AI 对筛查假阴性乳腺X线片中间期癌的检出分析
本研究回顾性分析 211517 例筛查乳腺X线片,纳入 96 例阴性或良性报告后一年内确诊的中间期乳腺癌。由两名放射科医生评估回顾可见性,并评价 FDA 批准 AI 工具在病灶层面的检出能力。结果显示 AI 检出 40%(38/96)的中间期癌,其中回顾可见者检出 81%,不可见者仅 7%。
要点
- AI 对回顾可见的中间期乳腺癌检出率高(81%),可作为次级阅片安全网。
- 真实中间期癌和乳腺X线隐匿癌 AI 几乎无法检出,仍需补充筛查手段。
- 回顾不可见不能排除为癌,报告中需提示补充超声或 MRI 的临床路径。
核心结论
- AI 检出大多数回顾可见的中间期乳腺癌,但对不可见癌检出能力有限。
- AI 检出与病灶回顾可见性、症状性表现、超声引导活检方式及更高 T 分期相关。
- AI 可作为可见癌的辅助安全网,但隐匿癌和真实中间期癌仍需补充筛查。
研究设计与样本
- 回顾性研究,纳入 2013 至 2020 年 211517 例筛查乳腺X线检查中的 96 例中间期乳腺癌。
- 两名放射科医生独立分类回顾可见性,不一致病例由第三人裁决;一致性以 Cohen's kappa 评估。
- 使用一种 FDA 批准的 AI 工具在病灶层面对假阴性筛查乳腺X线片进行评分。
关键影像表现/征象
- AI 检出 81%(34/42)回顾可见的中间期乳腺癌,仅检出 7%(4/54)回顾不可见者(p < 0.005)。
- 几乎全部漏诊病灶(91%,10/11)及全部最低限度可处理病灶(100%,13/13)被 AI 检出。
- 真实中间期癌(0%,0/14)、乳腺X线隐匿癌(10%,4/39)和技术性漏诊(0%,0/1)极少被 AI 检出。
诊断标准与关键数值
- AI 总体检出率 40%(38/96),回顾可见性 AI 检出优势比 OR 53.1(p < 0.001)。
- 症状性表现 OR 8.1(p = 0.007),超声引导活检 OR 5.7(p = 0.027),更高 T 分期 OR 3.4(p = 0.005)。
- 两名放射科医生对回顾可见性判断的一致性为 kappa = 0.69。
临床应用与阅片要点
- 在日常筛查报告中,AI 阴性不能排除中间期癌,尤其是致密乳腺或症状性病例。
- AI 阳性病灶应提高警惕并纳入报告,尤其是回顾可见的低限度病灶。
- 对 AI 未检出的可疑表现,仍需按临床指征补充超声或 MRI。
局限性
- 回顾性单中心研究设计可能影响结果的外推性。
- AI 评估为病灶层面分析,不能直接等同于筛查场景中的病例级诊断效能。
- 真实中间期癌和乳腺X线隐匿癌的样本量较小。
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
RATIONALE AND OBJECTIVES: Interval breast cancers (IBCs) occur after false-negative (FN) screening mammograms and can be diagnostically challenging. This study classified IBCs by retrospective visibility, evaluated an artificial intelligence (AI) tool's ability to detect IBCs, and identified factors associated with detection. MATERIALS AND METHODS: This retrospective study of 211,517 screening mammograms (2013-2020) included 96 IBCs diagnosed within a year of a negative/benign exam. Two radiologists independently classified IBC retrospective visibility; a third resolved discordant cases. Cohen's kappa assessed inter-reader agreement. A Food and Drug Administration-approved AI tool scored FN mammograms at the lesion level. Associations with AI detection were analyzed using chi-square/Fisher's exact tests and logistic regression.
RESULTS: AI detected 40% (38/96) of IBCs, including 81% (34/42) retrospectively visible and 7% (4/54) not visible (p < 0.005). Interreader agreement was substantial (kappa = 0.69). Nearly all misses (91%, 10/11), all minimal actionable findings (100%, 13/13), and most minimal nonactionable findings (61%, 11/18) were detected by AI. True interval (0%, 0/14), mammographically occult (10%, 4/39), and technical misses (0%, 0/1) were rarely detected. On univariate analysis, AI detection was significantly associated with retrospective visibility (odds ratio [OR] 53.1, p < 0.001), symptomatic presentation (OR 8.1, p = 0.007), ultrasound-guided biopsy (OR 5.7, p = 0.027), and higher T stage (OR 3.4, p = 0.005).
CONCLUSION: AI detected most retrospectively visible IBCs but had limited performance for cancers not visible. AI detection was associated with visibility, presentation, biopsy modality, and T stage. While AI offers a safety net for visible cancers, supplemental screening remains necessary for mammographically occult and true interval cancers.
引用
- Panigrahi, Oluyemi, Mullen, et al. Artificial Intelligence Detection of Interval Breast Cancers on False-negative Screening Mammograms: A Retrospective Lesion-level Analysis. Acad Radiol. 2026. PMID: 42642316. doi:10.1016/j.acra.2026.08.034
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。