\u4ef6\u6a21\u5757\u3002<\/p>\n \u5b9e\u9a8c\u8986\u76d6 16 \u4e2a\u6570\u636e\u96c6\u3001\u4e94\u79cd\u6a21\u6001\u548c\u516d\u7c7b\u5668\u5b98\/\u76ee\u6807\u3002\u8bad\u7ec3\u6570\u636e\u6548\u7387\u5b9e\u9a8c\u4f7f\u7528 10%\u300125%\u300150%\u3001100% \u6807\u6ce8\u6bd4\u4f8b\uff1b\u8de8\u57df\u6cdb\u5316\u5b9e\u9a8c\u5728 source dataset \u4e0a\u8bad\u7ec3\uff0c\u5728 unseen target datasets \u4e0a\u65e0\u5fae\u8c03\u6d4b\u8bd5\u3002\u4e3b\u8981\u6307\u6807\u4e3a DSC \u548c NSD\uff1b\u53ef\u9760\u6027\u7528 uncertainty-error correlation \u4e0e Brier score\u3002<\/p>\n \u5173\u952e\u7ed3\u679c\u5305\u62ec\uff1a \u53ef\u4fe1\u5ea6\u603b\u4f53\u8f83\u9ad8\uff0c\u4f46\u4e0d\u662f\u65e0\u6761\u4ef6\u63a5\u53d7\u3002\u5f3a\u9879\u662f\uff1a\u6570\u636e\u96c6\u6570\u91cf\u591a\uff0c\u5305\u542b polyp\u3001ultrasound\u3001MRI\u3001skin\u3001X-ray\u3001EUS\uff1bbaseline \u8986\u76d6 U-Net\/nnU-Net\/Transformer\/CLIP-based segmentation\uff1b\u6d88\u878d\u76f4\u63a5\u9488\u5bf9 PVL\u3001gating\u3001AttnPVL\u3001deterministic variant\u3001bidirectional interaction\u3001contrastive loss\u3001prompt style \u548c backbone\uff0c\u80fd\u652f\u6491\u201c\u6982\u7387\u5f0f\u56fe\u6587\u878d\u5408\u201d\u662f\u4e3b\u8981\u6536\u76ca\u6765\u6e90\u3002<\/p>\n \u4e3b\u8981 caveat \u6709\u56db\u4e2a\u3002\u7b2c\u4e00\uff0c\u6587\u672c prompt \u7684\u751f\u6210\u4f9d\u8d56 GPT-5 \u6a21\u677f\u548c mask\/image processing\uff1b\u867d\u7136\u9644\u5f55\u7ed9\u51fa\u7b97\u6cd5\uff0c\u4f46 prompt \u8d28\u91cf\u5bf9\u6027\u80fd\u5f71\u54cd\u6781\u5927\uff0c\u5b9e\u9645\u590d\u73b0\u65f6\u53ef\u80fd\u6210\u4e3a\u9690\u85cf\u53d8\u91cf\u3002\u7b2c\u4e8c\uff0cCLIP-based baseline \u662f\u5426\u90fd\u88ab\u540c\u7b49\u5145\u5206\u8c03\u53c2\u4ecd\u9700\u4ee3\u7801\u786e\u8ba4\uff1b\u8bba\u6587\u79f0\u540c\u7528 UniMedCLIP backbone\uff0c\u4f46\u4e0d\u540c\u65b9\u6cd5\u7684\u6700\u4f73\u8bad\u7ec3\u7b56\u7565\u53ef\u80fd\u4e0d\u540c\u3002\u7b2c\u4e09\uff0c\u63a8\u7406\u65f6 uncertainty \u9700\u8981\u591a\u6b21\u91c7\u6837\uff1b\u4e3b\u8ba1\u7b97\u8868\u6309 single sampled forward \u505a\u516c\u5e73 per-sample \u5bf9\u6bd4\uff0c\u4e0d\u80fd\u5b8c\u5168\u4ee3\u8868 uncertainty mode \u7684\u771f\u5b9e\u90e8\u7f72\u6210\u672c\u3002\u7b2c\u56db\uff0c\u8bba\u6587\u4e3b\u8981\u662f 2D benchmark\uff1b\u5bf9 3D CT\/MRI segmentation \u7684\u76f4\u63a5\u4ef7\u503c\u6709\u9650\u3002<\/p>\n \u4e0e U-Net\/nnU-Net \u7684\u5173\u7cfb\uff1aMedCLIPSeg \u4e0d\u4f9d\u8d56 U-Net \u5f0f encoder-decoder \u4e3b\u5e72\uff0c\u800c\u662f\u7528 CLIP patch token + text token \u76f8\u4f3c\u5ea6\u505a segmentation\uff1b\u5b83\u6311\u6218\u7684\u662f nnU-Net \u5728\u4f4e\u6807\u6ce8\u548c OOD \u573a\u666f\u4e0b\u7f3a\u5c11\u6587\u672c\u8bed\u4e49\u548c\u4e0d\u786e\u5b9a\u6027\u8868\u8fbe\u7684\u95ee\u9898\u3002<\/p>\n \u4e0e MedNeXt\/CNN-based segmentation \u7684\u5173\u7cfb\uff1aMedNeXt \u5f3a\u8c03\u5377\u79ef\u5f52\u7eb3\u504f\u7f6e\u548c\u5927\u6838\/3D \u6269\u5c55\uff0cMedCLIPSeg \u5f3a\u8c03\u9884\u8bad\u7ec3 VLM \u4e0e\u8de8\u6a21\u6001\u8bed\u4e49\u3002\u4e8c\u8005\u53ef\u4e92\u8865\uff0c\u4f8b\u5982\u7528 MedNeXt\/3D CNN \u4ea7\u751f\u5c40\u90e8\u7ed3\u6784 token\uff0c\u518d\u63a5\u6982\u7387\u6587\u672c adapter\u3002<\/p>\n \u4e0e UNETR\/Swin-UNet\/TransUNet\/TransFuse \u7684\u5173\u7cfb\uff1a\u8fd9\u4e9b\u65b9\u6cd5\u7528 Transformer \u6355\u6349\u957f\u7a0b\u4f9d\u8d56\uff0c\u4f46\u4ecd\u662f vision-only \u6216\u5f31\u6587\u672c\u6761\u4ef6\uff1bMedCLIPSeg \u7684\u957f\u7a0b\u8bed\u4e49\u6765\u81ea CLIP \u548c\u6587\u672c\u4ea4\u4e92\uff0c\u4e0d\u662f\u5355\u7eaf self-attention\u3002<\/p>\n \u4e0e Mamba\/VMamba\/SegMamba\/DAMamba \u7684\u5173\u7cfb\uff1aMamba \u7c7b\u65b9\u6cd5\u5173\u6ce8\u9ad8\u6548\u957f\u5e8f\u5217\u5efa\u6a21\uff0c\u5c24\u5176\u9002\u5408 3D\/\u9ad8\u5206\u8fa8\u7387\uff1bMedCLIPSeg \u4e0d\u89e3\u51b3\u7ebf\u6027\u590d\u6742\u5ea6\u5e8f\u5217\u626b\u63cf\u95ee\u9898\uff0c\u4f46\u5176 confidence-weighted