[{"data":1,"prerenderedAt":220},["ShallowReactive",2],{"blog-\u002Fblog\u002Fzh-hans\u002Fhow-to-evaluate-skin":3,"blog-\u002Fblog\u002Fzh-hans\u002Fhow-to-evaluate-skin-surround":194,"blog-\u002Fblog\u002Fzh-hans\u002Fhow-to-evaluate-skin-translations":203},{"id":4,"title":5,"author":6,"body":9,"date":178,"description":179,"draft":180,"extension":181,"image":182,"meta":183,"navigation":185,"path":186,"seo":187,"stem":188,"tags":189,"__hash__":193},"blogZhHans\u002Fblog\u002Fzh-hans\u002Fhow-to-evaluate-skin.md","用 AI 评估皮肤意味着什么",{"name":7,"picture":8},"Soo Ick Cho","\u002Fimages\u002Fblog\u002Fauthor-sooickcho.jpg",{"type":10,"value":11,"toc":170},"minimark",[12,16,21,24,27,31,34,37,40,43,46,49,53,56,62,65,68,120,123],[13,14,15],"p",{},"上一篇文章里，我谈到对皮肤的“测量”比想象中更不稳定。这样一来，自然会有人想：“既然设备靠不住，那让受过训练的医生亲自用眼睛看、做判断，不就行了吗？”很长一段时间里，我自己也是这么相信的。然而在做 AI 研究的过程中，我不得不面对一个事实：就连“医生的判断”，也无法归结为一个干脆利落的正确答案。",[17,18,20],"h2",{"id":19},"我的第一篇论文是鉴别唇部皮肤癌","我的第一篇论文，是鉴别“唇部皮肤癌”",[13,22,23],{},"在大学附属医院时，我的第一个研究课题，是用 AI 鉴别发生在嘴唇上的一种罕见皮肤癌。当时，用 AI 来“诊断”疾病的研究才刚刚出现，这项研究因为连并不常见的皮肤癌类型也能区分出来而获得认可，得以发表在皮肤科领域的主要学术期刊（《英国皮肤病学杂志》，BJD）上。",[13,25,26],{},"诊断相对来说比较容易处理，因为像“是不是皮肤癌”这样的问题，被认为是有正确答案的。（严格说来，诊断其实也更接近于概率。）不过，这类诊断研究当时已经在迅速趋于饱和，于是我把目光转向了下一个问题——“严重程度评估”。",[17,28,30],{"id":29},"严重程度并没有唯一的正确答案","“严重程度”并没有唯一的正确答案",[13,32,33],{},"接下来，我开展了用 AI 评估痤疮和特应性皮炎严重程度的研究。两项研究都因其新颖性获得认可，发表在不错的学术期刊上，但真正在研究过程中撞上的那堵墙，却是另一回事。",[13,35,36],{},"要训练 AI，就需要“标注数据（ground truth）”。于是我们把同一张照片拿给多位皮肤科医生看，请他们评定严重程度，结果各人的答案出现了分歧。大多数情况下答案是相近、重叠的，完全南辕北辙的情况很少见（比如在 4 个等级中，有人评为 3 级，有人评为 2 级），但即便如此，“唯一的正确答案”依然不存在。",[13,38,39],{},"与“是不是皮肤癌”不同，严重程度是一个多位专家的判断会散开成一个分布的问题。因此它需要一种不同于诊断的方法，而我的研究正是要解决这一点。",[17,41,42],{"id":42},"这不只是皮肤科的问题",[13,44,45],{},"后来在医疗 AI 公司（Lunit）开发病理分析 AI 时，我又遇到了同样的墙。我们用病理科医生在组织图像上标注的答案来训练模型，可就连这些答案，也因医生而异。面对同一块组织，有人判断为癌前病变，有人判断为恶性病变。",[13,47,48],{},"那时我才明白：我在皮肤科遇到的问题并不是皮肤科独有的，而是贯穿整个医学的普遍问题。经历了皮肤科 AI 研究和病理 AI 研究之后，我从经验中学会了该如何用 AI 来处理这种“散开的正确答案”。",[17,50,52],{"id":51},"所以怎样评估才是关键","所以，“怎样评估”才是关键",[13,54,55],{},"关键不在于给 AI 钉死一个严格的正确答案，而在于让它学习真实医疗现场中存在的判断分布——也就是说，以“概率”的形式学习。不是把某个人的判断当作错误而删掉，而是把多位专家判断所构成的幅度，原封不动地作为学习的素材。我在首尔大学医院和 Lunit 都以这种方式开发过 AI 模型，并确认这种方法更适合严重程度评估。如今，我正把当时积累的经验运用在 INSKINVIEW 的 AI 模型开发中。",[13,57,58],{},[59,60,61],"strong",{},"参考文献",[13,63,64],{},"以下是正文中提到的研究。",[13,66,67],{},"皮肤科 AI 研究",[69,70,71,90,105],"ul",{},[72,73,74,75,79,80],"li",{},"唇部皮肤癌鉴别 — ",[76,77,78],"em",{},"British Journal of Dermatology",", 2020;182:1388-1394.\n",[69,81,82],{},[72,83,84],{},[85,86,87],"a",{"href":87,"rel":88},"https:\u002F\u002Facademic.oup.com\u002Fbjd\u002Farticle-abstract\u002F182\u002F6\u002F1388\u002F6697632",[89],"nofollow",[72,91,92,93,96,97],{},"痤疮严重程度评估 — ",[76,94,95],{},"American Journal of Clinical Dermatology",", 2023;24:649-659.