{"id":1062,"date":"2026-05-15T08:35:50","date_gmt":"2026-05-15T00:35:50","guid":{"rendered":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/15\/2026-05-15-%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%9amedcore-%e8%be%b9%e7%95%8c%e4%bf%9d%e7%9c%9f-medsam-%e5%89%aa%e6%9e%9d\/"},"modified":"2026-05-15T08:35:50","modified_gmt":"2026-05-15T00:35:50","slug":"2026-05-15-%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%9amedcore-%e8%be%b9%e7%95%8c%e4%bf%9d%e7%9c%9f-medsam-%e5%89%aa%e6%9e%9d","status":"publish","type":"post","link":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/15\/2026-05-15-%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%9amedcore-%e8%be%b9%e7%95%8c%e4%bf%9d%e7%9c%9f-medsam-%e5%89%aa%e6%9e%9d\/","title":{"rendered":"2026-05-15 \u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u7cbe\u8bfb\uff1aMedCore \u8fb9\u754c\u4fdd\u771f MedSAM \u526a\u679d"},"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\u5728 2026-05-13 \u4e4b\u540e\u7684\u65b0\u589e\u5019\u9009\u4e2d\uff0c\u771f\u6b63\u7b26\u5408\u201c\u533b\u5b66\u56fe\u50cf\u5206\u5272\u4e3b\u4efb\u52a1\u30012025 \u5e74\u4ee5\u540e\u3001\u672a\u88ab\u672c\u5b9a\u65f6\u4efb\u52a1\u63a8\u8350\u8fc7\u3001\u4e14\u503c\u5f97\u7cbe\u8bfb\u201d\u7684\u65b0\u8bba\u6587\u53ea\u6709 1 \u7bc7\uff1a<strong>MedCore: Boundary-Preserving Medical Core Pruning for MedSAM<\/strong>\u3002\u7531\u4e8e\u6628\u5929\u5df2\u7ecf\u63a8\u8350\u8fc7 FEFormer \u4e0e USEMA\uff0c\u4eca\u5929\u6ca1\u6709\u5f3a\u884c\u51d1\u7b2c\u4e8c\u7bc7\u4f4e\u8d28\u91cf\u5019\u9009\uff1b\u7b2c\u4e8c\u7bc7\u4f4d\u7f6e\u6539\u4e3a\u201c\u5019\u9009\u4f46\u672a\u5165\u9009\u201d\u7684\u77ed\u8bc4\uff0c\u4ee5\u514d\u91cd\u590d\u63a8\u8350\u6216\u8bef\u5bfc\u3002<\/p>\n<p>\u6574\u4f53\u8d8b\u52bf\u662f\uff1a\u533b\u5b66\u56fe\u50cf\u5206\u5272\u7684\u65b0\u5de5\u4f5c\u7ee7\u7eed\u4ece\u201c\u5355\u7eaf\u63d0\u5347 Dice\u201d\u8f6c\u5411 <strong>foundation model \u538b\u7f29\u3001\u8fb9\u754c\u4fdd\u771f\u3001\u8de8\u57df\u6cdb\u5316\u548c\u90e8\u7f72\u6548\u7387<\/strong>\u3002MedCore \u5c24\u5176\u503c\u5f97\u5173\u6ce8\uff0c\u56e0\u4e3a\u5b83\u76f4\u63a5\u56f4\u7ed5 MedSAM\u3001polyp segmentation\u3001Boundary F1 \/ HD95 \u548c\u7ed3\u6784\u5316\u526a\u679d\u5c55\u5f00\uff0c\u6bd4\u4e00\u822c\u201c\u8f7b\u91cf\u5316\u6a21\u578b\u201d\u66f4\u8d34\u8fd1\u4e34\u5e8a\u8fb9\u754c\u8d28\u91cf\u95ee\u9898\u3002<\/p>\n<h2>\u68c0\u7d22\u8bf4\u660e<\/h2>\n<p>\u4eca\u65e5\u68c0\u7d22\u8303\u56f4\u8986\u76d6 arXiv 2026-05-13 \u81f3 2026-05-15 \u7684 <code>medical image segmentation<\/code>\u3001<code>polyp segmentation<\/code>\u3001<code>SAM medical segmentation<\/code>\u3001<code>3D medical segmentation<\/code>\u3001<code>foundation medical segmentation<\/code>\u3001<code>Mamba medical segmentation<\/code> \u7b49\u5173\u952e\u8bcd\uff0c\u5e76\u56de\u770b\u4e86 2026-05-11 \u4ee5\u6765\u7684\u5019\u9009\u3002\u4eca\u5929\u672a\u53d1\u73b0\u5df2\u6b63\u5f0f\u6807\u6ce8\u4e3a MICCAI \/ CVPR \/ ICCV \/ MedIA \/ TMI \u7b49\u9876\u4f1a\u9876\u520a\u63a5\u6536\u7684\u65b0\u589e\u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\uff0c\u56e0\u6b64\u5165\u9009\u8bba\u6587\u4e3a arXiv preprint\u3002\u6240\u6709\u5165\u9009\u6216\u8ba8\u8bba\u8bba\u6587\u5747\u4e3a 2025 \u5e74\u53ca\u4ee5\u540e\u3002<\/p>\n<p>\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 <strong>FEFormer: Frequency-enhanced Vision Transformer for Generic Knowledge Extraction and Adaptive Feature Fusion in Volumetric Medical Image Segmentation<\/strong>\u3001<strong>USEMA: a Scalable Efficient Mamba Like Attention for Medical Image Segmentation<\/strong>\u3001<strong>DINO-MVR<\/strong>\u3001<strong>XTinyU-Net<\/strong>\u3001<strong>Geometry-aware Prototype Learning for Cross-domain Few-shot Medical Image Segmentation<\/strong> \u7b49\u3002<\/p>\n<h2>WordPress \u53d1\u5e03<\/h2>\n<ul>\n<li>WordPress \u6587\u7ae0\u94fe\u63a5\uff1a&lt;\u5f85\u53d1\u5e03\u540e\u586b\u5199&gt;<\/li>\n<li>WordPress Post ID\uff1a&lt;\u5f85\u53d1\u5e03\u540e\u586b\u5199&gt;<\/li>\n<\/ul>\n<hr \/>\n<h2>\u8bba\u6587 1\uff1aMedCore: Boundary-Preserving Medical Core Pruning for MedSAM<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aMedCore: Boundary-Preserving Medical Core Pruning for MedSAM<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aCenwei Zhang, Suncheng Xiang, Lei You \/ \u7b2c\u4e00\u4f5c\u8005 Cenwei Zhang<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-13<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint, arXiv:2605.13688v1<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.13688<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.13688v1 \uff08\u5df2\u4e0b\u8f7d\uff1aMEDIA:\/tmp\/medseg_daily_20260515\/medcore_2605.13688.pdf\uff09<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/cenweizhang\/MedCore<\/li>\n<li>\u4efb\u52a1\uff1aMedSAM \u7ed3\u6784\u5316\u526a\u679d\uff1bprompt-driven medical image segmentation\uff1bpolyp segmentation\u3001breast ultrasound segmentation\u3001dermoscopy segmentation<\/li>\n<li>\u6570\u636e\u96c6\uff1aCVC-ClinicDB\u3001CVC-ColonDB\u3001Kvasir-SEG\u3001BUSI\u3001ISIC2018<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1afoundation model for medical segmentation\uff1bMedSAM