cross-modal adapter \u53ef\u4f5c\u4e3a Mamba encoder \u540e\u7684\u8bed\u4e49\u878d\u5408\u5934\uff0c\u7ed9 DAMamba \u7c7b\u65b9\u6cd5\u52a0\u5165\u6587\u672c\u6761\u4ef6\u548c uncertainty\u3002<\/p>\n \u4e0e\u533b\u5b66 foundation model\/SAM\/MedSAM \u7684\u5173\u7cfb\uff1aMedCLIPSeg \u4e0e SAM\/MedSAM \u90fd\u5c5e\u4e8e promptable\/foundation segmentation \u65b9\u5411\uff0c\u4f46\u5b83\u4f7f\u7528\u81ea\u7136\u8bed\u8a00 prompt \u800c\u4e0d\u662f\u70b9\u6846 prompt\uff0c\u5e76\u663e\u5f0f\u8f93\u51fa uncertainty map\u3002\u76f8\u6bd4 MedSAM \u66f4\u9002\u5408\u9700\u8981\u8bed\u4e49\u63cf\u8ff0\u548c\u8de8\u57df\u53ef\u9760\u6027\u8bc4\u4f30\u7684\u573a\u666f\u3002<\/p>\n \u5bf9 polyp segmentation\uff1a\u4ef7\u503c\u5f88\u9ad8\u3002\u8bba\u6587\u5305\u542b Kvasir-SEG \u548c\u591a\u4e2a CVC\/BKAI OOD \u6d4b\u8bd5\uff0c\u80fd\u76f4\u63a5\u4f5c\u4e3a polyp segmentation \u7684 related work \u548c\u5f3a baseline \u53c2\u8003\u3002\u5c24\u5176\u503c\u5f97\u501f\u9274\u7684\u662f prompt quality \u6d88\u878d\u3001\u8de8\u57df\u6d4b\u8bd5\u8bbe\u8ba1\u548c uncertainty map\uff0c\u53ef\u7528\u4e8e\u8bf4\u660e\u771f\u5b9e\u5185\u955c\u573a\u666f\u4e2d domain shift \u4e0e\u8fb9\u754c\u4e0d\u786e\u5b9a\u6027\u3002<\/p>\n \u5bf9 DAMamba \u6539\u9020\uff1a\u5efa\u8bae\u91cd\u70b9\u501f\u9274\u201c\u6982\u7387 cross-attention + \u6b8b\u5dee\u95e8\u63a7 + uncertainty head\u201d\uff0c\u800c\u4e0d\u662f\u5b8c\u6574\u7167\u642c CLIP \u6846\u67b6\u3002\u53ef\u4ee5\u8003\u8651\u5728 DAMamba decoder \u6216 skip fusion \u4e2d\u52a0\u5165\u8f7b\u91cf probabilistic adapter\uff0c\u7528\u65b9\u5dee\u60e9\u7f5a\u8c03\u8282\u8de8\u5c3a\u5ea6\/\u8de8\u6a21\u6001\u7279\u5f81\u878d\u5408\uff1b\u82e5\u6ca1\u6709\u6587\u672c\u8f93\u5165\uff0c\u4e5f\u53ef\u628a class prototype \u6216 anatomical prior token \u5f53\u4f5c query\/context\u3002<\/p>\n \u5bf9\u533b\u5b66\u5206\u5272\u6846\u67b6\u9009\u62e9\uff1a\u9002\u5408\u4f5c\u4e3a VLM\/foundation model segmentation \u65b9\u5411\u7684\u91cd\u8981 baseline\uff0c\u4e0d\u9002\u5408\u4f5c\u4e3a\u7eaf 3D CT segmentation backbone \u7684\u76f4\u63a5\u66ff\u4ee3\u3002<\/p>\n \u5bf9 introduction\/related work\uff1a\u975e\u5e38\u9002\u5408\u5f15\u7528\u5728\u201c\u533b\u5b66\u5206\u5272\u4ece architecture engineering \u8f6c\u5411 foundation\/VLM + uncertainty + OOD generalization\u201d\u7684\u8bba\u8ff0\u4e2d\u3002<\/p>\n \u5f3a\u70c8\u5efa\u8bae\u7cbe\u8bfb\u5168\u6587<\/strong>\u3002\u4f18\u5148\u8bfb Section 3.2 \u7684 PVL Adapter \u548c AttnPVL \u516c\u5f0f\u3001Section 4 \u7684 Table 1\u20135\u3001Appendix \u7684 dataset\/prompt generation \u548c compute cost\u3002\u82e5\u65f6\u95f4\u6709\u9650\uff0c\u53ef\u8df3\u8fc7\u90e8\u5206 related work\uff0c\u4f46\u4e0d\u8981\u8df3\u8fc7 prompt generation \u9644\u5f55\uff0c\u56e0\u4e3a prompt \u662f\u5f71\u54cd\u8be5\u65b9\u6cd5\u53ef\u4fe1\u5ea6\u548c\u590d\u73b0\u6027\u7684\u5173\u952e\u53d8\u91cf\u3002<\/p>\n SemiGDA \u7684\u6838\u5fc3\u4ef7\u503c\u662f\u628a\u534a\u76d1\u7763\u533b\u5b66\u5206\u5272\u4ece\u201c\u4f2a\u6807\u7b7e\/teacher-student \u4e00\u81f4\u6027\u201d\u6269\u5c55\u5230\u201c\u56fe\u50cf\u6f5c\u53d8\u91cf\u4e0e mask \u6f5c\u53d8\u91cf\u7684\u751f\u6210\u5f0f\u5206\u5e03\u5bf9\u9f50\u201d\uff0c\u5e76\u7528\u51bb\u7ed3 Stable Diffusion VAE decoder \u4e0e\u8f7b\u91cf skip adapter \u5728\u4f4e\u6807\u6ce8\u573a\u666f\u4e0b\u83b7\u5f97\u7a33\u5b9a\u63d0\u5347\u3002<\/p>\n \u8bba\u6587\u7814\u7a76\u534a\u76d1\u7763\u533b\u5b66\u56fe\u50cf\u5206\u5272\uff1a\u5728\u53ea\u6709\u5c11\u91cf\u6807\u6ce8 mask\u3001\u5927\u91cf\u672a\u6807\u6ce8\u56fe\u50cf\u65f6\uff0c\u5982\u4f55\u5229\u7528\u672a\u6807\u6ce8\u6570\u636e\u63d0\u5347\u5206\u5272\u3002\u533b\u5b66\u56fe\u50cf\u6807\u6ce8\u6602\u8d35\uff0c\u5c24\u5176 polyp\u3001pathology\u3001ultrasound \u7b49\u4efb\u52a1\u9700\u8981\u4e13\u5bb6\u52fe\u753b\u8fb9\u754c\uff0c\u56e0\u6b64 semi-supervised medical image