AI超声列线图术前预测乳腺癌腋窝淋巴结转移
本研究回顾性纳入148例经病理证实的浸润性乳腺癌女性,整合AI提取的超声形态特征与临床病理参数构建列线图。训练集118例、验证集30例,列线图AUC分别为0.855和0.810。在Youden最佳阈值0.410下,训练集和验证集准确率分别为79.7%和80.0%,决策曲线显示优于单纯放射科医师评估。
要点
- 列线图纳入病灶深度≥1.0 cm、长度≥2.2 cm、宽度≥2.1 cm及不规则形态。
- 同侧腋窝淋巴结长轴≥2.2 cm、短轴≥0.7 cm为独立预测因素。
- Ki-67>20%及放射科医师评估淋巴结受累亦纳入最终模型。
核心结论
- AI辅助超声列线图对乳腺癌同侧腋窝淋巴结转移具有稳健的预测性能,优于单纯放射科医师评估。
研究设计与样本
- 回顾性研究,纳入2020年5月至2022年1月间148例经病理证实的浸润性乳腺癌女性,平均年龄58±11岁。
- 所有患者术前接受超声检查,采用Samsung Medison的S-Detect提取AI超声特征。
- 按8:2比例分为内部训练集118例和验证集30例,采用logistic回归构建预测模型,以组织病理为参考标准。
关键影像表现/征象
- 病灶深度≥1.0 cm、病灶长度≥2.2 cm、病灶宽度≥2.1 cm与腋窝淋巴结转移相关。
- 同侧腋窝淋巴结长轴≥2.2 cm、短轴≥0.7 cm为独立预测因素。
- 病灶形态不规则及放射科医师评估腋窝淋巴结受累为影像相关预测指标。
诊断标准与关键数值
- 最终模型纳入8个独立预测因素:病灶深度≥1.0 cm、病灶长度≥2.2 cm、病灶宽度≥2.1 cm、腋窝淋巴结长轴≥2.2 cm、腋窝淋巴结短轴≥0.7 cm、病灶形态不规则、Ki-67>20%及放射科医师评估淋巴结受累。
- 列线图训练集AUC为0.855,验证集AUC为0.810,Hosmer-Lemeshow检验p>0.05提示校准良好。
- Youden最佳阈值为0.410,训练集准确率79.7%、敏感度73.8%、特异度84.8%、阳性预测值75.6%、阴性预测值86.8%、F1评分0.746。
- 验证集准确率80.0%、敏感度76.9%、特异度82.4%、阳性预测值76.9%、阴性预测值82.4%、F1评分0.769。
- 总体队列中57/148例(38.5%)经病理证实存在同侧腋窝淋巴结转移。
临床应用与阅片要点
- 日常阅片时可重点关注病灶三维径线阈值及腋窝淋巴结长短轴阈值,辅助术前风险分层。
- 列线图在20%至60%阈值区间内的净临床获益优于单纯放射科医师评估。
- 放射科医师对腋窝淋巴结受累的评估仍是模型组成之一,需与AI特征联合解读。
局限性
- 本研究为单中心回顾性设计,验证集样本量仅30例。
- AI特征提取仅来源于Samsung Medison的S-Detect软件,结果外推至其他超声平台需谨慎。
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)。
英文原文(摘要)· 点击展开
OBJECTIVE: To develop and validate a nomogram integrating artificial intelligence (AI)-extracted ultrasound features with clinic pathologic data for non-invasive preoperative prediction of ipsilateral axillary lymph node (IALN) metastasis in breast cancer women.
MATERIALS AND METHODS: This retrospective study, approved by the institutional review board with consent waived, included 148 women (mean age, 58 y ± 11) with histologically confirmed invasive breast cancer who underwent preoperative ultrasound between May 2020 and January 2022. A predictive model was developed by integrating AI-extracted ultrasound features (S-Detect, Samsung Medison) with tumor size, IALN size, the proliferation marker Ki-67, and radiologist assessment. Performance was assessed via logistic regression, receiver operating characteristic (ROC) analysis, calibration, and decision-curve analysis (DCA) on internal training (n = 118) and validation (n = 30) cohorts, with accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and harmonic mean of precision and recall (F1 score) calculated at the Youden-optimal threshold.
RESULTS: IALN metastasis was identified in 57 of 148 patients (38.5%). Eight independent predictors were included in the final model: lesion depth ≥1.0 cm, lesion length ≥2.2 cm, lesion width ≥2.1 cm, IALN long axis ≥2.2 cm, IALN short axis ≥0.7 cm, irregular lesion shape, Ki-67 >20%, and radiologist-assessed IALN involvement. The nomogram achieved AUCs of 0.855 (training) and 0.810 (validation) with excellent calibration (Hosmer-Lemeshow p > 0.05). At the Youden-optimal threshold (0.410), the nomogram achieved stable performance: accuracy 79.7%/80.0%, sensitivity 73.8%/76.9%, specificity 84.8%/82.4%, PPV 75.6%/76.9%, NPV 86.8%/82.4%, and F1 score 0.746/0.769 (training/validation). DCA demonstrated superior net clinical benefit compared to radiologist assessment, particularly within the 20%-60% threshold range.
CONCLUSION: An AI-assisted ultrasound nomogram demonstrated robust predictive performance for IALN metastasis, outperformed radiologist assessment, and may optimize preoperative risk stratification and surgical decision-making.
引用
- Wang, Yu, Chou, et al. A Predictive Nomogram Integrating AI-Assisted Morphological Feature Extraction with Clinical and Ultrasound Parameters for Preoperative Prediction of Axillary Lymph Node Metastasis in Breast Cancer. Ultrasound Med Biol. 2026. PMID: 42414194. doi:10.1016/j.ultrasmedbio.2026.05.023
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。