pruning\uff1bboundary-preserving compression\uff1bstructured pruning\uff1bboundary-aware Fisher \/ Cross-Fisher\uff1bpolyp segmentation<\/li>\n<\/ul>\n<h3>paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n<h4>1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n<p>MedCore \u7684\u6838\u5fc3\u4ef7\u503c\u662f\u628a MedSAM \u538b\u7f29\u4ece\u201c\u5220\u6389\u53c2\u6570\u4f46 Dice \u4e0d\u6389\u592a\u591a\u201d\u63d0\u5347\u4e3a\u201c\u8bc6\u522b\u5e76\u4fdd\u7559\u533b\u5b66\u9002\u914d\u540e\u771f\u6b63\u91cd\u8981\u3001\u4e14\u5bf9\u8fb9\u754c\u4f4d\u79fb\u9ad8\u6760\u6746\u7684 ViT \u7ed3\u6784\u201d\uff0c\u56e0\u6b64\u5bf9 polyp segmentation\u3001MedSAM \u90e8\u7f72\u3001\u8fb9\u754c\u8d28\u91cf\u8bc4\u4ef7\u548c\u8f7b\u91cf\u5316 foundation model \u90fd\u6709\u76f4\u63a5\u53c2\u8003\u4ef7\u503c\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>MedSAM \/ SAM \u7c7b promptable segmentation foundation model \u5728\u533b\u5b66\u56fe\u50cf\u4e2d\u5df2\u7ecf\u6709\u8f83\u5f3a mask \u751f\u6210\u80fd\u529b\uff0c\u4f46 ViT-B image encoder \u4ecd\u7136\u5f88\u91cd\uff1a\u539f\u59cb MedSAM \u5728\u8bba\u6587\u5b9e\u9a8c\u4e2d\u7ea6 <strong>89.7M \u53c2\u6570\u3001926.5G FLOPs<\/strong>\u3002\u8fd9\u5bf9\u5b9e\u65f6\u5185\u955c\u3001\u5e8a\u65c1\u8d85\u58f0\u3001\u8d44\u6e90\u53d7\u9650\u8bbe\u5907\u6216\u5927\u89c4\u6a21\u4e34\u5e8a\u90e8\u7f72\u5e76\u4e0d\u53cb\u597d\u3002<\/p>\n<p>\u666e\u901a\u538b\u7f29\u65b9\u6cd5\u5f80\u5f80\u53ea\u770b Dice \/ IoU \u6216 loss \u589e\u91cf\uff0c\u5bb9\u6613\u9690\u85cf\u533b\u5b66\u5206\u5272\u4e2d\u7279\u522b\u5371\u9669\u7684\u5931\u8d25\u6a21\u5f0f\uff1a\u9884\u6d4b\u533a\u57df\u5927\u4f53\u91cd\u5408\uff0cDice \u4ecd\u7136\u9ad8\uff0c\u4f46\u8fb9\u754c\u53d8\u539a\u3001\u65ad\u88c2\u3001\u504f\u79fb\u6216\u7ec6\u5c0f\u7ed3\u6784\u6d88\u5931\u3002\u5bf9\u4e8e\u606f\u8089\u3001\u4e73\u817a\u75c5\u7076\u3001\u76ae\u80a4\u75c5\u7076\u7b49\u4efb\u52a1\uff0c\u8fb9\u754c\u76f4\u63a5\u5f71\u54cd\u5c3a\u5bf8\u4f30\u8ba1\u3001\u5207\u9664\u8303\u56f4\u3001\u98ce\u9669\u5224\u65ad\u548c\u4e34\u5e8a\u89e3\u91ca\u3002<\/p>\n<p>paper map \u53ef\u6982\u62ec\u4e3a\uff1a\u8bba\u6587\u7814\u7a76 <strong>\u5982\u4f55\u5728\u4fdd\u7559 MedSAM \u533b\u5b66\u9002\u914d\u80fd\u529b\u548c\u8fb9\u754c\u8d28\u91cf\u7684\u524d\u63d0\u4e0b\u505a\u7ed3\u6784\u5316\u526a\u679d<\/strong>\uff1b\u4e3b\u52a8\u4f5c\u662f\u63d0\u51fa MedCore\uff0c\u7528 zero intervention\u3001reset-to-SAM intervention\u3001boundary-aware Fisher \/ Cross-Fisher \u4e0e boundary leverage \u539f\u7406\u4e3a attention heads \u548c MLP groups \u6253\u5206\uff1b\u4f5c\u8005\u58f0\u79f0\u53ef\u5728 polyp benchmarks \u4e0a\u51cf\u5c11 60.0% \u53c2\u6570\u548c 58.4% FLOPs\uff0c\u540c\u65f6\u4fdd\u6301\u751a\u81f3\u63d0\u5347 Dice\u3001BF1 \u548c HD95\uff1b\u8bc1\u636e\u4e3b\u8981\u6765\u81ea\u4e09\u7c7b\u606f\u8089\u6570\u636e\u96c6\u7684 macro-average \u8868\u3001\u8de8\u6a21\u6001 BUSI\/ISIC \u6d4b\u8bd5\u3001\u7ec4\u4ef6\u6d88\u878d\u548c head-vs-MLP boundary leverage \u5206\u6790\uff1b\u4e3b\u8981\u5931\u8d25\u98ce\u9669\u662f\u7ed3\u679c\u5f3a\u4f9d\u8d56 recovery fine-tuning\u3001prompt \/ split \/ calibration \u8bbe\u7f6e\uff0c\u4e14\u4f5c\u4e3a arXiv preprint \u4ecd\u9700\u4ee3\u7801\u548c\u72ec\u7acb\u590d\u73b0\u786e\u8ba4\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u6279\u8bc4\u7684\u4e0d\u8db3\u4e3b\u8981\u6709\u4e09\u7c7b\uff1a<\/p>\n<ol>\n<li><strong>MedSAM \/ SAM foundation model \u592a\u91cd<\/strong>\uff1aMedSAM \u7684 ViT image encoder \u8ba1\u7b97\u91cf\u5927\uff0c\u96be\u4ee5\u7528\u4e8e\u5b9e\u65f6 endoscopy\u3001point-of-care ultrasound \u7b49\u573a\u666f\u3002<\/li>\n<li><strong>\u666e\u901a\u526a\u679d \/ \u538b\u7f29\u53ea\u770b endpoint checkpoint \u6216\u53c2\u6570\u91cd\u8981\u6027<\/strong>\uff1a\u591a\u6570 pruning \u65b9\u6cd5\u53ea\u4f30\u8ba1\u4ece\u5f53\u524d\u6a21\u578b\u5220\u6389\u67d0\u7ec4\u53c2\u6570\u540e\u7684 loss \u589e\u91cf\uff0c\u5ffd\u7565\u4e86 MedSAM \u662f\u4ece SAM \u9002\u914d\u800c\u6765\u7684\uff0c\u67d0\u4e9b\u7ed3\u6784\u867d\u7136\u5728\u53c2\u6570\u91cf\u4e0a\u4e0d\u5927\uff0c\u5374\u627f\u8f7d\u4e86\u533b\u5b66\u9002\u914d\u540e\u7684\u80fd\u529b\u3002<\/li>\n<li><strong>Dice \u4e0d\u80fd\u5145\u5206\u4ee3\u8868\u533b\u5b66\u8fb9\u754c\u8d28\u91cf<\/strong>\uff1a\u538b\u7f29\u540e\u6a21\u578b\u53ef\u80fd\u4fdd\u6301\u7c97\u533a\u57df\u91cd\u53e0\uff0c\u5374\u9020\u6210\u8fb9\u754c\u4f4d\u79fb\u3002\u8bba\u6587\u660e\u786e\u628a BF1 \u548c HD95 \u4f5c\u4e3a key metrics\uff0c\u5e76\u63d0\u51fa boundary leverage \u6765\u89e3\u91ca\u4e3a\u4ec0\u4e48\u5c0f\u7684 logit perturbation \u4f1a\u5728\u4f4e\u68af\u5ea6\u8fb9\u754c\u5904\u5bfc\u81f4\u660e\u663e\u51e0\u4f55\u504f\u79fb\u3002<\/li>\n<\/ol>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>\u8def\u7ebf\u8bb0\u5f55\uff1aPrimary adapter = method-algorithm\uff1bSecondary adapter = benchmark-evaluation\uff08\u8f7b\u91cf\u4f7f\u7528\uff0c\u56e0\u4e3a\u5b9e\u9a8c\u548c\u8fb9\u754c\u6307\u6807\u662f load-bearing evidence\uff09\uff1bEvidence packs = general\u3001experimental-eval\u3001ablation-and-mechanism-isolation\u3001reproducibility-and-compute\uff1bRoute confidence = \u9ad8\u3002\u9009\u62e9\u8be5\u8def\u7ebf\u662f\u56e0\u4e3a\u8bba\u6587\u4e3b\u8981\u8d21\u732e\u662f