segmentation\uff08SMIS\uff09\u5177\u6709\u5b9e\u9645\u610f\u4e49\u3002\u73b0\u6709 SMIS \u591a\u7528 pseudo-labeling\u3001Mean Teacher\u3001cross-consistency\u3001dual-stream mutual learning \u7b49\u5224\u522b\u5f0f\u8303\u5f0f\uff1b\u4f5c\u8005\u8ba4\u4e3a\u8fd9\u4e9b\u65b9\u6cd5\u8fc7\u5ea6\u4f9d\u8d56 per-pixel classification\uff0c\u5bb9\u6613\u53d7 noisy pseudo-label\u3001teacher \u7d2f\u79ef\u8bef\u5dee\u548c\u6709\u9650 label \u4e0b\u7684\u8fc7\u62df\u5408\u5f71\u54cd\uff0c\u4e14\u4e0d\u5584\u4e8e\u5efa\u6a21\u5168\u5c40\u7ed3\u6784\u4e0e mask \u5206\u5e03\u3002<\/p>\n Paper map\uff1a\u8bba\u6587\u7814\u7a76\u4f4e\u6807\u6ce8\u534a\u76d1\u7763\u533b\u5b66\u5206\u5272\uff0c\u8bbe\u7f6e\u4e3a 10%\/30% label \u751a\u81f3\u66f4\u4f4e label ratio \u7684 2D \u533b\u5b66\u56fe\u50cf\u3002\u4e3b\u62db\u662f\u628a\u8f93\u5165\u56fe\u50cf\u548c ground-truth mask \u90fd\u9001\u5165 VAE\/encoder latent space\uff0c\u5c06\u56fe\u50cf latent distribution \u6620\u5c04\u5e76\u7ea6\u675f\u5230 mask prior distribution\uff0c\u518d\u7528\u51bb\u7ed3 VAE decoder \u751f\u6210 mask\uff0c\u540c\u65f6\u7528 Consistency-Driven Skip Adapter \u6062\u590d\u591a\u5c3a\u5ea6\u7ec6\u8282\u3002\u5b83\u58f0\u79f0\u4f18\u4e8e SOTA SMIS\uff0c\u8bc1\u636e\u6765\u81ea CVC-ClinicDB\u3001Kvasir\u3001CVC-300\u3001ISIC-2018\u3001BCSS\u3001BUSI \u4e0a\u4e0e 11 \u4e2a\u534a\u76d1\u7763\u65b9\u6cd5\u6bd4\u8f83\u53ca\u6a21\u5757\/\u635f\u5931\u6d88\u878d\u3002\u5173\u952e\u6280\u672f\u5bf9\u8c61\u662f DAM\u3001latent mapping model\u3001mask\/image prior distributions\u3001CDSA\u3001ACR \u548c supervised\/unsupervised distribution\/segmentation losses\u3002\u771f\u6b63\u7684\u77e5\u8bc6\u8d1f\u8f7d\u5728\u201cStable Diffusion VAE prior \u662f\u5426\u771f\u7684\u9002\u914d\u533b\u5b66 mask \u751f\u6210\u201d\u4ee5\u53ca\u201c\u5206\u5e03\u5bf9\u9f50\u7ea6\u675f\u662f\u5426\u6bd4\u666e\u901a\u4e00\u81f4\u6027\u66f4\u6709\u6548\u201d\u3002\u4e3b\u8981\u5931\u8d25\u98ce\u9669\u662f\u65b9\u6cd5\u4f9d\u8d56 SD VAE \u7684 latent prior \u4e0e 224\u00d7224 2D \u8f93\u5165\uff0c\u4e14\u7f3a\u5c11\u5bf9 3D\/\u8de8\u57df\u6cdb\u5316\/\u7edf\u8ba1\u663e\u8457\u6027\u548c\u8ba1\u7b97\u6210\u672c\u7684\u5145\u5206\u5c55\u5f00\u3002<\/p>\n Route record\uff1aPrimary adapter = method-algorithm\uff1bSecondary adapter = benchmark-evaluation\uff1bEvidence packs = general\u3001experimental-eval\u3001ablation-and-mechanism-isolation\u3001reproducibility-and-compute\uff1bRoute confidence = \u4e2d\u9ad8\u3002\u8be5\u8def\u7ebf\u5fe0\u5b9e\u4e8e\u8bba\u6587\uff1a\u4e3b\u8981\u662f\u65b0\u534a\u76d1\u7763\u7b97\u6cd5\uff0c\u53ef\u4fe1\u5ea6\u53d6\u51b3\u4e8e\u591a\u6570\u636e\u96c6\u6bd4\u8f83\u548c\u6d88\u878d\u3002<\/p>\n \u4f5c\u8005\u8ba4\u4e3a SMIS \u73b0\u6709\u65b9\u6cd5\u6709\u4e09\u7c7b\u95ee\u9898\u3002\u7b2c\u4e00\uff0cpseudo-labeling \u65b9\u6cd5\u521d\u59cb\u9884\u6d4b\u566a\u58f0\u4f1a\u88ab\u53cd\u590d\u5f3a\u5316\uff0c\u5bfc\u81f4\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u3002\u7b2c\u4e8c\uff0cMean Teacher\/consistency learning \u867d\u7136\u5229\u7528\u6270\u52a8\u4e0d\u53d8\u6027\uff0c\u4f46 teacher \u7684\u9519\u8bef\u4f1a\u7d2f\u79ef\uff0c\u5c24\u5176\u5728\u5c11 label \u4e0b\u5f71\u54cd student\u3002\u7b2c\u4e09\uff0cdual-stream mutual learning \u591a\u4ecd\u505c\u7559\u5728\u5224\u522b\u5f0f per-pixel \u5206\u7c7b\uff0c\u6ca1\u6709\u663e\u5f0f\u5efa\u6a21 image structure \u4e0e mask distribution \u7684\u5173\u7cfb\u3002\u4f5c\u8005\u8fd8\u6307\u51fa GAN\/VAE\/diffusion \u7b49\u751f\u6210\u6a21\u578b\u5728\u533b\u5b66\u56fe\u50cf\u4e2d\u5e38\u7528\u4e8e\u6570\u636e\u589e\u5f3a\u6216\u91cd\u5efa\uff0c\u4f46\u201c\u76f4\u63a5\u628a\u751f\u6210\u6a21\u578b\u7528\u4e8e segmentation mask generation\u201d\u7684\u63a2\u7d22\u4ecd\u6709\u9650\uff1b\u5bf9\u6297\u8bad\u7ec3\u53c8\u6709\u6536\u655b\u96be\u9898\u3002<\/p>\n SemiGDA \u7684\u6574\u4f53\u6846\u67b6\u7531\u4e09\u5757\u7ec4\u6210\uff1aDual-distribution Alignment Module\uff08DAM\uff09\u3001Consistency-Driven