人工智能系统诊断黑色素瘤:系统评价与Meta分析
该研究系统检索2006年1月至2024年5月Web of Science、Scopus和MEDLINE数据库,纳入39项原始研究的78个诊断测试数据,采用Meta分析评估人工智能算法基于皮肤镜图像诊断黑色素瘤的预测性能。汇总SROC模型AUC为0.96(95%CI: 0.94-0.98),灵敏度0.89(95%CI: 0.85-0.91),特异度0.92(95%CI: 0.90-0.94);亚组分析显示HYBRID模型汇总AUC达0.98(95%CI: 0.96-0.99)。
要点
- AI算法诊断黑色素瘤汇总AUC 0.96,灵敏度0.89,特异度0.92。
- HYBRID模型亚组汇总AUC达0.98,预测性能突出。
- AI算法可辅助医生诊断黑色素瘤,HYBRID模型前景可观。
核心结论
- 人工智能算法和机器学习方法应广泛用于医学以辅助医生诊断黑色素瘤。
- HYBRID模型算法预测疾病的能力有前景,其性能表现较高。
研究设计与样本
- 系统评价与Meta分析,遵循PRISMA清单。
- 检索Web of Science、Scopus和MEDLINE数据库,时间范围为2006年1月至2024年5月。
- 共评估2722篇文章,最终纳入39项原始研究的78个诊断测试数据。
主要发现
- 汇总SROC整体模型AUC为0.96(95%CI: 0.94-0.98)。
- 汇总灵敏度为0.89(95%CI: 0.85-0.91)。
- 汇总特异度为0.92(95%CI: 0.90-0.94)。
- 亚组分析中HYBRID模型汇总AUC为0.98(95%CI: 0.96-0.99)。
关键数值与推荐
- 汇总SROC整体模型AUC 0.96(95%CI: 0.94-0.98),灵敏度0.89(95%CI: 0.85-0.91),特异度0.92(95%CI: 0.90-0.94)。
- HYBRID模型亚组汇总AUC 0.98(95%CI: 0.96-0.99)。
临床怎么用
- 临床医生可参考本Meta分析结果,了解人工智能算法尤其是HYBRID模型在黑色素瘤诊断中的预测性能。
- AI算法和机器学习方法可辅助医生诊断黑色素瘤。
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
BACKGROUND: One of the most promising and rapidly advancing research areas in recent years is using dermoscopic images for automatic diagnosis with artificial intelligence and machine learning methods. OBJECTIVES: This study aimed to synthesize the existing studies for the clinical use of applications made with artificial intelligence methods and to summarize the predictive performance of deep learning and hybrid models-based algorithms in all these studies with a large-scale meta-analysis. METHOD: The literature review was conducted between January 2006 and May 2024, and meta-analysis data were created by scanning the Web of Science (WOS), Scopus and MEDLINE databases. This study followed the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) checklist. RESULTS: A total of 2722 articles were evaluated. Data from 78 diagnostic tests from 39 primary studies meeting the inclusion and exclusion criteria were assessed. The pooled SROC overall model AUC was 0.96 [95% CI: 0.94-0.98], sensitivity was 0.89 [95% CI: 0.85-0.91] and specificity was 0.92 [95% CI: 0.90-0.94]. In the subgroup analyses, the pooled AUC was 0.98 [95% CI: 0.96-0.99] for HYBRID models. CONCLUSIONS: Recent studies have suggested that artificial intelligence algorithms and machine learning methods should be used extensively in medicine to assist physicians, especially in diagnosing melanoma. The ability of HYBRID model algorithms to predict diseases is promising. In particular, the performance of HYBRID models was found to be high. This information can assist clinicians in interpreting the most appropriate algorithms for diagnosing melanoma.
引用
- Ertürk Zararsız G, Yerlitaş Taştan SI, Çelik Gürbulak E, et al. Diagnosis melanoma with artificial intelligence systems: A meta-analysis study and systematic review. J Eur Acad Dermatol Venereol. 2025. PMID: 40476369. doi:10.1111/jdv.20781
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。