AI乳腺超声工具在妊娠及哺乳期患者的诊断准确性
本研究回顾性纳入2018-2021年两家学术中心504例妊娠或哺乳期女性共639个乳腺超声病灶,以活检或至少2年随访为参考标准。AI决策支持工具与放射科医师对照,两者敏感度均为80.0%(4/5)。排除乳腺囊肿、积液和皮肤病变后,AI与医师活检建议率相似(27.4% vs 23.8%,P=0.18)。
要点
- 妊娠及哺乳期乳腺病灶恶性率仅0.8%(5/639),AI与医师敏感度相同。
- AI将更多良性病灶归为BI-RADS 2、更少归为BI-RADS 3。
- 排除乳腺囊肿、积液、皮肤病变后AI与医师活检率无显著差异。
核心结论
- 妊娠及哺乳期患者中,AI乳腺超声决策支持工具敏感度与放射科医师相当。
- 排除AI工具适用范围外的病灶后,AI归为BI-RADS 2更多、BI-RADS 3更少。
研究设计与样本
- 回顾性研究,纳入2018-2021年两家学术医学中心连续妊娠或哺乳期有乳腺主诉的患者。
- 共504例女性(平均33岁,范围19-45岁)、639个乳腺超声病灶,病灶BI-RADS 2-5类。
- 参考标准为活检病理或至少2年随访,AI工具与放射科医师判读结果用McNemar检验比较。
关键影像表现/征象
- 5个恶性病灶中4个被放射科医师和AI均判为BI-RADS 4或5,1个均判为BI-RADS 3。
- 良性病灶中AI活检建议率高于医师(37.1% vs 21.0%,P<0.001)。
- 排除乳腺囊肿、积液、皮肤病变后AI与医师活检建议率差异无统计学意义(27.4% vs 23.8%,P=0.18)。
- AI将更多良性病灶归为BI-RADS 2(53.1% vs 32.9%,P<0.001),更少归为BI-RADS 3(19.5% vs 43.3%,P<0.001)。
诊断标准与关键数值
- 恶性检出敏感度:AI工具80.0%(4/5),放射科医师80.0%(4/5)。
- 良性病灶活检建议率:AI 37.1%,放射科医师21.0%,P<0.001。
- 排除乳腺囊肿、积液、皮肤病变后活检建议率:AI 27.4%,放射科医师23.8%,P=0.18。
- BI-RADS 2分类率:AI 53.1%,放射科医师32.9%,P<0.001。
- BI-RADS 3分类率:AI 19.5%,放射科医师43.3%,P<0.001。
临床应用与阅片要点
- 妊娠及哺乳期乳腺超声使用AI辅助判读时,需先排除乳腺囊肿、积液和皮肤病变等工具适用范围外病灶。
- AI倾向于减少BI-RADS 3评估、增加BI-RADS 2评估,可能降低不必要的短期随访。
- AI总体活检建议率高于放射科医师,临床应用时需结合病灶类型和患者背景综合判断。
局限性
- 回顾性设计,存在选择性偏倚可能。
- 恶性病灶仅5个(0.8%),敏感度估计置信区间宽。
- 数据来自两家学术医学中心,外部推广性有限。
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: To assess the diagnostic performance of an artificial intelligence (AI)-based decision support tool for breast ultrasound in pregnant and lactating patients and compare BI-RADS assessments with radiologist interpretation.
MATERIALS AND METHODS: This retrospective study included consecutive pregnant or lactating patients presenting with breast complaints at two academic medical centers from 2018-2021. Eligible patients underwent ultrasound with one or more findings assessed as BI-RADS category 2-5 and had biopsy or at least two years of follow-up. Examinations were analyzed using an AI-based decision support tool and compared with radiologist interpretations using the reference standard. Diagnostic performance was compared using McNemar's test.
RESULTS: A total of 504 women (mean age, 33 years; range, 19-45) with 639 breast ultrasound findings were included. Five findings (0.8%) were malignant, and 634 (99.2%) were benign. Both radiologists and the AI tool classified four malignancies as BI-RADS 4 or 5 and one as BI-RADS 3, yielding a sensitivity of 80.0% (4/5). Among benign lesions, the AI tool recommended more biopsies than radiologists (37.1% vs 21.0%, p < 0.001). After excluding galactoceles, fluid collections, and skin lesions, biopsy recommendation rates were similar (27.4% vs 23.8%, p = 0.18). The AI tool categorized more benign lesions as BI-RADS 2 (53.1% vs 32.9%, p < 0.001) and fewer as BI-RADS 3 (19.5% vs 43.3%, p < 0.001).
CONCLUSION: In pregnant and lactating patients, an AI-based decision support tool demonstrated sensitivity comparable to that of radiologists. After excluding lesions outside the AI tool's intended use, the AI tool assigned more BI-RADS 2 and fewer BI-RADS 3 assessments.
KEY POINTS: Question Can an AI-based decision support tool for breast ultrasound perform comparably to radiologists in pregnant and lactating patients? Findings The AI tool demonstrated sensitivity comparable to radiologists and categorized fewer lesions as BI-RADS 3; biopsy rates were similar after excluding lesions outside intended use. Clinical relevance AI-based decision support for breast ultrasound may reduce unnecessary follow-up imaging in pregnant and lactating patients after excluding lesions outside the AI tool's intended use, while demonstrating sensitivity comparable to radiologists.
引用
- Dwan, Lamb, Haver, et al. Diagnostic accuracy of an artificial intelligence-based breast ultrasound tool in pregnant and lactating patients. Eur Radiol. 2026. PMID: 42268305. doi:10.1007/s00330-026-12659-5
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。