Skip Adapter\uff08CDSA\uff09\u548c Annotation Conversion\/Reversion\uff08ACR\uff09\u3002\u8bad\u7ec3\u6570\u636e\u5305\u542b labeled set (D_L={(x_i^l,y_i^l)}) \u4e0e unlabeled set (D_U={x_i^u})\uff0c\u5176\u4e2d (N_l \\ll N_u)\u3002<\/p>\n \u6d41\u7a0b\u5982\u4e0b\uff1a DAM\uff08Dual-distribution Alignment Module\uff09<\/strong>\uff1a\u8f93\u5165\u662f\u56fe\u50cf (x) \u4e0e\u6709\u6807\u6ce8\u6837\u672c\u7684 mask (g)\u3002\u51bb\u7ed3 VAE encoder \u7ed9\u51fa (p(z_v|x)=\\mathcal{N}(z_v;\\mu_{z_v},\\sigma_{z_v}))\uff0clatent mapping model \u7ed9\u51fa (p(\\tilde z_v|z_v))\uff0ctrainable encoder \u7ed9\u51fa (p(z_r|x))\uff0cmask \u7ecf VAE encoder \u7ed9\u51fa (p(z_g|g))\u3002\u5bf9 labeled data\uff0c\u635f\u5931\u4e3a (|\\tilde z_v^l-z_g|_2^2+|z_r^l-z_g|_2^2)\uff1b\u5bf9 unlabeled data\uff0c\u7ea6\u675f (|\\tilde z_v^u-z_r^u|_2^2)\u3002\u5b83\u89e3\u51b3\u7684\u662f\u201c\u4ec5\u9760\u6700\u7ec8 mask loss \u5bf9\u4f4e\u6807\u6ce8\u6837\u672c\u76d1\u7763\u592a\u5f31\u201d\u7684\u95ee\u9898\uff0c\u628a\u76d1\u7763\u63d0\u524d\u5230 latent distribution \u5c42\u3002\u521b\u65b0\u6027\u8f83\u660e\u786e\uff0c\u4f46\u516c\u5f0f\u4e2d\u4f7f\u7528 MSE \u5bf9\u9f50 Gaussian latent \u7684\u5747\u503c\/\u6837\u672c\u8868\u5f81\uff0c\u4e25\u683c\u6982\u7387\u610f\u4e49\u5e76\u4e0d\u5145\u5206\uff1b\u66f4\u50cf feature distribution regularization\u3002<\/p>\n Latent Mapping Model<\/strong>\uff1a\u4f7f\u7528 self-attention \u5c06 image latent \u6620\u5c04\u5230 mask-compatible latent manifold\u3002\u5b83\u662f DAM \u7684\u5173\u952e\uff0c\u56e0\u4e3a\u76f4\u63a5\u7528\u56fe\u50cf VAE latent \u751f\u6210 mask \u901a\u5e38\u4e0d\u6210\u7acb\u3002\u9002\u5408\u8fc1\u79fb\u5230\u5176\u4ed6\u6846\u67b6\u4e2d\u7684\u5730\u65b9\u662f\u201cimage feature \u2192 mask prior\/prototype space\u201d\u7684\u6620\u5c04\u601d\u60f3\uff0c\u800c\u975e\u5fc5\u987b\u4f7f\u7528 SD VAE\u3002<\/p>\n CDSA\uff08Consistency-Driven Skip Adapter\uff09<\/strong>\uff1a\u8f93\u5165\u662f\u4e24\u4e2a\u5206\u652f\u7684\u591a\u5c3a\u5ea6 feature bank\uff1aVAE\/image-distribution encoder \u7684 (S_v={\\mathcal{E}^{(i)}(x)}{i=1}^{N_f})\uff0ctrainable\/mask-distribution encoder \u7684 (S_r={E^{(i)}(x)}<\/em>)\u3002Image Skip Adapter \u548c Mask Skip Adapter \u4f7f\u7528\u8f7b\u91cf\u5377\u79ef\u5c42\u63a5\u5165\u51bb\u7ed3 VAE decoder \u7684 skip\/upsampling \u4f4d\u7f6e\u3002\u5bf9 labeled data\uff0c\u4e24\u4e2a adapter \u8f93\u51fa\u90fd\u7528 Dice loss \u5bf9\u9f50 GT\uff1b\u5bf9 unlabeled data\uff0c\u4e24\u4e2a\u8f93\u51fa\u4e92\u76f8 Dice consistency\u3002\u5b83\u89e3\u51b3 VAE decoder \u751f\u6210 mask \u65f6\u7ec6\u8282\u4e0d\u8db3\u548c\u8fb9\u754c\u7c97\u7cd9\u7684\u95ee\u9898\u3002\u8fd9\u4e2a\u6a21\u5757\u5bf9 U-Net\/nnU-Net\/DAMamba \u6539\u9020\u5f88\u6709\u53c2\u8003\u4ef7\u503c\uff1a\u53ef\u4ee5\u628a\u201c\u4e24\u4e2a\u6765\u6e90\u7684 skip feature + \u8f93\u51fa\u4e00\u81f4\u6027\u201d\u8fc1\u79fb\u5230 encoder-decoder segmentation framework \u4e2d\u3002}^{N_f<\/p>\n ACR\uff08Annotation Conversion and Reversion\uff09<\/strong>\uff1a\u628a mask \u50cf\u7d20\u503c\u5148\u5f52\u4e00\u5316\u5230 [0,1]\uff0c\u518d\u6620\u5c04\u5230 [-1,1]\uff0c\u4f7f GT mask \u7b26\u5408 VAE \u8f93\u5165\u5206\u5e03\uff1b\u8f93\u51fa\u540e\u518d\u53cd\u53d8\u6362\u3002\u8fd9\u662f\u5de5\u7a0b\u4e0a\u5fc5\u8981\u7684 compatibility trick\uff0c\u521b\u65b0\u6027\u4e0d\u5f3a\uff0c\u4f46\u5982\u679c\u590d\u73b0\u751f\u6210\u5f0f\u5206\u5272\u5f88\u5173\u952e\u3002<\/p>\n Overall loss<\/strong>\uff1a(L_{total}=L_{sup}+\\lambda_u L_{unsup})\uff0c\u5176\u4e2d (L_{sup}=L^p_{sup}+L^s_{sup})\uff0c(L_{unsup}=L^p_{unsup}+L^s_{unsup})\u3002\u5b83\u5c06 latent distribution alignment \u4e0e segmentation output consistency \u7ed1\u5b9a\u5728\u4e00\u8d77\uff0c\u662f\u8bba\u6587\u76f8\u5bf9\u4f20\u7edf SMIS \u7684\u4e3b\u8981\u533a\u522b\u3002<\/p>\n \u5bf9 polyp segmentation\uff1a\u975e\u5e38\u76f8\u5173\uff0c\u5305\u542b