MedSAM \u526a\u679d\u7b97\u6cd5\uff0c\u4f46\u5176\u53ef\u4fe1\u5ea6\u9ad8\u5ea6\u4f9d\u8d56\u5b9e\u9a8c\u3001\u516c\u5e73\u6bd4\u8f83\u3001\u6d88\u878d\u4e0e\u8ba1\u7b97\u91cf\u62a5\u544a\u3002<\/p>\n<p>MedCore \u7684\u6574\u4f53\u6d41\u7a0b\u5982\u4e0b\uff1a<\/p>\n<ol>\n<li>\u4ece\u5b98\u65b9 <strong>MedSAM ViT-B checkpoint<\/strong> \u51fa\u53d1\uff0c\u4fdd\u7559\u539f\u59cb <strong>SAM ViT-B checkpoint<\/strong> \u4f5c\u4e3a\u533b\u5b66\u9002\u914d\u524d\u7684\u53c2\u8003 <code>\u03b8S<\/code>\uff0cMedSAM \u53c2\u6570\u4e3a <code>\u03b8M<\/code>\u3002<\/li>\n<li>\u5c06\u53ef\u526a\u679d\u53c2\u6570\u5206\u7ec4\u4e3a ViT image encoder \u4e2d\u7684 <strong>attention heads<\/strong> \u548c <strong>MLP connection groups<\/strong>\uff1bprompt encoder \u51bb\u7ed3\u3002<\/li>\n<li>\u5bf9\u6bcf\u4e2a\u7ed3\u6784\u7ec4 <code>g<\/code> \u4f30\u8ba1\u4e24\u4e2a\u53cd\u4e8b\u5b9e\u91cd\u8981\u6027\uff1a<br \/>\n   - <strong>zero intervention<\/strong>\uff1a\u628a <code>g<\/code> \u7f6e\u96f6\uff0c\u8861\u91cf\u5b83\u5bf9\u5f53\u524d MedSAM \u9884\u6d4b\u7684\u8d21\u732e\uff1b<br \/>\n   - <strong>reset-to-SAM intervention<\/strong>\uff1a\u628a <code>g<\/code> \u4ece MedSAM \u6743\u91cd\u91cd\u7f6e\u56de\u539f\u59cb SAM \u6743\u91cd\uff0c\u8861\u91cf\u5b83\u5728 SAM\u2192MedSAM \u533b\u5b66\u9002\u914d\u4e2d\u83b7\u5f97\u7684\u529f\u80fd\u91cd\u8981\u6027\u3002<\/li>\n<li>\u4e3a\u907f\u514d\u9010\u7ec4\u771f\u5b9e\u5e72\u9884\u8fc7\u6162\uff0c\u7528 <strong>boundary-weighted Fisher<\/strong> \u8fd1\u4f3c zero cost\uff0c\u7528 <strong>Cross-Fisher<\/strong> \u8fd1\u4f3c reset cost\u3002<\/li>\n<li>\u5f15\u5165 <strong>distribution-aware aggregation<\/strong>\uff1a\u6309\u6570\u636e\u96c6 \/ \u6a21\u6001\u8ba1\u7b97 group score\uff0c\u518d\u52a0\u5165\u65b9\u5dee\u9879\uff0c\u907f\u514d\u526a\u6389\u53ea\u5728\u67d0\u4e9b\u57df\u4e2d\u91cd\u8981\u7684\u7ed3\u6784\u3002<\/li>\n<li>\u7528 block sensitivity \u505a\u975e\u5747\u5300\u9884\u7b97\u5206\u914d\uff0c\u4fdd\u62a4\u6700\u654f\u611f\u7684\u6df1\u5c42 block\uff0c\u5c24\u5176\u662f\u9760\u8fd1 mask decoder \u7684\u90e8\u5206\u3002<\/li>\n<li>\u91c7\u7528 <strong>head-to-MLP cascade pruning<\/strong>\uff1a\u5148\u526a attention heads\uff0c\u77ed\u6062\u590d\uff0c\u518d\u526a MLP groups\uff1b\u8fd9\u6837\u907f\u514d\u540c\u65f6\u7834\u574f\u4e24\u7c7b\u529f\u80fd\u7ed3\u6784\uff0c\u4e5f\u4fbf\u4e8e\u5206\u6790 head \/ MLP \u7684\u8fb9\u754c\u5f71\u54cd\u5dee\u5f02\u3002<\/li>\n<li>\u526a\u679d\u540e\u505a recovery fine-tuning\uff0c\u635f\u5931\u5305\u542b segmentation loss\u3001boundary-weighted BCE\u3001feature distillation\u3001boundary-region logit distillation \u548c high-frequency mask discrepancy\u3002<\/li>\n<\/ol>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<ul>\n<li>\n<p><strong>Dual-intervention score<\/strong>\uff1a\u8f93\u5165\u662f\u7ed3\u6784\u7ec4 <code>g<\/code>\u3001MedSAM \u6743\u91cd <code>\u03b8M<\/code> \u548c\u539f\u59cb SAM \u6743\u91cd <code>\u03b8S<\/code>\uff1b\u8f93\u51fa\u662f group priority\u3002<code>zero cost<\/code> \u8861\u91cf\u201c\u5f53\u524d\u9884\u6d4b\u662f\u5426\u9700\u8981\u8fd9\u4e2a\u7ec4\u201d\uff0c<code>reset cost<\/code> \u8861\u91cf\u201c\u533b\u5b66\u9002\u914d\u662f\u5426\u6539\u53d8\u5e76\u4f9d\u8d56\u8fd9\u4e2a\u7ec4\u201d\u3002\u8fd9\u6bd4\u53ea\u5728 MedSAM endpoint \u4e0a\u505a Fisher \u66f4\u8d34\u5408 MedSAM \u7684\u6765\u6e90\u3002\u521b\u65b0\u6027\u8f83\u5f3a\uff0c\u5c24\u5176\u9002\u5408\u4efb\u4f55\u4ece\u901a\u7528 foundation model \u9002\u914d\u5230\u533b\u5b66\u6a21\u578b\u540e\u7684\u538b\u7f29\u3002<\/p>\n<\/li>\n<li>\n<p><strong>Boundary leverage principle<\/strong>\uff1a\u8bba\u6587\u628a\u9884\u6d4b\u8fb9\u754c\u770b\u6210 logit map \u7684 zero level set\u3002\u82e5\u538b\u7f29\u64cd\u4f5c <code>G<\/code> \u9020\u6210 logit perturbation <code>\u03b4G(x)<\/code>\uff0c\u4e00\u9636\u8fb9\u754c\u6cd5\u5411\u4f4d\u79fb\u8fd1\u4f3c\u4e3a <code>-\u03b4G(x)\/||\u2207s\u03b8(x)||2<\/code>\u3002\u76f4\u89c9\u662f\uff1a\u8fb9\u754c\u5904 logit slope \u8d8a\u5c0f\uff0c\u540c\u6837\u7684 logit \u6270\u52a8\u9020\u6210\u7684\u8fb9\u754c\u4f4d\u79fb\u8d8a\u5927\u3002\u8fd9\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48 Dice \u8fd8\u53ef\u4ee5\u4f46 BF1\/HD95 \u5148\u5d29\u3002\u8be5\u90e8\u5206\u662f\u8bba\u6587\u6700\u6709\u7406\u8bba\u89e3\u91ca\u529b\u7684\u5730\u65b9\u3002<\/p>\n<\/li>\n<li>\n<p><strong>Boundary-aware Fisher \/ Cross-Fisher<\/strong>\uff1aboundary Fisher \u7528 boundary-weighted BCE + Dice \u8ba1\u7b97 Fisher\uff0c\u4f7f\u526a\u679d\u6253\u5206\u504f\u5411\u4fdd\u7559\u8fb9\u754c\u654f\u611f\u53c2\u6570\u3002Cross-Fisher \u5219\u7528 SAM \u4e0e MedSAM Fisher \u7684\u51e0\u4f55\u5e73\u5747\u6765\u8861\u91cf reset-to-SAM \u7684\u4ee3\u4ef7\uff0c\u907f\u514d\u88ab\u5927\u4f46\u65e0\u529f\u80fd\u610f\u4e49\u7684 weight shift \u4e3b\u5bfc\u3002\u5bf9\u8fb9\u754c\u654f\u611f\u533b\u5b66\u4efb\u52a1\u5f88\u53ef\u590d\u7528\u3002<\/p>\n<\/li>\n<li>\n<p><strong>Distribution-aware aggregation<\/strong>\uff1a\u5bf9\u4e0d\u540c\u6570\u636e\u96c6 \/ \u6a21\u6001\u8ba1\u7b97 score\uff0c\u5e76\u52a0\u5165\u8de8\u57df\u65b9\u5dee\u9879\u3002\u82e5\u67d0\u4e2a group \u53ea\u5728 BUSI \u6216 ISIC \u7b49\u7279\u5b9a\u5206\u5e03\u4e0a\u91cd\u8981\uff0c\u65b9\u5dee\u9879\u4f1a\u4f7f\u526a\u679d\u66f4\u4fdd\u5b88\u3002\u8fd9\u5bf9\u8de8\u533b\u9662\u3001\u591a\u6a21\u6001\u533b\u5b66\u6570\u636e\u5f88\u91cd\u8981\u3002<\/p>\n<\/li>\n<li>\n<p><strong>Head-to-MLP cascade pruning<\/strong>\uff1a\u8bba\u6587\u53d1\u73b0 MedSAM \u5904\u4e8e <strong>head-fragile boundary regime<\/strong>\uff1aattention heads \u7684 boundary leverage \u660e\u663e\u9ad8\u4e8e MLP groups\u3002\u56e0\u6b64\u5b9e\u9645\u526a\u679d\u7b56\u7565\u5e94\u66f4\u4fdd\u5b88\u5730\u526a heads\uff0c\u66f4\u591a\u538b\u7f29\u6765\u81ea MLP\u3002\u8fd9\u4e2a\u7ecf\u9a8c\u5bf9\u538b\u7f29 ViT-based MedSAM \/ SAMed \/ Medical SAM Adapter \u90fd\u6709\u542f\u53d1\u3002<\/p>\n<\/li>\n<li>\n<p><strong>Recovery fine-tuning<\/strong>\uff1a\u526a\u679d\u540e\u7528 <code>Lrec = Lseg + \u03bb1Lbd + \u03bb2Lfeat + \u03bb3Llogit + \u03bb4Lfreq<\/code> \u8ba9\u5269\u4f59\u7ed3\u6784\u91cd\u65b0\u534f\u8c03\u3002\u9700\u8981\u6ce8\u610f\uff1a\u4e3b\u7ed3\u679c\u5305\u542b recovery fine-tuning\uff0c\u56e0\u6b64\u4e0d\u80fd\u628a\u63d0\u5347\u5b8c\u5168\u5f52\u56e0\u4e8e pruning score\uff1b\u5b83\u662f\u201c\u526a\u679d\u9009\u62e9 + \u6062\u590d\u8bad\u7ec3\u201d\u7684\u7ec4\u5408\u7cfb\u7edf\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u662f\u5426\u9002\u5408 polyp segmentation \/ 3D segmentation<\/strong>\uff1a\u5bf9 polyp segmentation \u975e\u5e38\u76f4\u63a5\uff0c\u56e0\u4e3a\u4e3b\u8868\u5c31\u662f CVC-ClinicDB \/ CVC-ColonDB \/ Kvasir-SEG macro average\uff0c\u5e76\u4e14\u8fb9\u754c\u8d28\u91cf\u662f\u606f\u8089\u4efb\u52a1\u7684\u6838\u5fc3\u3002\u5bf9 3D medical segmentation\uff0c\u601d\u60f3\u53ef\u8fc1\u79fb\uff0c\u4f46 MedSAM \u672c\u8eab\u662f 2D promptable model\uff1b\u82e5\u8fc1\u79fb\u5230 MedSAM2 \/ 3D SAM \/ SegMamba\uff0c\u9700\u8981\u91cd\u65b0\u5b9a\u4e49 3D boundary band\u3001surface distance\u30013D Fisher \u548c volume-level prompts\u3002<\/p>\n<\/li>\n<\/ul>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u5b9e\u9a8c\u4f7f\u7528\u5b98\u65b9 MedSAM ViT-B \u4f5c\u4e3a base model\uff0cSAM ViT-B \u4f5c\u4e3a reference checkpoint\u3002\u6570\u636e\u8986\u76d6\u4e94\u4e2a\u533b\u5b66\u5206\u5272\u6570\u636e\u96c6\uff1a<\/p>\n<ul>\n<li><strong>Polyp endoscopy<\/strong>\uff1aCVC-ClinicDB\u3001CVC-ColonDB\u3001Kvasir-SEG\uff1b\u4e3b\u8868\u62a5\u544a\u4e09\u8005 macro-average\u3002<\/li>\n<li><strong>Breast ultrasound<\/strong>\uff1aBUSI\u3002<\/li>\n<li><strong>Dermoscopy<\/strong>\uff1aISIC2018\u3002<\/li>\n<\/ul>\n<p>\u8bc4\u4ef7\u6307\u6807\u5305\u62ec Dice\u3001IoU\u3001Boundary F1\uff08BF1\uff09\u3001HD95\u3001\u53c2\u6570\u91cf\u548c FLOPs\u3002Fisher estimation \u6bcf\u4e2a\u6570\u636e\u96c6\u91c7\u6837 128 calibration images\uff0cbatch size 1\u3002<\/p>\n<p>\u5173\u952e\u7ed3\u679c\uff1a<\/p>\n<ul>\n<li>\u539f\u59cb MedSAM\uff1a<strong>89.7M \u53c2\u6570\u3001926.5G FLOPs\u3001Dice 0.9191\u3001IoU 0.8648\u3001BF1 0.5321\u3001HD95 21.29<\/strong>\u3002<\/li>\n<li>MedCore one-time <code>h50_m70<\/code>\uff1a<strong>35.8M \u53c2\u6570\u3001385.2G FLOPs\u3001Dice 0.9549\u3001IoU 0.9169\u3001BF1 0.6388\u3001HD95 5.14<\/strong>\u3002\u8bba\u6587\u79f0\u76f8\u5f53\u4e8e\u51cf\u5c11 <strong>60.0% \u53c2\u6570\u300158.4% FLOPs<\/strong>\u3002<\/li>\n<li>MedCore sequential <code>h84_m95<\/code>\uff1a<strong>12.1M \u53c2\u6570\u300190.4G FLOPs\u3001Dice 0.9550\u3001IoU 0.9174\u3001BF1 0.6462\u3001HD95 5.12<\/strong>\uff0c\u5c5e\u4e8e\u66f4\u6fc0\u8fdb\u538b\u7f29\u4f46\u4ecd\u4fdd\u6301\u5f3a\u8fb9\u754c\u6307\u6807\u7684\u914d\u7f6e\u3002<\/li>\n<li>\u5bf9\u6bd4 EfficientSAM-s\/t\u3001SlimSAM\u3001Swin-Unet\u3001nnU-Net\u3001EMCAD\u3001MK-UNet\u3001SAMed\u3001QMedSAM\uff0cMedCore \u5728\u4e09\u7c7b\u606f\u8089\u6570\u636e macro-average \u4e0a\u62a5\u544a\u4e86\u6700\u597d\u7684 Dice\/BF1\/HD95 \u7ec4\u5408\u3002<\/li>\n<li>\u8de8\u6a21\u6001\u5b9e\u9a8c\u4e2d\uff0cCVC-ClinicDB \u4e0e ISIC2018 \u4e0a boundary metrics \u660e\u663e\u6539\u5584\uff1bBUSI \u4e0a moderate compression \u63a5\u8fd1\u6216\u7565\u4f18\uff0c\u4f46 aggressive compression \u4f1a\u5bfc\u81f4 BF1\/HD95 \u4e0b\u964d\uff0c\u8bf4\u660e\u65b9\u6cd5\u4ecd\u9700\u76ee\u6807\u57df\u9a8c\u8bc1\u3002<\/li>\n<li>\u7ec4\u4ef6\u6d88\u878d\u663e\u793a\uff0c\u53bb\u6389 boundary Fisher \u5728 aggressive one-time setting <code>h70_m95<\/code> \u4e0a BF1 \u4ece 0.4050 \u964d\u5230 0.2781\uff0c\u5f71\u54cd\u6700\u5927\uff1b\u53bb\u6389 reset-to-SAM \u6216 variance aggregation \u5728\u9ad8\u538b\u7f29\u65f6\u4e5f\u66f4\u660e\u663e\u3002<\/li>\n<li>Head-vs-MLP \u5206\u6790\u663e\u793a\uff0cattention head pruning \u7684 median boundary leverage \u4e3a <strong>3.961<\/strong>\uff0cMLP pruning \u4e3a <strong>1.372<\/strong>\uff0c\u6bd4\u503c <strong>2.887<\/strong>\uff1bBF1 damage density \u548c HD95 damage density \u4e5f\u5206\u522b\u7ea6\u4e3a MLP \u7684 <strong>2.607\u00d7<\/strong> \u548c <strong>2.432\u00d7<\/strong>\u3002<\/li>\n<\/ul>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u4e4b\u5904\uff1a<\/p>\n<ul>\n<li>\u8bba\u6587\u6ca1\u6709\u53ea\u62a5\u544a Dice\uff0c\u800c\u662f\u628a BF1 \u548c HD95 \u653e\u5728\u6838\u5fc3\u4f4d\u7f6e\uff0c\u7b26\u5408\u533b\u5b66\u8fb9\u754c\u4fdd\u771f\u9700\u6c42\u3002<\/li>\n<li>\u4e3b\u5b9e\u9a8c\u76f4\u63a5\u5305\u542b