基于B型超声与彩色多普勒血流成像视频的多模态特征解耦模型用于淋巴结病诊断:一项回顾性、前瞻性、多中心研究
该研究基于6个中心7371例患者的B型超声与彩色多普勒血流成像视频构建并验证深度学习模型诊断淋巴结病,内部、回顾性外部和前瞻性外部验证队列AUC分别为0.956、0.928和0.912;模型辅助使低年资放射科医师在回顾性外部队列的平均AUC从0.739升至0.891,在前瞻性外部队列从0.767升至0.899。
要点
- 模型在内部、回顾性外部、前瞻性外部验证队列AUC分别为0.956、0.928、0.912。
- 回顾性外部队列中,低年资医师平均AUC由0.739升至0.891(模型辅助)。
- 前瞻性外部队列中,低年资医师平均AUC由0.767升至0.899(模型辅助)。
核心结论
- 基于多模态超声视频的深度学习模型提高了淋巴结病的诊断准确性,并显示出作为无创、高效临床决策工具的较大潜力。
研究设计与样本
- 回顾性及前瞻性研究,纳入2016年1月至2025年8月中国5个地区6个中心的7371例患者(男性3824例[51.9%],女性3547例[48.1%],中位年龄52岁[9-94岁]),均接受多模态超声检查。
- 从所有患者的B型超声和彩色多普勒血流成像视频中提取147420个关键帧用于模型训练和验证。
- 模型性能以受试者工作特征曲线下面积(AUC)、准确率(ACC)、敏感度(SEN)、特异度(SPE)和精确率(PRE)评估。
- 通过比较不同经验水平放射科医师独立诊断与模型辅助诊断的性能,验证模型的临床实用价值。
诊断标准与关键数值
- 模型在内部验证队列的AUC为0.956(95%CI:0.925-0.981)。
- 模型在回顾性外部验证队列的AUC为0.928(95%CI:0.884-0.965)。
- 模型在前瞻性外部验证队列的AUC为0.912(95%CI:0.863-0.952)。
- 回顾性外部队列中,低年资放射科医师平均AUC从0.739(95%CI:0.676-0.801)升至模型辅助下的0.891(95%CI:0.846-0.940)。
- 前瞻性外部队列中,低年资放射科医师平均AUC从0.767(95%CI:0.705-0.829)升至模型辅助下的0.899(95%CI:0.853-0.944)。
临床应用与阅片要点
- 模型在多中心数据集中区分良恶性淋巴结病的表现优于高年资放射科医师,可作为临床决策支持工具。
- 模型辅助可提高低年资放射科医师的诊断性能,并减少对操作者的依赖。
- 该模型有助于在不同资源条件的医疗机构中促进淋巴结病的早期发现和精准管理。
局限性
- 原文未报告局限性。
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
OBJECTIVES: This study developed and validated a deep learning model for diagnosing lymphadenopathy (LA) using B-mode ultrasound (BUS) and color Doppler flow imaging (CDFI) videos.
MATERIALS AND METHODS: A retrospective and prospective study was conducted from January 2016 to August 2025, including 7371 patients (3824 male [51.9%], 3547 females [48.1%], median age, 52 years [9-94 years]) who underwent multimodal ultrasound examinations across six centers from five regions of China. A total of 147,420 key frames were extracted from BUS and CDFI videos of all patients for model training and validation. Besides, patient clinical information was integrated to enhance the diagnostic performance of the model. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy (ACC), sensitivity (SEN), specificity (SPE), and precision (PRE). The clinical practical value of the model was verified by comparing with the performance of independent diagnosis and model-assisted diagnosis of radiologists with different levels of experience.
RESULTS: This model achieved AUCs of 0.956 (95% CI: 0.925-0.981), 0.928 (95% CI: 0.884-0.965), and 0.912 (95% CI: 0.863-0.952) in the internal, retrospective external, and prospective external validation cohorts, respectively. In the retrospective external cohort, the average AUC of junior radiologists improved from 0.739 (95% CI: 0.676-0.801) to 0.891 (95% CI: 0.846-0.940) with the assistance of the model. In the prospective external cohort, their average AUC improved from 0.767 (95% CI: 0.705-0.829) to 0.899 (95% CI: 0.853-0.944).
CONCLUSION: This multimodal video-based deep learning model enhances LA diagnostic accuracy and shows strong potential as a noninvasive, efficient tool for clinical decision-making.
KEY POINTS: Question Why is multimodal ultrasound video needed in clinical practice to enhance the automated assessment of LA? Findings The proposed multimodal ultrasound video AI model achieved high diagnostic accuracy and robustness, outperforming senior radiologists in distinguishing benign from malignant LA across multicenter datasets. Clinical relevance This model offers a reliable, noninvasive clinical decision-support tool that enhances diagnostic performance, reduces operator dependence, and facilitates early detection and precise management of LA across healthcare institutions with varying resources.
引用
- Cao R, Zhu Y, Zhao H, et al. A multimodal feature disentanglement model for lymphadenopathy diagnosis based on BUS and CDFI ultrasound videos: a retrospective, prospective, multicenter study. Eur Radiol. 2026. PMID: 41831029. doi:10.1007/s00330-026-12409-7
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。