CVC-ClinicDB\u3001Kvasir\u3001CVC-300\u3002\u5bf9 3D medical segmentation\uff1a\u76ee\u524d\u4e3b\u8981\u662f 2D 224\u00d7224 \u8bbe\u7f6e\uff0c\u4e0d\u80fd\u76f4\u63a5\u7528\u4e8e 3D\uff1b\u4f46 DAM\/CDSA \u601d\u8def\u53ef\u8fc1\u79fb\u5230 3D VAE \u6216 3D latent prior\u3002<\/p>\n \u8bba\u6587\u5728\u56db\u7c7b\u533b\u5b66\u5206\u5272\u4efb\u52a1\u4e0a\u8bc4\u4f30\uff1acolonoscopy\uff08CVC-ClinicDB\u3001Kvasir\u3001CVC-300\uff09\u3001ISIC-2018\u3001BCSS pathology\u3001BUSI breast ultrasound\u3002\u6307\u6807\u4e3a Dice\u3001IoU\u300195HD\uff1blabel ratio \u4e3b\u8981\u4e3a 10% \u548c 30%\uff0c\u5e76\u8865\u5145\u4e0d\u540c labeled ratio \u66f2\u7ebf\u3002\u5b9e\u73b0\u4e0a\u7528 PyTorch 2.4.1\u3001CUDA 11.2\u3001\u4e24\u5f20 NVIDIA 4090\uff1bStable Diffusion VAE \u6743\u91cd\u4f5c\u4e3a\u9884\u8bad\u7ec3 encoder\/decoder\uff1bbatch size 4\uff082 labeled + 2 unlabeled\uff09\uff1b\u8f93\u5165 resize \u5230 224\u00d7224\uff1b\u5148\u9884\u8bad\u7ec3 mapping network \u4e0e encoder 200 epochs\uff0c\u518d\u5168\u6a21\u578b\u8bad\u7ec3 350 epochs\uff1b\u63a8\u7406\u65f6\u53d6\u4e24\u4e2a\u9884\u6d4b\u5747\u503c\u3002<\/p>\n \u4e3b\u8981\u7ed3\u679c\uff1a \u53ef\u4fe1\u5ea6\u4e2d\u9ad8\u3002\u4f18\u70b9\u662f baseline \u8f83\u5f3a\u4e14\u8986\u76d6\u8fd1\u5e74 SMIS \u65b9\u6cd5\uff0c\u5982 UA-MT\u3001DTC\u3001MC-Net\u3001URPC\u3001MCF\u3001CauSSL\u3001CDMA\u3001BS-Net\u3001PMT\u3001VCLIPSeg\u3001UnCo\u3001SKCDF\u3001CSCPA\uff1b\u4efb\u52a1\u8986\u76d6 endoscopy polyp\u3001skin lesion\u3001pathology\u3001ultrasound\uff0c\u6bd4\u8f83\u8d34\u8fd1\u533b\u5b66\u5206\u5272\u4e3b\u6d41\u573a\u666f\uff1b\u6d88\u878d\u56f4\u7ed5 DAM\u3001CDSA\u3001skip adapters \u548c loss functions\uff0c\u80fd\u652f\u6491\u6838\u5fc3\u6a21\u5757\u786e\u5b9e\u6709\u8d21\u732e\u3002<\/p>\n \u4e0d\u8db3\u4e5f\u660e\u663e\u3002\u7b2c\u4e00\uff0c\u8bba\u6587\u6ca1\u6709\u5145\u5206\u62a5\u544a\u7edf\u8ba1\u663e\u8457\u6027\u3001\u65b9\u5dee\u6216\u591a\u968f\u673a\u79cd\u5b50\uff1b\u534a\u76d1\u7763\u4f4e\u6807\u6ce8\u5212\u5206\u5bf9\u7ed3\u679c\u654f\u611f\uff0c\u5355\u6b21 split \u53ef\u80fd\u9ad8\u4f30\u6536\u76ca\u3002\u7b2c\u4e8c\uff0c\u8f93\u5165\u7edf\u4e00 resize \u5230 224\u00d7224\uff0c\u53ef\u80fd\u727a\u7272\u5c0f\u75c5\u7076\/\u7ec6\u8fb9\u754c\u4fe1\u606f\uff1b\u5bf9\u9ad8\u5206\u8fa8\u7387 pathology \u4e0e\u5185\u955c\u8fb9\u754c\u8bc4\u4ef7\u8981\u8c28\u614e\u3002\u7b2c\u4e09\uff0cStable Diffusion VAE \u662f\u81ea\u7136\u56fe\u50cf\u9884\u8bad\u7ec3\u5148\u9a8c\uff0c\u4e3a\u4ec0\u4e48\u5b83\u7684 latent manifold \u9002\u5408\u533b\u5b66 mask \u751f\u6210\uff0c\u8bba\u6587\u66f4\u591a\u9760\u5b9e\u9a8c\u8bf4\u660e\uff0c\u7406\u8bba\u89e3\u91ca\u6709\u9650\u3002\u7b2c\u56db\uff0c\u8ba1\u7b97\u6210\u672c\u4e0d\u4f4e\uff1a200 epoch \u9884\u8bad\u7ec3 + 350 epoch full training + \u53cc\u5206\u652f + VAE decoder\uff1b\u4f46\u8bba\u6587\u4e3b\u6587\u6ca1\u6709\u7ed9\u51fa\u5145\u5206 FLOPs\/\u53c2\u6570\/\u8bad\u7ec3\u65f6\u95f4\u5bf9\u6bd4\u3002\u7b2c\u4e94\uff0c\u5b9e\u9a8c\u662f 2D \u5206\u5272\uff0c\u6ca1\u6709\u9a8c\u8bc1 3D CT\/MRI\uff0c\u4e5f\u6ca1\u6709\u8de8\u57df OOD \u6d4b\u8bd5\u3002<\/p>\n \u4e0e U-Net\/nnU-Net \u7684\u5173\u7cfb\uff1aSemiGDA \u4e0d\u662f\u81ea\u914d\u7f6e U-Net\uff0c\u4e5f\u4e0d\u662f\u5355\u7eaf encoder-decoder \u7ed3\u6784\u6539\u9020\uff1b\u5b83\u628a segmentation \u770b\u6210 mask generation\uff0c\u628a VAE latent prior \u548c mask prior \u5bf9\u9f50\u653e\u5728\u6838\u5fc3\u4f4d\u7f6e\u3002nnU-Net \u53ef\u4f5c\u4e3a\u5f3a supervised baseline\uff0c\u4f46 SemiGDA \u9762\u5411\u4f4e\u6807\u6ce8\u534a\u76d1\u7763\u573a\u666f\u3002<\/p>\n \u4e0e MedNeXt\/CNN-based segmentation \u7684\u5173\u7cfb\uff1atrainable encoder \u53ef\u89c6\u4f5c CNN\/ResNet \u5206\u652f\uff0cCDSA \u7684\u591a\u5c3a\u5ea6 skip adapter \u4e0e CNN encoder-decoder \u5f88\u63a5\u8fd1\uff1b\u5982\u679c\u628a ResNet \u6362\u6210 MedNeXt block\uff0c\u53ef\u80fd\u5f62\u6210\u66f4\u5f3a\u7684\u4f4e\u6807\u6ce8\u534a\u76d1\u7763 