polyp segmentation \u4e09\u4e2a\u5e38\u7528\u6570\u636e\u96c6\uff0c\u5bf9\u7528\u6237\u7684 polyp \u7814\u7a76\u975e\u5e38\u76f8\u5173\u3002<\/li>\n<li>\u6d88\u878d\u4e0d\u4ec5\u9a8c\u8bc1\u6a21\u5757\u6709\u65e0\uff0c\u8fd8\u628a boundary Fisher\u3001reset score\u3001distribution variance \u5206\u5f00\u770b\uff0c\u80fd\u652f\u6491\u4f5c\u8005\u7684\u673a\u5236\u89e3\u91ca\u3002<\/li>\n<li>head \/ MLP boundary leverage sweep \u4e0e\u6700\u7ec8 BF1\/HD95 damage density \u65b9\u5411\u4e00\u81f4\uff0c\u8bf4\u660e\u7406\u8bba\u6307\u6807\u548c\u5b9e\u9645\u8fb9\u754c\u6307\u6807\u6709\u4e00\u5b9a\u5bf9\u5e94\u5173\u7cfb\u3002<\/li>\n<li>PDF \u4e2d\u7ed9\u51fa\u5b98\u65b9 GitHub \u94fe\u63a5\uff0c\u590d\u73b0\u5e0c\u671b\u6bd4\u8bb8\u591a arXiv \u9884\u5370\u672c\u66f4\u597d\u3002<\/li>\n<\/ul>\n<p>\u9700\u8981\u8c28\u614e\u4e4b\u5904\uff1a<\/p>\n<ul>\n<li>\u4e3b\u7ed3\u679c\u5305\u542b post-pruning recovery fine-tuning\uff0c\u4e14\u539f\u59cb MedSAM \u4e0e\u538b\u7f29\u540e\u6a21\u578b\u7684\u63d0\u5347\u4e0d\u5e94\u88ab\u7b80\u5355\u7406\u89e3\u4e3a\u201c\u526a\u679d\u672c\u8eab\u8ba9\u6a21\u578b\u66f4\u5f3a\u201d\uff1b\u53ef\u80fd\u662f recovery pipeline\u3001\u8bad\u7ec3\u6570\u636e\u3001prompt \u8bbe\u7f6e\u5171\u540c\u4f5c\u7528\u3002<\/li>\n<li>\u6bcf\u4e2a\u6570\u636e\u96c6\u7684 train\/test split \u53ea\u8bf4 fixed random seed\uff0c\u5177\u4f53 split \u6587\u4ef6\u3001prompt \u751f\u6210\u7b56\u7565\u3001box prompt \u6765\u6e90\u3001\u662f\u5426\u4e0e baseline \u5b8c\u5168\u4e00\u81f4\uff0c\u9700\u8981\u770b\u4ee3\u7801\u786e\u8ba4\u3002<\/li>\n<li>\u4e0e nnU-Net\u3001EMCAD\u3001MK-UNet \u7b49\u4e13\u7528 medical segmentation model \u7684\u6bd4\u8f83\u53ef\u80fd\u5b58\u5728\u4efb\u52a1\u5b9a\u4e49\u5dee\u5f02\uff1aMedSAM \u662f prompt-driven\uff0cnnU-Net \u662f\u81ea\u52a8\u5206\u5272\uff1b\u82e5 prompt \u6765\u81ea ground-truth box \u6216\u5f3a oracle\uff0c\u4f1a\u6539\u53d8\u516c\u5e73\u6027\u3002<\/li>\n<li>FLOPs \u4ecd\u7136\u4e0d\u4f4e\uff1a<code>h50_m70<\/code> \u8fd8\u6709 385.2G\uff1b\u5373\u4f7f <code>h84_m95<\/code> \u4e3a 90.4G\uff0c\u4e5f\u672a\u62a5\u544a\u771f\u5b9e FPS\u3001\u663e\u5b58\u3001\u7aef\u4fa7\u5ef6\u8fdf\u3002<\/li>\n<li>\u6ca1\u6709 3D \u6570\u636e\u3001\u89c6\u9891 polyp \u6216\u8de8\u533b\u9662\u5916\u90e8\u5927\u89c4\u6a21\u9a8c\u8bc1\uff1b\u4e34\u5e8a\u90e8\u7f72\u4ecd\u9700 prospective validation\u3002<\/li>\n<\/ul>\n<p>\u56e0\u6b64\uff0cMedCore \u7684\u7ed3\u8bba\u5e94\u8868\u8ff0\u4e3a\uff1a\u5b83\u5f3a\u6709\u529b\u5730\u652f\u6301\u201cMedSAM \u538b\u7f29\u65f6\u5fc5\u987b\u663e\u5f0f\u4fdd\u62a4\u533b\u5b66\u9002\u914d\u7ed3\u6784\u548c\u8fb9\u754c\u9ad8\u6760\u6746\u7ed3\u6784\u201d\uff0c\u4f46\u662f\u5426\u80fd\u4f5c\u4e3a\u5b9e\u9645\u90e8\u7f72\u6a21\u578b\uff0c\u8fd8\u8981\u770b prompt \u534f\u8bae\u3001\u7aef\u4fa7\u901f\u5ea6\u548c\u72ec\u7acb\u590d\u73b0\u3002<\/p>\n<h4>8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n<ul>\n<li><strong>U-Net \/ nnU-Net<\/strong>\uff1aMedCore \u4e0d\u662f U-Net \u6539\u8fdb\uff0c\u800c\u662f MedSAM\/SAM \u538b\u7f29\u65b9\u6cd5\u3002\u5b83\u4e0e nnU-Net \u7684\u5173\u7cfb\u4e3b\u8981\u4f53\u73b0\u5728\u5b9e\u9a8c\u6bd4\u8f83\u548c\u90e8\u7f72\u5b9a\u4f4d\uff1annU-Net \u662f fully supervised automatic segmentation \u5f3a\u57fa\u7ebf\uff0cMedCore \u662f promptable foundation model \u7684\u8f7b\u91cf\u5316\u7248\u672c\u3002<\/li>\n<li><strong>MedNeXt \/ CNN segmentation<\/strong>\uff1a\u6ca1\u6709\u76f4\u63a5\u6bd4\u8f83 MedNeXt\uff1b\u4f46 boundary-aware recovery\u3001high-frequency discrepancy loss \u7b49\u601d\u60f3\u53ef\u8fc1\u79fb\u5230 MedNeXt \u538b\u7f29\u6216\u84b8\u998f\u3002<\/li>\n<li><strong>UNETR \/ Swin-UNETR \/ TransUNet<\/strong>\uff1aMedCore \u7684\u526a\u679d\u5bf9\u8c61\u662f ViT image encoder\uff0c\u56e0\u6b64\u4e0e\u8fd9\u4e9b Transformer-based segmentation \u7684\u7ed3\u6784\u5316 pruning \u6709\u65b9\u6cd5\u76f8\u901a\u6027\uff0c\u5c24\u5176\u662f attention head vs MLP group \u7684\u526a\u679d\u9884\u7b97\u5206\u914d\u3002<\/li>\n<li><strong>Mamba \/ VMamba \/ SegMamba \/ DAMamba<\/strong>\uff1a\u8bba\u6587\u6ca1\u6709 SSM\/Mamba \u6a21\u5757\uff0c\u4f46 boundary leverage \u7684\u601d\u60f3\u53ef\u7528\u4e8e DAMamba\uff1a\u5206\u6790\u4e0d\u540c scan direction\u3001state-space branch\u3001MLP branch \u5bf9 boundary logit perturbation \u7684\u5f71\u54cd\uff0c\u51b3\u5b9a\u526a\u679d\u6216\u8f7b\u91cf\u5316\u7b56\u7565\u3002<\/li>\n<li><strong>Foundation model for medical segmentation<\/strong>\uff1a\u8fd9\u662f\u8bba\u6587\u6700\u76f4\u63a5\u7684\u4f4d\u7f6e\u3002\u5b83\u5ef6\u7eed MedSAM \/ SAMed \/ Medical SAM Adapter \/ EfficientSAM \/ QMedSAM \u65b9\u5411\uff0c\u4f46\u8d21\u732e\u70b9\u4e0d\u662f\u9002\u914d\u65b0\u4efb\u52a1\uff0c\u800c\u662f\u201c\u5982\u4f55\u5728\u538b\u7f29\u5df2\u9002\u914d\u6a21\u578b\u65f6\u4e0d\u7834\u574f\u533b\u5b66\u6838\u5fc3\u548c\u8fb9\u754c\u201d\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5bf9\u7528\u6237\u7684 polyp segmentation\u3001DAMamba \u6539\u9020\u548c\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u9009\u62e9\uff0cMedCore \u4ef7\u503c\u8f83\u9ad8\uff1a<\/p>\n<ul>\n<li><strong>polyp segmentation<\/strong>\uff1a\u5f3a\u76f8\u5173\u3002\u4e3b\u5b9e\u9a8c\u5c31\u662f