基于智能手机的眼表恶性肿瘤主动自筛系统:非随机临床试验
该非随机临床试验在中国开展,先以12年多中心裂隙灯图像训练并验证深度学习模型,再优化为智能手机拍照并经移动应用部署。614名参与者完成居家自筛,智能手机系统区分恶性与良性病变的AUC为0.905,实际筛查中病理确诊20例恶性肿瘤、19例为新诊断。
要点
- 智能手机AI自筛区分眼表恶性与良性病变的AUC达0.905,接近裂隙灯模型。
- 614名居家自筛者中病理确诊20例恶性肿瘤,其中19例为新诊断且无需眼球摘除。
- 群体水平AUC 0.977、敏感度89.3%、特异度95.9%,高危者被优先转诊。
核心结论
- 整合智能手机成像、AI诊断与媒体触达,可为罕见眼表恶性肿瘤提供可扩展、可及且可能临床有效的人群筛查策略。
研究设计与样本
- 中国开展的非随机临床试验,2022年12月至2023年6月,先用12年多中心裂隙灯图像训练和验证深度学习模型,再优化适配智能手机摄影并通过广泛传播的移动应用部署。
- 参与者使用含实时AI拍照指导的CaptureTumor应用拍摄疑似病变,应用即时给出良恶性二分类与多分类风险分层,并将高危病例优先转诊。
- 主要结局为区分恶性与良性病变的AUC,次要结局包括敏感度、特异度和病理确诊恶性肿瘤数。
主要发现
- 电视、社交媒体和互联网医院的多媒体传播触达256 053人,614人通过应用完成居家自筛。
- 参与者中位(IQR)年龄46(11)岁(范围4-87岁),女性301人(49%)、男性313人(51%)。
- 图像质量优化后,智能手机CaptureTumor的AUC为0.905(95%CI,0.837-0.973),与裂隙灯模型AUC 0.945(95%CI,0.918-0.972)相当。
- 实际筛查中614名参与者病理确诊20例恶性肿瘤,20例中19例(95%)为新诊断,无病例需眼球摘除。
- 群体水平CaptureTumor的AUC为0.977(95%CI,0.964-0.990),敏感度89.3%(95%CI,86.7%-91.9%),特异度95.9%(95%CI,94.2%-97.6%)。
关键数值与推荐
- 智能手机CaptureTumor区分眼表恶性与良性病变AUC 0.905(95%CI,0.837-0.973)。
- 裂隙灯模型AUC 0.945(95%CI,0.918-0.972)。
- 群体水平AUC 0.977(95%CI,0.964-0.990),敏感度89.3%(95%CI,86.7%-91.9%),特异度95.9%(95%CI,94.2%-97.6%)。
- 614名自筛者中病理确诊20例恶性肿瘤,19例(95%)为新诊断,无病例需眼球摘除。
临床怎么用
- 可借助智能手机应用配合AI实时拍照指导,让公众居家拍摄疑似眼表病变并获得良恶性风险分层。
- 对应用判定的高危病例优先转诊,以实现眼表恶性肿瘤的早期发现和及时诊疗。
- 以媒体和互联网医院触达结合移动应用,构建闭环移动健康筛查模式,弥补罕见眼病早期发现的缺口。
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: Ocular surface malignancies pose risks to vision and survival yet are frequently misdiagnosed as benign lesions because of their subtle presentation and the lack of widely accessible screening tools, potentially resulting in treatment delays and the need for extensive surgical intervention. OBJECTIVE: To develop and validate a smartphone-based, media-facilitated artificial intelligence (AI) system for proactive self-screening of ocular surface malignancies in the general population. DESIGN, SETTING, AND PARTICIPANTS: A nonrandomized clinical trial was conducted across China from December 2022 to June 2023. A deep learning model was initially trained and validated using 12 years of multicenter slitlamp images. The system was then optimized for smartphone-based photography and deployed through a widely disseminated mobile application. Data analysis was performed from July 2023 to June 2024. INTERVENTION: Participants used the CaptureTumor standardized smartphone application, incorporating real-time AI-guided photography instructions, to capture images of suspected lesions. The application provided immediate binary (benign vs malignant) and multiclass risk stratification and triaged high-risk cases for expedited clinical referral. MAIN OUTCOMES AND MEASURES: The primary outcome was area under the receiver operating characteristic curve (AUC) for differentiating malignant from benign lesions. Secondary outcomes included sensitivity, specificity, and the number of histopathologically confirmed malignancies detected. RESULTS: Multimedia outreach via television, social media, and internet hospitals reached 256 053 individuals, with 614 completing at-home self-screening through the app. Of these participants, the median (IQR) age was 46 (11) years (range, 4-87 years); 301 (49%) were female and 313 (51%) male. After optimizing the image quality, the smartphone-based CaptureTumor achieved an AUC of 0.905 (95% CI, 0.837-0.973), comparable with the performance of the slitlamp-based model (AUC = 0.945; 95% CI, 0.918-0.972). During real-world screening, 20 malignancies were pathologically confirmed among the 614 participants, with 19 of 20 participants (95%) newly diagnosed, and no cases requiring enucleation. At the population level, CaptureTumor demonstrated an AUC of 0.977 (95% CI, 0.964-0.990), a sensitivity of 89.3% (95% CI, 86.7%-91.9%), and a specificity of 95.9% (95% CI, 94.2%-97.6%). CONCLUSIONS AND RELEVANCE: This trial found that the integration of smartphone-enabled imaging, AI-driven diagnostics, and targeted media outreach established a scalable, accessible, and potentially clinically effective strategy for population-level screening of rare ocular malignancies. This closed-loop mobile health model potentially addresses gaps in early detection and equitable care delivery for vision- and life-threatening rare diseases. TRIAL REGISTRATION: ClinicalTrials.gov Identifier: NCT05645341.
引用
- Wang R, Bi S, Lin D, et al. Smartphone-Based Proactive Self-Screening for Ocular Surface Malignancies: A Nonrandomized Clinical Trial. JAMA Ophthalmol. 2026. PMID: 42240998. doi:10.1001/jamaophthalmol.2026.1609
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。