backbone\u3002<\/p>\n \u4e0e UNetR\/Swin-UNet\/TransUNet\/TransFuse \u7684\u5173\u7cfb\uff1a\u8fd9\u4e9b\u65b9\u6cd5\u4e3b\u8981\u6539\u5584 encoder long-range dependency\uff1bSemiGDA \u7684 self-attention latent mapping \u53ea\u7528\u4e8e image-to-mask latent transformation\uff0c\u4e0d\u662f\u5b8c\u6574 Transformer segmentation backbone\u3002\u53ef\u4ee5\u628a DAM \u63a5\u5230 Transformer encoder \u8f93\u51fa\u4e0a\u505a mask latent prior alignment\u3002<\/p>\n \u4e0e Mamba\/VMamba\/SegMamba\/DAMamba \u7684\u5173\u7cfb\uff1aSemiGDA \u4e0e Mamba \u6ca1\u6709\u76f4\u63a5\u5173\u7cfb\uff0c\u4f46 CDSA\/DAM \u5bf9 DAMamba \u6539\u9020\u6709\u542f\u53d1\uff1aDAMamba \u7684\u72b6\u6001\u7a7a\u95f4\u5206\u652f\u53ef\u63d0\u4f9b (z_r)\uff0c\u53e6\u4e00\u4e2a frozen\/generative prior \u5206\u652f\u63d0\u4f9b (z_v)\uff0c\u901a\u8fc7 distribution consistency \u7ea6\u675f unlabeled data\uff1b\u4e5f\u53ef\u4ee5\u628a Mamba encoder \u7684\u591a\u5c3a\u5ea6\u7279\u5f81\u4f5c\u4e3a skip adapter \u8f93\u5165\uff0c\u589e\u5f3a\u4f4e\u6807\u6ce8\u9c81\u68d2\u6027\u3002<\/p>\n \u4e0e foundation model\/SAM\/MedSAM \u7684\u5173\u7cfb\uff1aSemiGDA \u4f7f\u7528 Stable Diffusion VAE \u4f5c\u4e3a\u751f\u6210\u5f0f foundation prior\uff0c\u800c\u4e0d\u662f SAM \u5f0f promptable mask decoder\uff1b\u76f8\u6bd4 SAM\/MedSAM\uff0c\u5b83\u66f4\u504f\u8bad\u7ec3\u8303\u5f0f\u548c\u534a\u76d1\u7763\u5b66\u4e60\uff0c\u4e0d\u662f\u4ea4\u4e92\u5f0f\u5206\u5272\u6a21\u578b\u3002\u4e0e foundation model \u7684\u5173\u7cfb\u5728\u4e8e\u201c\u501f\u7528\u5927\u6a21\u578b latent space\u201d\uff0c\u800c\u4e0d\u662f\u7aef\u5230\u7aef\u5927\u6a21\u578b\u5206\u5272\u3002<\/p>\n \u5bf9 polyp segmentation\uff1a\u4ef7\u503c\u5f88\u9ad8\u3002\u5b83\u76f4\u63a5\u5728 CVC-ClinicDB\u3001Kvasir\u3001CVC-300 \u4e0a\u6d4b\u8bd5\uff0c\u4e14 10%\/30% labeled setting \u9002\u5408\u7814\u7a76\u201c\u5c11\u6807\u6ce8\u606f\u8089\u5206\u5272\u201d\u3002\u5982\u679c\u7528\u6237\u5173\u6ce8 polyp segmentation\uff0c\u53ef\u628a\u5b83\u4f5c\u4e3a semi-supervised baseline \u6216\u4f4e\u6807\u6ce8\u5b9e\u9a8c\u53c2\u8003\u3002<\/p>\n \u5bf9 DAMamba \u6539\u9020\uff1a\u5efa\u8bae\u501f\u9274\u4e24\u4e2a\u601d\u60f3\u3002\u7b2c\u4e00\uff0clatent\/distribution-level consistency \u6bd4\u5355\u7eaf output pseudo-label \u66f4\u7a33\u5b9a\uff0c\u53ef\u7528\u4e8e DAMamba \u534a\u76d1\u7763\u7248\u672c\uff1b\u7b2c\u4e8c\uff0cCDSA \u7684\u53cc skip adapter \u53ef\u8fc1\u79fb\u5230 Mamba\/U-Net decoder \u4e2d\uff0c\u7528\u4e0d\u540c\u5206\u652f\u7279\u5f81\u7684\u4e00\u81f4\u6027\u63d0\u5347\u8fb9\u754c\u8d28\u91cf\u3002\u4e0d\u8981\u76f4\u63a5\u7167\u642c SD VAE\uff0c\u9664\u975e\u76ee\u6807\u4efb\u52a1\u662f 2D \u4e14\u53ef\u63a5\u53d7 224\u00d7224 \u8f93\u5165\u4e0e\u8f83\u9ad8\u8bad\u7ec3\u6210\u672c\u3002<\/p>\n \u5bf9 related work\uff1a\u9002\u5408\u653e\u5728\u201csemi-supervised medical segmentation\u201d\u548c\u201cgenerative prior for segmentation\u201d\u4e24\u6bb5\uff1b\u5982\u679c\u8bba\u6587\u4e3b\u9898\u662f Mamba \u6216 U-Net architecture\uff0c\u5219\u53ef\u4f5c\u4e3a\u4f4e\u6807\u6ce8\u8bad\u7ec3\u7b56\u7565\u800c\u975e\u4e3b\u5e72\u67b6\u6784 baseline\u3002<\/p>\n \u5efa\u8bae\u7cbe\u8bfb<\/strong>\uff0c\u5c24\u5176\u9002\u5408\u6b63\u5728\u505a\u4f4e\u6807\u6ce8 polyp segmentation \u6216\u60f3\u628a DAMamba \u6269\u5c55\u5230 semi-supervised setting \u7684\u573a\u666f\u3002\u9605\u8bfb\u4f18\u5148\u7ea7\u4e3a Section 3.1 DAM\u3001Section 3.2 CDSA\u3001Table 1\u20135 \u548c Fig. 5\/6\uff1b\u5982\u679c\u53ea\u505a fully supervised 3D segmentation\uff0c\u53ef\u7565\u8bfb\u5b9e\u9a8c\u8bbe\u7f6e\u5e76\u91cd\u70b9\u5438\u6536 consistency\/design ideas\u3002<\/p>\n \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 2026-05-27 \u5f53\u65e5\u521a\u53d1\u5e03\u4e14\u8d28\u91cf\u8db3\u591f\u7a33\u5b9a\u7684\u533b\u5b66\u56fe\u50cf\u5206\u5272\u9876\u4f1a\/\u9876\u520a …<\/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-1078","post","type-post","status-publish","format-standard","hentry","category-85"],"views":12,"_links":{"self":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1078","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=1078"}],"version-history":[{"count":1,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1078\/revisions"}],"predecessor-version":[{"id":1079,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1078\/revisions\/1079"}],"wp:attachment":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/media?parent=1078"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/categories?post=1078"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/tags?post=1078"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
\n1. CLIP tokenization and encoding<\/strong>\uff1a\u56fe\u50cf\u5f97\u5230 (Z_v \\in \\mathbb{R}^{B\\times(P+1)\\times D})\uff0c\u6587\u672c\u5f97\u5230 (Z_t \\in \\mathbb{R}^{B\\times L\\times D})\u3002
\n2. Down projection<\/strong>\uff1a\u628a\u89c6\u89c9\/\u6587\u672c token \u6295\u5f71\u5230\u8f83\u4f4e\u7ef4\u5171\u4eab\u7a7a\u95f4 (D_s)\uff0c\u964d\u4f4e adapter \u6210\u672c\u3002
\n3. Probabilistic bidirectional attention<\/strong>\uff1a\u5728\u89c6\u89c9\u5230\u6587\u672c\u3001\u6587\u672c\u5230\u89c6\u89c9\u4e24\u4e2a\u65b9\u5411\u4e0a\u6267\u884c AttnPVL\uff1bKey \u548c Value \u4e0d\u518d\u662f\u786e\u5b9a\u5411\u91cf\uff0c\u800c\u662f\u5747\u503c\u548c\u65b9\u5dee\u3002
\n4. Segmentation via pixel-text similarity<\/strong>\uff1a\u6700\u7ec8 visual patch token \u7ecf\u8fc7\u4e0a\u91c7\u6837\u5757\uff0c\u6587\u672c [EOS] token \u7ecf MLP mask head \u540e\uff0c\u4e0e\u6bcf\u4e2a\u50cf\u7d20\/patch \u505a dot product \u5f97\u5230 mask logits\u3002
\n5. Training objective<\/strong>\uff1aDice+BCE \u5206\u5272\u635f\u5931\u4e0e soft patch-level contrastive loss \u7ed3\u5408\uff0c\u540e\u8005\u7528\u6587\u672c\u76f8\u4f3c\u5ea6\u6784\u9020 soft target\uff0c\u907f\u514d\u628a\u8bed\u4e49\u76f8\u8fd1 prompt \u5f3a\u884c\u5f53\u4f5c\u5b8c\u5168\u4e0d\u540c\u7c7b\u522b\u3002<\/p>\n5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n
6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n
\n- \u6570\u636e\u6548\u7387 Table 1\uff1aMedCLIPSeg \u5728 10% \u6570\u636e\u4e0b DSC 81.10\u3001NSD 83.94\uff0c\u4f18\u4e8e CAT-Seg \u7684 78.76\/81.50\uff0c\u4e5f\u660e\u663e\u4f18\u4e8e nnU-Net \u7684 73.45\/77.37\uff1b100% \u6570\u636e\u4e0b MedCLIPSeg \u4e3a 88.66\/91.35\u3002
\n- \u8de8\u57df\u6cdb\u5316 Table 2\uff1a\u5728 breast ultrasound\u3001polyp endoscopy\u3001brain MRI\u3001skin dermatoscopy \u591a\u4e2a source-target \u8bbe\u7f6e\u4e2d\uff0cMedCLIPSeg \u591a\u6570\u6307\u6807\u9886\u5148\u3002\u4f8b\u5982 Kvasir-SEG source \u4e0a DSC 90.15\uff0cCVC-ColonDB target 71.90\uff0cClinicDB target 80.80\uff0cCVC300 target 80.82\uff0cBKAI target 79.15\u3002
\n- \u5173\u952e\u6d88\u878d Table 3\uff1a\u53bb\u6389 PVL Adapter \u540e OOD DSC \u4ece 79.02 \u964d\u5230 55.23\uff1b\u786e\u5b9a\u6027 MedCLIPSeg \u7684 OOD DSC \u4e3a 63.12\uff0c\u8bf4\u660e\u6982\u7387\u6ce8\u610f\u529b\u5bf9 OOD \u7684\u8d21\u732e\u662f\u4f5c\u8005\u6700\u5f3a\u8bc1\u636e\u4e4b\u4e00\u3002
\n- Prompt \u6d88\u878d Table 4\uff1a\u539f\u59cb concise prompt \u7684 HM DSC 83.76\uff1bunderdescriptive prompt \u964d\u81f3 56.82\uff0c\u8bf4\u660e\u6a21\u578b\u9ad8\u5ea6\u4f9d\u8d56 prompt \u8d28\u91cf\u3002
\n- Backbone \u6d88\u878d Table 5\uff1aUniMedCLIP \u4f18\u4e8e CLIP\u3001PubMedCLIP\u3001BiomedCLIP\uff0c\u8bf4\u660e\u9884\u8bad\u7ec3 backbone \u5bf9\u7ed3\u679c\u5f71\u54cd\u663e\u8457\u3002
\n- Reliability\uff1auncertainty \u4e0e segmentation error \u7684 Spearman correlation \u5728 ID\/OOD \u5206\u522b\u4e3a 87.57%\/80.41%\uff1bBrier score \u4ece deterministic baseline \u7684 23.9%\/25.3% \u964d\u5230 11.1%\/11.8%\u3002<\/p>\n7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n
8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n
