Woo Ok Hee, Song Sung Eun, Choe Su Jin, Kim Minhye, Cho Kyu Ran, Seo Bo Kyoung
Department of Radiology, Korea University Guro Hospital, Seoul, Korea.
Department of Radiology, Korea University Anam Hospital, Korea University College of Medicine, 73 Goryeodae-ro, Seongbuk-gu, Seoul 02841, Korea.
Radiology. 2025 Jun;315(3):e242408. doi: 10.1148/radiol.242408.
Background Little is known about the features of invasive breast cancers missed by artificial intelligence (AI) on mammograms. Purpose To assess the false-negative rate (FNR) of AI mammogram evaluation according to molecular subtype and to investigate the features of and reasons for AI-missed cancers. Materials and Methods This retrospective study identified consecutive patients diagnosed with breast cancer between January 2014 and December 2020. Commercial AI software was used to read the mammograms, and abnormality score (AS) was acquired. AI-missed cancers were defined as those for which AI did not identify a precise location matching the reference standard. The FNR was calculated by counting AI-missed cancers according to molecular subtype (hormone receptor-positive [luminal] vs human epidermal growth factor receptor 2 [HER2]-enriched vs triple-negative). Three blinded radiologists classified AI-missed cancers as either actionable or under threshold, and reasons for misses were determined through nonblinded reviews. Features were compared according to AI detection with the χ test. Results A total of 1082 consecutive women diagnosed with 1097 cancers (mean age, 54.3 years ± 11 [SD]) were included. AI missed 14% (154 of 1097) of cancers. The FNR was lowest in the HER2-enriched subtype (9% [36 of 398] in the HER2-enriched subtype, 17.2% [106 of 616] in the luminal subtype, and 14.5% [12 of 83] in the triple-negative subtype; = .001). Compared with AI-detected cancers, AI-missed cancers were associated with younger age, a tumor size less than or equal to 2 cm, a lower histologic grade, fewer lymph node metastases, more Breast Imaging Reporting and Data System category 4 findings, lower Ki-67 expression, and nonmammary zone locations (all, < .05). In blinded reviews, 61.7% (95 of 154) of AI-missed cancers were actionable; the reasons for misses were dense breasts ( = 56), nonmammary zone locations ( = 22), architectural distortions ( = 12), and amorphous microcalcifications ( = 5). Conclusion To reduce AI-missed cancers on mammograms, attention should be given to luminal cancer, dense breasts, nonmammary zone locations, architectural distortions, and amorphous calcifications. Published under a CC BY 4.0 license. See also the editorial by Mullen in this issue.
关于人工智能（AI）在乳房X光检查中漏诊的浸润性乳腺癌的特征，人们知之甚少。目的：评估人工智能乳房X光检查评估根据分子亚型的假阴性率（FNR），并调查人工智能漏诊癌症的特征及原因。材料与方法：这项回顾性研究纳入了2014年1月至2020年12月期间连续诊断为乳腺癌的患者。使用商用人工智能软件读取乳房X光片，并获取异常评分（AS）。人工智能漏诊的癌症定义为人工智能未识别出与参考标准匹配的精确位置的癌症。通过根据分子亚型（激素受体阳性[管腔型]与人表皮生长因子受体2[HER2]富集型与三阴性）计算人工智能漏诊的癌症来计算FNR。三位盲法放射科医生将人工智能漏诊的癌症分类为可操作的或低于阈值的，并通过非盲法审查确定漏诊的原因。根据人工智能检测结果，使用χ检验比较特征。结果：共纳入1082名连续诊断为1097例癌症的女性（平均年龄，54.3岁±11[标准差]）。人工智能漏诊了14%（1097例中的154例）的癌症。FNR在HER2富集亚型中最低（HER2富集亚型中为9%[398例中的36例]，管腔型中为17.2%[616例中的106例]，三阴性亚型中为14.5%[83例中的12例]；P = .001）。与人工智能检测到的癌症相比，人工智能漏诊的癌症与年龄较小、肿瘤大小小于或等于2 cm、组织学分级较低、淋巴结转移较少、更多的乳腺影像报告和数据系统4类发现、较低的Ki-67表达以及非乳腺区域位置相关（所有P < .05）。在盲法审查中，61.7%（154例中的95例）的人工智能漏诊癌症是可操作的；漏诊的原因是乳房致密（P = 56）、非乳腺区域位置（P = 22）、结构扭曲（P = 12）和无定形微钙化（P = 5）。结论：为了减少乳房X光检查中人工智能漏诊的癌症，应关注管腔型癌症、乳房致密、非乳腺区域位置、结构扭曲和无定形钙化。根据知识共享署名4.0许可发布。另见本期穆伦的社论。