CVC-ClinicDB\u3001CVC-ColonDB\u3001Kvasir-SEG\uff1b\u4e14\u5173\u6ce8 BF1\/HD95\uff0c\u6b63\u597d\u5bf9\u5e94\u606f\u8089\u8fb9\u754c\u5f31\u3001\u5f62\u6001\u53d8\u5316\u5927\u3001\u4e34\u5e8a\u5c3a\u5bf8\u4f30\u8ba1\u654f\u611f\u7684\u95ee\u9898\u3002<\/li>\n<li><strong>DAMamba \u6539\u9020<\/strong>\uff1a\u53ef\u501f\u9274 boundary leverage \u5206\u6790\u65b9\u6cd5\u3002\u6bd4\u5982\u5728 DAMamba \u4e2d\u6bd4\u8f83\u4e0d\u540c\u6a21\u5757\uff08DAMamba block\u3001CNN branch\u3001decoder fusion\u3001MLP\uff09\u5bf9 boundary band logit perturbation \u7684\u5f71\u54cd\uff0c\u7528\u5b83\u6307\u5bfc\u8f7b\u91cf\u5316\u3001\u526a\u679d\u6216\u6a21\u5757\u4fdd\u7559\u3002<\/li>\n<li><strong>MedSAM \/ SAM baseline<\/strong>\uff1a\u5982\u679c\u7528\u6237\u540e\u7eed\u505a polyp segmentation foundation model \u5bf9\u6bd4\uff0cMedCore \u53ef\u4f5c\u4e3a\u201ccompressed MedSAM + boundary-preserving\u201d\u76f8\u5173\u5de5\u4f5c\u3002<\/li>\n<li><strong>related work<\/strong>\uff1a\u9002\u5408\u653e\u5728 foundation model compression\u3001boundary-aware segmentation\u3001efficient medical segmentation \u4e09\u4e2a\u5c0f\u8282\u3002<\/li>\n<li><strong>\u590d\u73b0\u5efa\u8bae<\/strong>\uff1a\u82e5\u4ee3\u7801\u53ef\u8fd0\u884c\uff0c\u4f18\u5148\u590d\u73b0 <code>h50_m70<\/code> \u5728 Kvasir-SEG \/ ClinicDB \u4e0a\u7684 Dice\u3001BF1\u3001HD95\uff0c\u800c\u4e0d\u662f\u4e00\u5f00\u59cb\u8ffd\u6c42\u6700\u6fc0\u8fdb <code>h84_m95<\/code>\u3002<\/li>\n<\/ul>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5f3a\u70c8\u5efa\u8bae\u7cbe\u8bfb\u3002<\/strong> \u5bf9 polyp segmentation \u548c MedSAM \u8f7b\u91cf\u5316\u975e\u5e38\u76f8\u5173\uff0c\u4e14\u8bba\u6587\u7684 boundary leverage \u89e3\u91ca\u6bd4\u4e00\u822c\u538b\u7f29\u8bba\u6587\u66f4\u6709\u673a\u5236\u4ef7\u503c\u3002\u5efa\u8bae\u91cd\u70b9\u8bfb Section 3 \u7684 dual intervention\u3001boundary-aware Fisher \/ Cross-Fisher\u3001boundary leverage theorem\uff0c\u4ee5\u53ca Table 1\u3001Table 3\u3001Table 4\u3002\u8bfb\u5b9e\u9a8c\u65f6\u8981\u7279\u522b\u533a\u5206\u201c\u526a\u679d\u6253\u5206\u8d21\u732e\u201d\u548c\u201crecovery fine-tuning \u8d21\u732e\u201d\uff0c\u5e76\u68c0\u67e5\u4ee3\u7801\u4e2d\u7684 prompt\u3001split \u548c calibration \u8bbe\u7f6e\u3002<\/p>\n<hr \/>\n<h2>\u8bba\u6587 2\uff1aFrequency Adapter with SAM for Generalized Medical Image Segmentation\uff08\u5019\u9009\u4f46\u4eca\u65e5\u4e0d\u4f5c\u4e3a\u6b63\u5f0f\u5165\u9009\u7cbe\u8bfb\uff09<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aFrequency Adapter with SAM for Generalized Medical Image Segmentation<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aPhuoc-Nguyen Bui, Van-Nguyen Pham, Duc-Tai Le, Junghyun Bum, Hyunseung Choo \/ \u7b2c\u4e00\u4f5c\u8005 Phuoc-Nguyen Bui<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-11<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint, arXiv:2605.09925v1<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.09925<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.09925v1<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u8bba\u6587\u79f0 code and pre-trained models will be made available on GitHub\uff1b\u672a\u83b7\u53d6\u5230\u660e\u786e\u53ef\u8bbf\u95ee\u5b98\u65b9\u4ed3\u5e93<\/li>\n<li>\u4efb\u52a1\uff1asingle-source domain generalization for medical image segmentation\uff1bfundus optic disc\/cup segmentation\uff1bprostate MRI segmentation<\/li>\n<li>\u6570\u636e\u96c6\uff1aRIGA+ fundus multi-domain dataset\uff1bmulti-site prostate MRI dataset\uff08A\/RUNMC, B\/BMC, C\/I2CVB, D\/UCL, E\/BIDMC, F\/HK\uff09<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1aSAM adaptation\uff1bLoRA\uff1bfrequency adapter\uff1bdomain generalization\uff1bfoundation model for medical segmentation<\/li>\n<\/ul>\n<h3>paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n<h4>1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n<p>FSAM \u6709\u4e00\u4e2a\u6e05\u695a\u4f46\u76f8\u5bf9\u8f7b\u91cf\u7684\u60f3\u6cd5\uff1a\u5728 SAM + LoRA \u533b\u5b66\u9002\u914d\u4e2d\u52a0\u5165\u9891\u57df adapter\uff0c\u8ba9\u6a21\u578b\u5728\u5355\u6e90\u8bad\u7ec3\u3001\u8de8\u57df\u6d4b\u8bd5\u65f6\u66f4\u5173\u6ce8\u5bf9 scanner \/ protocol \/ texture shift \u66f4\u7a33\u5065\u7684\u9891\u7387\u7ed3\u6784\uff1b\u4f46\u5b83\u4e0e\u7528\u6237\u7684 polyp \/ DAMamba \u4e3b\u7ebf\u5173\u8054\u5f31\u4e8e MedCore\uff0c\u4e5f\u4e0d\u662f 5 \u6708 13 \u65e5\u4e4b\u540e\u7684\u65b0\u5019\u9009\uff0c\u6240\u4ee5\u4eca\u5929\u53ea\u4f5c\u4e3a\u5907\u9009\u77ed\u8bc4\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76 single-source domain generalization\uff1a\u53ea\u7528\u4e00\u4e2a\u6709\u6807\u6ce8\u6e90\u57df\u8bad\u7ec3\uff0c\u6d4b\u8bd5\u65f6\u76f4\u63a5\u9762\u5bf9\u672a\u89c1\u8fc7\u7684\u76ee\u6807\u57df\u3002\u533b\u5b66\u56fe\u50cf\u8de8\u57df\u5dee\u5f02\u6765\u81ea\u6210\u50cf\u8bbe\u5907\u3001\u626b\u63cf\u534f\u8bae\u3001\u533b\u9662\u4eba\u7fa4\u3001\u5206\u8fa8\u7387\u3001\u5bf9\u6bd4\u5ea6\u548c\u7eb9\u7406\u3002\u4f20\u7edf U-Net \u6216\u666e\u901a SAM adaptation \u53ef\u80fd\u5728\u6e90\u57df\u8868\u73b0\u597d\uff0c\u4f46\u8de8\u57df Dice \u4e0b\u964d\u660e\u663e\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u8ba4\u4e3a\u4f20\u7edf DG \u65b9\u6cd5\u4f9d\u8d56\u56e0\u679c\u7279\u5f81\u3001\u5bf9\u6297\u4e00\u81f4\u6027\u3001style augmentation \u6216 handcrafted augmentation\uff0c\u672a\u5145\u5206\u5229\u7528 SAM \u8fd9\u7c7b foundation model\uff1b\u5df2\u6709 SAM-based DG \u65b9\u6cd5\u5982 DeSAM\u3001DAPSAM \u4e3b\u8981\u5728\u7a7a\u95f4\u57df\u505a\u7279\u5f81\u89e3\u8026\u6216 prototype prompt\uff0c\u5ffd\u7565\u9891\u7387\u57df\u5dee\u5f02\uff0c\u800c\u9891\u7387\u6210\u5206\u5f80\u5f80\u5bf9\u5e94\u533b\u5b66\u56fe\u50cf\u4e2d\u7684\u7eb9\u7406\u3001\u8fb9\u754c\u3001\u626b\u63cf\u566a\u58f0\u548c\u5206\u8fa8\u7387\u5dee\u5f02\u3002<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>FSAM \u628a\u8f93\u5165\u56fe\u50cf\u505a FFT\uff0c\u4f7f\u7528 amplitude component \u751f\u6210 frequency input\uff1b\u539f\u56fe\u548c\u9891\u57df\u8f93\u5165\u90fd\u7ecf\u8fc7 SAM ViT-B patch embedding\u3002SAM encoder \u4e3b\u4f53\u51bb\u7ed3\u6216\u90e8\u5206\u901a\u8fc7 LoRA \u9002\u914d\uff0cLoRA rank \u8bbe\u4e3a 4\uff1b\u6bcf\u4e2a SAM ViT block \u5bf9\u5e94\u4e00\u4e2a <strong>Frequency Adapter<\/strong>\uff0c\u7531 down projection\u3001GELU\u3001up projection \u7ec4\u6210\u3002\u6700\u540e encoder embedding \u8fdb\u5165 prototype-based automated prompt generator \u548c SAM decoder\uff0c\u8bad\u7ec3\u635f\u5931\u4e3a CE + Dice\u3002<\/p>\n<p>\u8def\u7ebf\u8bb0\u5f55\uff1aPrimary adapter = method-algorithm\uff1bSecondary adapter = \u65e0\uff1bEvidence packs = general\u3001experimental-eval\u3001robustness-and-ood\uff1bRoute confidence = \u4e2d\u3002\u9009\u62e9\u8be5\u8def\u7ebf\u662f\u56e0\u4e3a\u8bba\u6587\u65b9\u6cd5\u7b80\u5355\uff0c\u4e3b\u8981\u8bc1\u636e\u662f\u8de8\u57df\u6cdb\u5316\u8868\u683c\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<ul>\n<li><strong>Frequency input \/ FFT amplitude<\/strong>\uff1a\u628a\u7a7a\u95f4\u56fe\u50cf\u53d8\u6362\u5230\u9891\u57df\uff0c\u53d6 amplitude \u8868\u793a\uff0c\u76ee\u6807\u662f\u6355\u6349\u8de8\u57df\u66f4\u7a33\u5065\u7684\u7ed3\u6784\u4e0e\u7eb9\u7406\u7edf\u8ba1\u3002<\/li>\n<li><strong>Frequency Adapter<\/strong>\uff1a\u6bcf\u4e2a ViT block \u4e2d\u6ce8\u5165\u4e00\u4e2a\u8f7b\u91cf bottleneck adapter\uff0c\u7ed3\u6784\u4e3a linear down-projection \u2192 GELU \u2192 linear up-projection\uff0c\u7528\u4e8e\u628a frequency embedding \u878d\u5165 SAM encoder\u3002<\/li>\n<li><strong>LoRA on SAM attention<\/strong>\uff1a\u53ea\u66f4\u65b0\u4f4e\u79e9\u77e9\u9635\uff0c\u51cf\u5c11\u8bad\u7ec3\u53c2\u6570\uff0c\u4fdd\u7559 SAM \u9884\u8bad\u7ec3\u77e5\u8bc6\u3002<\/li>\n<li><strong>Automated prompt generator<\/strong>\uff1a\u6cbf\u7528 DAPSAM \u7c7b memory-based prototype prompt\uff0c\u4e0d\u4f9d\u8d56\u4eba\u5de5 box\/point prompt\uff0c\u751f\u6210 domain-adaptive prompt \u7ed9 SAM decoder\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u6a21\u5757\u5bf9 polyp segmentation \u6709\u4e00\u5b9a\u542f\u53d1\uff1a\u53ef\u628a\u9891\u57df adapter \u7528\u4e8e\u8de8\u4e2d\u5fc3\u606f\u8089\u6570\u636e\u6cdb\u5316\uff1b\u4f46\u8bba\u6587\u6ca1\u6709\u606f\u8089\u5b9e\u9a8c\uff0c\u4e5f\u6ca1\u6709 Mamba \/ DAMamba \u76f8\u5173\u7ed3\u6784\u3002<\/p>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<ul>\n<li><strong>RIGA+ fundus<\/strong>\uff1aBinRushed \/ Magrabia \u4f5c\u6e90\u57df\uff0cMessidor Base 1\/2\/3 \u4f5c\u76ee\u6807\u57df\uff0c\u8bc4\u4ef7 optic disc \/ optic cup DSC\u3002Magrabia \u4f5c\u6e90\u57df\u65f6\uff0cFSAM average OD\/OC \u4e3a <strong>96.39 \/ 88.63<\/strong>\uff0c\u9ad8\u4e8e CCSDG \u7684 <strong>94.98 \/ 85.53<\/strong>\uff0c\u4e5f\u7565\u9ad8\u4e8e DAPSAM \u7684 <strong>96.30 \/ 88.15<\/strong>\u3002BinRushed \u4f5c\u6e90\u57df\u65f6\uff0cFSAM average OD\/OC \u4e3a <strong>96.25 \/ 87.53<\/strong>\uff0c\u63a5\u8fd1\u4f46\u7565\u4f4e\u4e8e DAPSAM \u7684 <strong>96.26 \/ 87.77<\/strong>\u3002<\/li>\n<li><strong>Prostate multi-site MRI<\/strong>\uff1a\u516d\u57df leave-one-domain style \u8bc4\u4f30\uff0cFSAM average DSC <strong>82.74<\/strong>\uff0c\u9ad8\u4e8e DAPSAM <strong>81.31<\/strong>\u3001SAMed <strong>78.51<\/strong>\u3001CSDG <strong>70.06<\/strong>\u3001MaxStyle <strong>68.27<\/strong>\u3002<\/li>\n<li><strong>\u8bad\u7ec3\u8bbe\u7f6e<\/strong>\uff1aSAM ViT-B\uff0cLoRA rank 4\uff0cAdamW lr 5e-4\uff0cweight decay 0.1\uff0cearly stop at 160 epochs \/ max 200 epochs\uff0cDice + CE loss\uff0c\u56fe\u50cf resize \u5230 RIGA 512\u00d7512\u3001prostate 384\u00d7384\u3002<\/li>\n<\/ul>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u4f18\u70b9\u662f\uff1a\u4efb\u52a1\u5b9a\u4e49\u6e05\u695a\uff0c\u4e24\u4e2a\u5178\u578b DG benchmark\uff0c\u548c\u4f20\u7edf DG + SAM-based DG \u90fd\u6709\u6bd4\u8f83\u3002\u7f3a\u70b9\u662f\uff1a\u6ca1\u6709\u4ee3\u7801\u4ed3\u5e93\u786e\u8ba4\uff1b\u53ea\u62a5\u544a DSC\uff0c\u7f3a\u5c11 HD95\/BF1\/ASD \u7b49\u8fb9\u754c\u6307\u6807\uff1b\u6ca1\u6709\u5145\u5206\u6d88\u878d LoRA-only vs frequency adapter-only vs prompt generator-only\uff1bRIGA \u4e2d BinRushed \u6e90\u57df\u4e0b FSAM \u5e76\u4e0d\u4f18\u4e8e DAPSAM\uff1b\u6ca1\u6709 polyp\u3001CT\u30013D \u6216\u89c6\u9891\u4efb\u52a1\u9a8c\u8bc1\u3002\u56e0\u6b64\u5b83\u662f\u4e00\u4e2a\u53ef\u5173\u6ce8\u7684 SAM-DG \u5c0f\u8bba\u6587\uff0c\u4f46\u4e0d\u662f\u4eca\u5929\u6700\u503c\u5f97\u6b63\u5f0f\u7cbe\u8bfb\u7684\u7b2c\u4e8c\u7bc7\u3002<\/p>\n<h4>8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n<ul>\n<li><strong>U-Net \/ nnU-Net<\/strong>\uff1a\u4e0d\u662f U-Net \u6539\u8fdb\uff0c\u800c\u662f SAM adaptation\uff1b\u4f20\u7edf DG baselines \u591a\u4e3a U-Net \u7cfb\u3002<\/li>\n<li><strong>MedNeXt \/ UNETR \/ Swin-UNETR<\/strong>\uff1a\u65e0\u76f4\u63a5\u7ed3\u6784\u5173\u7cfb\u3002<\/li>\n<li><strong>Mamba \/ DAMamba<\/strong>\uff1a\u65e0\u76f4\u63a5\u5173\u7cfb\uff1b\u9891\u57df adapter \u601d\u60f3\u53ef\u4e0e DAMamba \u7684 encoder \/ decoder \u7279\u5f81\u589e\u5f3a\u7ed3\u5408\uff0c\u4f46\u9700\u8981\u91cd\u65b0\u8bbe\u8ba1\u3002<\/li>\n<li><strong>Foundation model segmentation<\/strong>\uff1a\u5c5e\u4e8e SAM \/ SAMed \/ MedSAM \/ DAPSAM \u7684 domain generalization \u5206\u652f\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5bf9\u7528\u6237\u5f53\u524d polyp \/ DAMamba \u65b9\u5411\u4ef7\u503c\u4e2d\u7b49\u504f\u4f4e\uff1a\u5b83\u6ca1\u6709\u606f\u8089\u5b9e\u9a8c\uff0c\u4e5f\u4e0d\u63d0\u4f9b\u65b0 backbone\uff1b\u4f46\u5982\u679c\u540e\u7eed\u5173\u6ce8\u8de8\u4e2d\u5fc3 polyp generalization\uff0c\u53ef\u4ee5\u501f\u9274\u201cfrequency adapter + LoRA + automatic prompt\u201d\u7684\u8bbe\u5b9a\uff0c\u5728 Kvasir-SEG \u2192 ClinicDB \/ ColonDB \/ ETIS \u6216\u53cd\u5411\u8de8\u57df\u4e0a\u9a8c\u8bc1\u3002<\/p>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u53ef\u7565\u8bfb\u3002<\/strong> \u5efa\u8bae\u53ea\u8bfb\u65b9\u6cd5\u56fe\u3001Frequency Adapter \u516c\u5f0f\u548c\u4e24\u4e2a\u5b9e\u9a8c\u8868\u3002\u82e5\u505a SAM-based domain generalization\uff0c\u53ef\u7ee7\u7eed\u8ddf\u8e2a\uff1b\u82e5\u5f53\u524d\u91cd\u70b9\u662f DAMamba\u3001polyp segmentation backbone \u6216 3D segmentation\uff0c\u5219\u4f18\u5148\u7ea7\u4f4e\u4e8e MedCore\u3001FEFormer\u3001USEMA\u3002<\/p>\n<hr \/>\n<h2>\u4eca\u65e5\u63a8\u8350\u4f18\u5148\u7ea7<\/h2>\n<ol>\n<li>\n<p><strong>MedCore: Boundary-Preserving Medical Core Pruning for MedSAM<\/strong><br \/>\n   \u4eca\u5929\u6700\u503c\u5f97\u8bfb\u3002\u5b83\u76f4\u63a5\u8fde\u63a5 MedSAM\u3001polyp segmentation\u3001\u8fb9\u754c\u8d28\u91cf\u3001\u6a21\u578b\u538b\u7f29\u4e0e\u90e8\u7f72\uff0c\u4e14\u6709\u7406\u8bba\u89e3\u91ca\u3001\u6d88\u878d\u548c\u5b98\u65b9\u4ee3\u7801\u94fe\u63a5\u3002<\/p>\n<\/li>\n<li>\n<p><strong>Frequency Adapter with SAM for Generalized Medical Image Segmentation<\/strong><br \/>\n   \u4f5c\u4e3a\u5907\u9009\u8ddf\u8e2a\uff0c\u4e0d\u4f5c\u4e3a\u4eca\u65e5\u6b63\u5f0f\u7b2c\u4e8c\u7bc7\u63a8\u8350\u3002\u5b83\u5bf9 SAM \u8de8\u57df\u6cdb\u5316\u6709\u53c2\u8003\u4ef7\u503c\uff0c\u4f46\u7f3a\u5c11 polyp \/ Mamba \/ 3D \u76f8\u5173\u5b9e\u9a8c\uff0c\u4e14\u4ee3\u7801\u5c1a\u672a\u786e\u8ba4\u3002<\/p>\n<\/li>\n<\/ol>\n<h2>\u4eca\u65e5 PDF \u83b7\u53d6\u60c5\u51b5<\/h2>\n<ul>\n<li>\u8bba\u6587 1\uff1a\u5df2\u9644 PDF \/ \u63d0\u4f9b PDF \u94fe\u63a5\uff1aMEDIA:\/tmp\/medseg_daily_20260515\/medcore_2605.13688.pdf\uff1bhttps:\/\/arxiv.org\/pdf\/2605.13688v1<\/li>\n<li>\u8bba\u6587 2\uff1a\u63d0\u4f9b PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.09925v1\uff08\u5df2\u4e0b\u8f7d\u5230\u672c\u5730\u7528\u4e8e\u9605\u8bfb\uff0c\u4f46\u56e0\u4e0d\u4f5c\u4e3a\u6b63\u5f0f\u5165\u9009\u7cbe\u8bfb\uff0c\u4e0d\u9644\u9644\u4ef6\uff09<\/li>\n<\/ul>\n<h2>\u4eca\u65e5\u53ef\u6267\u884c\u5efa\u8bae<\/h2>\n<ol>\n<li><strong>\u4f18\u5148\u7cbe\u8bfb\u5e76\u590d\u73b0 MedCore \u7684 boundary leverage \u4e0e h50_m70 \u914d\u7f6e\u3002<\/strong> \u5148\u5728 Kvasir-SEG \/ ClinicDB \u4e0a\u6838\u67e5 Dice\u3001BF1\u3001HD95 \u4e0e FLOPs\uff0c\u518d\u8003\u8651\u66f4\u6fc0\u8fdb\u7684 h84_m95\u3002<\/li>\n<li><strong>\u628a MedCore \u7684 boundary-aware pruning \u601d\u60f3\u8fc1\u79fb\u5230 DAMamba\u3002<\/strong> \u53ef\u4ee5\u8bbe\u8ba1\u4e00\u4e2a\u5c0f\u5b9e\u9a8c\uff1a\u5bf9 DAMamba \u4e0d\u540c\u6a21\u5757\u505a boundary band logit perturbation \u5206\u6790\uff0c\u6bd4\u8f83 CNN branch\u3001Mamba branch\u3001decoder fusion \u5bf9\u606f\u8089\u8fb9\u754c\u7684\u5f71\u54cd\u3002<\/li>\n<li><strong>related work \u4e2d\u65b0\u589e\u4e00\u4e2a\u5c0f\u8282\uff1aBoundary-preserving compression of medical foundation models\u3002<\/strong> MedCore \u53ef\u4f5c\u4e3a\u6838\u5fc3\u5f15\u7528\uff1bFSAM \u53ef\u653e\u5728 SAM domain generalization \/ frequency adapter \u7684\u8865\u5145\u5f15\u7528\u4e2d\uff0c\u4f46\u4e0d\u5efa\u8bae\u4f5c\u4e3a\u5f3a baseline\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\u5728 2026-05-13 \u4e4b\u540e\u7684\u65b0\u589e\u5019\u9009\u4e2d\uff0c\u771f\u6b63\u7b26\u5408\u201c\u533b\u5b66\u56fe\u50cf\u5206\u5272\u4e3b\u4efb\u52a1\u30012025 &#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-1062","post","type-post","status-publish","format-standard","hentry","category-85"],"views":6,"_links":{"self":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1062","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=1062"}],"version-history":[{"count":0,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1062\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/media?parent=1062"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/categories?post=1062"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/tags?post=1062"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}