9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n
10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n
\n\u8bba\u6587 2\uff1aSemiGDA: Generative Dual-distribution Alignment for Semi-Supervised Medical Image Segmentation<\/h2>\n
\u57fa\u672c\u4fe1\u606f<\/h3>\n
\n
paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n
1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n
2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n
3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n
4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n
\n1. \u51bb\u7ed3 VAE \u5206\u652f<\/strong>\uff1a\u8f93\u5165\u56fe\u50cf (x) \u7ecf\u8fc7\u9884\u8bad\u7ec3 Stable Diffusion VAE encoder (\\mathcal{E})\uff0c\u5f97\u5230\u56fe\u50cf\u5148\u9a8c\u5206\u5e03 (p(z_v|x))\u3002
\n2. latent mapping<\/strong>\uff1a\u4f7f\u7528 self-attention latent mapping model (\\mathcal{M}) \u628a (z_v) \u6620\u5c04\u4e3a (\\tilde z_v)\uff0c\u671f\u671b\u5176\u63a5\u8fd1 mask latent distribution\u3002
\n3. trainable encoder \u5206\u652f<\/strong>\uff1a\u540c\u4e00\u56fe\u50cf\u8fd8\u8fdb\u5165\u53ef\u8bad\u7ec3 encoder (E)\uff08ResNet backbone\uff09\uff0c\u5f97\u5230 (p(z_r|x))\uff0c\u7528\u4e8e\u63d0\u53d6\u66f4\u5224\u522b\u5f0f\/\u7ec6\u7c92\u5ea6\u7ed3\u6784\u7279\u5f81\u3002
\n4. mask prior supervision<\/strong>\uff1a\u5bf9 labeled data\uff0cground-truth mask (g) \u4e5f\u7ecf VAE encoder \u5f97\u5230 (p(z_g|g))\uff0c\u4f5c\u4e3a\u56fe\u50cf\u5206\u652f\u548c trainable \u5206\u652f\u7684 latent \u5bf9\u9f50\u76ee\u6807\u3002
\n5. VAE decoder + skip adapters<\/strong>\uff1a\u6620\u5c04\u540e\u7684 latent \u9001\u5165\u51bb\u7ed3 VAE decoder\uff1bCDSA \u5728 decoder skip \u4f4d\u7f6e\u5f15\u5165 Image Skip Adapter \u548c Mask Skip Adapter\uff0c\u878d\u5408\u591a\u5c3a\u5ea6\u4fe1\u606f\u5e76\u5bf9 unlabeled data \u65bd\u52a0\u8f93\u51fa\u4e00\u81f4\u6027\u3002
\n6. loss<\/strong>\uff1a\u603b\u635f\u5931\u4e3a supervised distribution loss + supervised segmentation loss + (\\lambda_u) times unsupervised distribution\/output consistency loss\uff1b(\\lambda_u) \u7528 Gaussian warm-up\uff0c(\\beta=0.1)\u3002<\/p>\n5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n
6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n
\n- Colonoscopy\/ISIC Table 1\uff1a\u5728 CVC-300 10% labeled \u4e0b Ours Dice 84.34\u3001IoU 76.28\u300195HD 3.19\uff0c\u660e\u663e\u9ad8\u4e8e UnCo 77.56 Dice\u3001CSCPA 76.97 Dice\uff1b\u5728 Kvasir 10% labeled \u4e0b Ours Dice 83.03\uff0c\u9ad8\u4e8e UnCo 81.19\u3001CSCPA 81.60\uff1bISIC-2018 10% labeled \u4e0b Ours Dice 86.28\uff0c\u7565\u9ad8\u4e8e CSCPA 85.75\u3002
\n- BCSS\/BUSI Table 2\uff1aBCSS 10% labeled \u4e0b Ours Dice 74.05\u3001IoU 62.68\u300195HD 7.05\uff0c\u4f18\u4e8e CSCPA 71.95 Dice\uff1bBUSI 10% labeled \u4e0b Ours Dice 75.57\u3001IoU 65.72\uff0c\u663e\u8457\u9ad8\u4e8e CSCPA 65.16 Dice\u3002
\n- Ablation Table 3\uff1abaseline \u5728 BUSI 10% Dice 70.48\uff1b\u52a0\u5165 DAM \u540e 73.07\uff1b\u52a0\u5165 CDSA \u540e 75.25\uff1b\u5b8c\u6574\u6a21\u578b 75.57\u3002ClinicDB\/Kvasir \u4e5f\u5448\u7a33\u5b9a\u4e0a\u5347\u3002
\n- Skip adapter ablation Table 4\uff1aBUSI 10% \u4ece\u65e0 adapter \u7684 73.07 \u63d0\u5347\u5230\u53cc adapter \u7684 75.57\uff1bKvasir 10% \u4ece 80.02 \u63d0\u5347\u5230 83.03\u3002
\n- Loss ablation Table 5\uff1aKvasir 10% \u5b8c\u6574\u635f\u5931 Dice 83.03\uff0c\u800c\u53ea\u4fdd\u7559\u90e8\u5206 loss \u4f1a\u964d\u5230 78.61\u201382.64 \u533a\u95f4\uff1b\u8bf4\u660e unsupervised distribution\/output consistency \u5bf9\u4f4e\u6807\u6ce8\u6709\u6548\u3002<\/p>\n7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n
8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n
9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n
10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n
\n\u4eca\u65e5\u63a8\u8350\u4f18\u5148\u7ea7<\/h2>\n
\n
\u4eca\u65e5 PDF \u83b7\u53d6\u60c5\u51b5<\/h2>\n
\n
\u4eca\u65e5\u53ef\u6267\u884c\u5efa\u8bae<\/h2>\n
\n