整合人工智能驱动的超声和分子检测，以改善不确定性质甲状腺结节的诊断。
完整摘要 · 中文译文
背景：不确定甲状腺结节（Bethesda III/IV）难以管理：超声和细针穿刺抽吸（FNA）常产生非确定性结果，因此许多患者因最终证明为良性的病变接受诊断性腺叶切除术。AI增强超声与分子生物标志物平台的联合提供互补的诊断能力。然而，关于其联合应用的证据仍然有限，来源于少量早期研究，尚未在大型或前瞻性队列中得到验证。方法：对PubMed（2012-2025）的系统检索识别出评估常规超声+FNA、AI辅助超声、分子生物标志物平台（Afirma、ThyroSeq、ThyGeNEXT、ThyraMIR）以及将AI增强超声影像组学与分子生物标志物检测相结合的整合多模态方法在不确定甲状腺结节中诊断性能的研究。检索限于英文出版物。应用纳入和排除标准后，共纳入49项研究。提取的结局包括灵敏度、特异度、阴性预测值（NPV）和阳性预测值（PPV）。研究质量采用QUADAS-2进行评价，模型报告透明度以TRIPOD和STARD-AI指南为基准进行评价。由于方法学异质性，未进行荟萃分析或统计合并；汇总值为未加权描述性均值，并以观察范围呈现。结果：常规超声联合FNA细胞学表现出高灵敏度（71-99.6%），但变异 特异性（14.1-92%），反映出Bethesda III/IV类结节存在持续的诊断不确定性。与传统方法相比，人工智能增强超声往往能提高特异性并减少假阳性结果，但各研究间表现差异显著。分子生物标志物平台总体表现出较高的诊断性能，描述性（未加权）平均敏感性为86.8%，特异性为78.5%，阴性预测值为91.8%，使许多患者免于诊断性叶切除术。仅少数研究评估了人工智能+生物标志物联合模型；这些初步结果提示具有潜在的叠加价值，但在得出结论前仍需进一步的前瞻性验证。 结论：分子生物标志物检测和人工智能辅助超声在提高不确定甲状腺结节诊断准确性方面均显示出前景，早期证据表明，整合这些手段可能进一步优化恶性风险评估。然而，当前证据尚属初步且异质性较大，联合方法获益尚未得到证实。在常规临床采用前，需要进行采用标准化方案和透明报告的大型多中心前瞻性试验，以确认上述发现并评估真实世界中的成本效益。
完整摘要 · 英文原文
BACKGROUND: Indeterminate thyroid nodules (Bethesda III/IV) are difficult to manage: ultrasound and fine-needle aspiration (FNA) often yield non-definitive results, so many patients undergo diagnostic lobectomy for lesions that prove benign. The combination of AI-enhanced ultrasound and molecular biomarker platforms provides complementary diagnostic capabilities. However, evidence on their combined application remains limited, derives from a small number of early studies, and has not yet been validated in large or prospective cohorts. METHODS: A systematic search of PubMed (2012-2025) identified studies evaluating the diagnostic performance of conventional ultrasound + FNA, AI-assisted ultrasound, molecular biomarker platforms (Afirma, ThyroSeq, ThyGeNEXT, ThyraMIR), and integrated multimodal approaches combining AI-enhanced ultrasound radiomics with molecular biomarker testing in indeterminate thyroid nodules. The search was limited to English-language publications. After applying inclusion and exclusion criteria, 49 studies were included. Extracted outcomes included sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV). Study quality was appraised using QUADAS-2, and model reporting transparency was benchmarked against TRIPOD and STARD-AI guidelines. Because of methodological heterogeneity, no meta-analysis or statistical pooling was performed; summary values are unweighted descriptive means presented with observed ranges. RESULTS: Conventional ultrasound combined with FNA cytology demonstrated high sensitivity (71-99.6%) but variable specificity (14.1-92%), reflecting persistent diagnostic uncertainty in Bethesda III/IV nodules. AI-enhanced ultrasound tended to improve specificity and reduce false-positive findings compared with conventional methods, though performance varied considerably across studies. Molecular biomarker platforms showed generally high diagnostic performance, with a descriptive (unweighted) mean sensitivity of 86.8%, specificity 78.5%, and NPV 91.8%, enabling many patients to avoid diagnostic lobectomy. Only a few studies have evaluated integrated AI + biomarker models; these preliminary results suggest potential additive value but require further prospective validation before conclusions can be drawn. CONCLUSIONS: Molecular biomarker testing and AI-assisted ultrasound each show promise for enhancing diagnostic accuracy in indeterminate thyroid nodules, and early evidence suggests that integrating these modalities could further refine malignancy-risk assessment. However, the current evidence is preliminary and heterogeneous, and the benefit of combined approaches has not yet been confirmed. Large, multicenter prospective trials with standardized protocols and transparent reporting are needed to confirm these findings and to assess real-world cost-effectiveness before routine clinical adoption.