文献速递皮肤肿瘤

人工智能辅助黑色素瘤诊断的前瞻性证据:系统综述与荟萃分析

Prospective Evidence on Artificial Intelligence-Assisted Melanoma Diagnostics: A Systematic Review and Meta-Analysis.
PubMed 文献 · JAMA Dermatol · 2026-01-01

皮肤科医生合并敏感度78.6%(95%CI 67.5%-88.1%)、特异度75.2%(95%CI 63.3%-84.3%),AI单独敏感度80.9%(95%CI 63.6%-94.5%)、特异度75.6%(95%CI 64.5%-85.6%);唯一报告AI辅助皮肤科医生的研究敏感度91.9%、特异度83.7%。作者认为AI诊断性能与皮肤科医生相当,作为决策支持工具可能提升性能,但研究偏倚风险高、泛化性有限,需在未选择人群的临床环境中进一步验证。

要点

  • 前瞻性证据显示AI单独诊断黑色素瘤的敏感度与特异度与皮肤科医生相近。
  • 唯一一项AI辅助皮肤科医生的研究报告敏感度91.9%、特异度83.7%。
  • 多数研究在患者选择与指数检验领域存在高偏倚风险,泛化性有限。

核心结论

  • 在前瞻性研究中,AI系统对黑色素瘤的诊断性能与皮肤科医生相当,作为决策支持工具可能提升诊断性能。
  • 当前研究偏倚风险高且泛化性有限,需在临床未选择人群中开展更广泛验证。

研究设计与样本

  • 系统综述与荟萃分析,检索PubMed、Embase、Web of Science和Google Scholar,检索时间自建库至2025年7月9日。
  • 纳入前瞻性、使用皮肤镜图像、以组织病理学为标准、报告或可计算皮肤科医生/AI/AI辅助皮肤科医生诊断性能指标的研究。
  • 排除非皮肤镜对照与回顾性设计;组织病理学确诊黑色素瘤≤20例的研究被预先排除于定量合成之外。

主要发现

  • 皮肤科医生合并敏感度78.6%(95%CI 67.5%-88.1%)、特异度75.2%(95%CI 63.3%-84.3%)。
  • AI单独合并敏感度80.9%(95%CI 63.6%-94.5%)、特异度75.6%(95%CI 64.5%-85.6%)。
  • 唯一报告AI辅助皮肤科医生的研究敏感度91.9%、特异度83.7%。
  • 直接临床比较中,AI显示更高特异度与相似敏感度。
  • 多数研究在患者选择与指数检验领域为高偏倚风险,主要因预先选择疑似黑色素瘤病灶及二分类判定。

关键数值与推荐

  • 纳入标准:前瞻性设计、皮肤镜图像、组织病理学参考标准、报告或可计算目标诊断性能指标。
  • 排除标准:非皮肤镜对照、回顾性设计、组织病理学确诊黑色素瘤≤20例的研究不纳入定量合成。
  • 偏倚风险评估工具:QUADAS-2与QUADAS-C。

临床怎么用

  • AI可作为黑色素瘤诊断的决策支持工具,与皮肤科医生联合使用可能提高诊断性能。
  • 解读现有证据时需考虑研究偏倚风险高与泛化性有限,不宜直接外推至未选择人群。

局限性

  • 多数研究在患者选择与指数检验领域存在高偏倚风险。
  • 研究多预先选择疑似黑色素瘤病灶并采用二分类判定,限制泛化性。
  • 仅一项研究报告AI辅助皮肤科医生的诊断性能,相关证据有限。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

IMPORTANCE: Dermoscopy is a standard of care for melanoma diagnostics, and artificial intelligence (AI) systems are increasingly investigated as decision-support tools. Prospective evidence is essential to assess their performance compared to dermatologists. OBJECTIVES: To evaluate the diagnostic performance of dermatologists, AI systems, and dermatologists assisted by AI in prospective studies of melanoma detection, and to assess the readiness of AI for clinical use. DATA SOURCES: PubMed, Embase, Web of Science, and Google Scholar were searched from inception through July 9, 2025. STUDY SELECTION: Eligible studies were prospective, used dermoscopic images, and reported or allowed calculation of performance metrics for dermatologists, AI, or dermatologists assisted by AI against a histopathologic reference standard. Nondermoscopic comparators and retrospective designs were excluded. Studies with 20 or fewer histopathologically confirmed melanomas were excluded a priori from quantitative synthesis. DATA EXTRACTION AND SYNTHESIS: Two reviewers independently screened and extracted data and discrepancies or missing values were clarified among all authors. Risk of bias and applicability were assessed with QUADAS-2 and QUADAS-C. Study-level sensitivity and specificity were summarized and plotted; head-to-head comparisons were analyzed descriptively. MAIN OUTCOMES AND MEASURES: Diagnostic outcomes were sensitivity, specificity, accuracy, and balanced accuracy for melanoma detection. RESULTS: Eleven prospective studies with a total of more than 2500 patients and 50 participant-dermatologists were included in the analyses. Dermatologists achieved a pooled sensitivity of 78.6% (95% CI, 67.5%-88.1%) and specificity of 75.2% (95% CI, 63.3%-84.3%), whereas AI alone reached 80.9% (95% CI, 63.6%-94.5%) sensitivity and 75.6% (95% CI, 64.5%-85.6%) specificity. In the single study reporting AI-assisted dermatologists, sensitivity was 91.9% and specificity was 83.7%. In direct clinical comparisons, AI demonstrated higher specificity and similar sensitivity. Most studies were at high risk of bias in patient selection and index test domains, primarily due to the preselection of lesions suspected of melanoma and binary classifications. CONCLUSIONS AND RELEVANCE: In the systematic review and meta-analysis of prospective settings, AI systems perform at comparable levels to dermatologists for melanoma diagnostics and may enhance performance when used as a decision-support tool. However, the frequent risk of bias and limited generalizability of current studies highlight the need for broader validation in unselected patient populations in the clinical setting.

来源:JAMA Dermatol,查看原文。版权归原作者所有。

引用

  1. Laiouar-Pedari S, Kühn A, Wies C, et al. Prospective Evidence on Artificial Intelligence-Assisted Melanoma Diagnostics: A Systematic Review and Meta-Analysis. JAMA Dermatol. 2026. PMID: 41879756. doi:10.1001/jamadermatol.2026.0217

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。