\n",[69,98,99],{},[72,100,101],{},[85,102,103],{"href":103,"rel":104},"https:\u002F\u002Flink.springer.com\u002Farticle\u002F10.1007\u002Fs40257-023-00777-5",[89],[72,106,107,108,111,112],{},"特应性皮炎严重程度评估 — ",[76,109,110],{},"IEEE Journal of Biomedical and Health Informatics",", 2023;27:166-175.\n",[69,113,114],{},[72,115,116],{},[85,117,118],{"href":118,"rel":119},"https:\u002F\u002Fieeexplore.ieee.org\u002Fabstract\u002Fdocument\u002F9932872\u002F",[89],[13,121,122],{},"病理科 AI 研究",[69,124,125,140,155],{},[72,126,127,128,131,132],{},"非小细胞肺癌病理分析 AI（PD-L1 判读） — ",[76,129,130],{},"European Journal of Cancer",", 2022;170:17-26.\n",[69,133,134],{},[72,135,136],{},[85,137,138],{"href":138,"rel":139},"https:\u002F\u002Fwww.sciencedirect.com\u002Fscience\u002Farticle\u002Fabs\u002Fpii\u002FS0959804922002155",[89],[72,141,142,143,146,147],{},"乳腺癌病理分析 AI（肿瘤浸润淋巴细胞评估） — ",[76,144,145],{},"npj Breast Cancer",", 2023;9:71.\n",[69,148,149],{},[72,150,151],{},[85,152,153],{"href":153,"rel":154},"https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41523-023-00577-4",[89],[72,156,157,158,161,162],{},"乳腺癌病理分析 AI（HER2·ER·PR 判读） — ",[76,159,160],{},"Breast Cancer Research",", 2024;26:31.\n",[69,163,164],{},[72,165,166],{},[85,167,168],{"href":168,"rel":169},"https:\u002F\u002Flink.springer.com\u002Farticle\u002F10.1186\u002Fs13058-024-01784-y",[89],{"title":171,"searchDepth":172,"depth":172,"links":173},"",2,[174,175,176,177],{"id":19,"depth":172,"text":20},{"id":29,"depth":172,"text":30},{"id":42,"depth":172,"text":42},{"id":51,"depth":172,"text":52},"2026-07-19","皮肤的严重程度，并没有唯一的正确答案。这是一位皮肤科专科医生在大学附属医院和医疗 AI 公司一路走来所面对的问题——“医生的判断有多可信”——以及他用 AI 来求解这个问题的故事。",false,"md","\u002Fimages\u002Fblog\u002FChatGPT Image 2026년 7월 19일 오후 10_37_40.png",{"slug":184},"how-to-evaluate-skin",true,"\u002Fblog\u002Fzh-hans\u002Fhow-to-evaluate-skin",{"title":5,"description":179},"blog\u002Fzh-hans\u002Fhow-to-evaluate-skin",[190,191,192],"皮肤评估","医疗 AI","INSKINVIEW","ePNUN6CgCumwo75FTn6BqPOBJaeKniSQ4o38HW91xmk",[195,199],{"title":196,"path":197,"stem":198,"children":-1},"为什么要用智能手机做肌肤分析","\u002Fblog\u002Fzh-hans\u002Fwhy-smartphone","blog\u002Fzh-hans\u002Fwhy-smartphone",{"title":200,"path":201,"stem":202,"children":-1},"把皮肤变成数据来测量：一位皮肤科医生的记录","\u002Fblog\u002Fzh-hans\u002Fmeasuring-skin-as-data","blog\u002Fzh-hans\u002Fmeasuring-skin-as-data",[204,208,212,216],{"code":205,"name":206,"path":207},"ko","한국어","\u002Fblog\u002Fhow-to-evaluate-skin",{"code":209,"name":210,"path":211},"en","English","\u002Fblog\u002Fen\u002Fhow-to-evaluate-skin",{"code":213,"name":214,"path":215},"ja","日本語","\u002Fblog\u002Fja\u002Fhow-to-evaluate-skin",{"code":217,"name":218,"path":219},"zh-hant","繁體中文","\u002Fblog\u002Fzh-hant\u002Fhow-to-evaluate-skin",1791649183025]