文献速递乳腺乳腺X线

东西方数字乳腺断层摄影AI病灶检测迁移学习研究

International Application of Artificial Intelligence for Lesion Detection on Digital Breast Tomosynthesis: Comparing Western and Eastern Databases.
PubMed 文献 · Acad Radiol · 2026-08-25

基于数字乳腺断层摄影(DBT)数据,比较西方(94例)与东方(157例)数据库训练的人工智能病灶检测模型跨库表现。研究发现两库病灶类型差异显著(P<0.001),直接跨库应用模型性能显著下降(P<0.001)。采用迁移学习后模型在目标数据集上的IoU、敏感度、精确度均显著提升(P<0.001)。

要点

  • 不同人群DBT病灶类型差异大,东方数据钙化占比高达60%,跨库直接应用AI模型性能会显著下降。
  • 迁移学习可改善模型跨库泛化,东西双向迁移后敏感度可达约0.90,阅片工作中需关注AI训练数据来源。
  • 评估DBT AI检测性能时需同时关注IoU、敏感度与精确度,不宜仅凭单一指标判断模型可用性。

核心结论

  • 东西方数据库间DBT病灶类型分布差异显著(P<0.001),直接跨库应用检测模型性能显著下降(P<0.001)。
  • 迁移学习可显著提升模型在目标数据集上的检测性能(P<0.001),有利于DBT病灶检测AI的国际应用。

研究设计与样本

  • 研究采用西方数据库94例(源自Cancer Imaging Archive Breast Cancer Screening DBT)与东方数据库157例(源自一个匿名医学中心)的DBT数据。
  • 所有乳腺病灶被分层为六种类型,采用YOLO病灶检测模型,分别用单一数据库训练并在另一数据库测试,再以迁移学习优化后评估。
  • 评估指标为IoU(Intersection over Union)、敏感度与精确度。

关键影像表现/征象

  • 东方数据库与西方数据库的病灶类型构成存在显著差异(P<0.001),其中东方数据库钙化病灶占比达60%。

诊断标准与关键数值

  • 东方到西方的迁移学习模型在西方目标数据集上IoU为0.53±0.23,敏感度为0.90±0.05,精确度为0.93±0.04。
  • 西方到东方的迁移学习模型在东方目标数据集上IoU为0.56±0.22,敏感度为0.89±0.05,精确度为0.87±0.05。
  • 迁移学习后模型在目标数据集上的检测性能较直接跨库应用显著改善(P<0.001)。

临床应用与阅片要点

  • 在将DBT AI病灶检测工具引入本地区人群前,应确认训练数据的人群来源与本地区影像特征匹配程度。
  • 钙化在东方人群占比更高,阅片时面对钙化为主的DBT病例需警惕AI漏检风险。
  • 报告中可参考AI检测结果,但需结合IoU、敏感度与精确度综合评估模型对当前病例的可靠性。

局限性

  • 研究样本量有限(西方94例、东方157例),且东方数据仅来自单一匿名医学中心。

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: The international application of artificial intelligence (AI) for lesion detection based on digital breast tomosynthesis (DBT) is limited due to disease variations among populations. We hypothesized that lesion detection models trained on either the Western or Eastern DBT dataset would exhibit reduced performance on another dataset. We proposed transfer learning to enhance lesion detection across DBT databases.

MATERIALS AND METHODS: The Western database (94 patients) was obtained from the Cancer Imaging Archive Breast Cancer Screening DBT, and the Eastern database (157 patients) was collected from one anonymous medical center. All breast lesions were stratified into six types. The YOLO (You Only Look Once) lesion detection models were trained using one of the databases with/without transfer learning to another dataset. Intersection over Union (IoU), sensitivity, and precision were used to evaluate model performance and generality.

RESULTS: Significant differences were observed in lesion types (P < 0.001) between the two databases, with calcification more prevalent in the Eastern database (60%). A significant reduction in lesion detection performance (P < 0.001) was observed when applying detection models to another dataset. After transfer learning, the detection models showed significant improvements in detection performance on the target datasets (P < 0.001), with the Eastern-to-Western model achieving an IoU of 0.53 ± 0.23, sensitivity of 0.90 ± 0.05, and precision of 0.93 ± 0.04 on the Western target dataset, and the Western-to-Eastern model achieving an IoU of 0.56 ± 0.22, sensitivity of 0.89 ± 0.05, and precision of 0.87 ± 0.05 on the Eastern target dataset.

CONCLUSION: The difference in DBT lesion types between Western and Eastern databases significantly impacted model performance. The applied transfer learning effectively enhanced the model performance in the target dataset, which could benefit the international application of DBT lesion detection.

来源:Acad Radiol,查看原文。版权归原作者所有。

引用

  1. Tang, Lin, Lin, et al. International Application of Artificial Intelligence for Lesion Detection on Digital Breast Tomosynthesis: Comparing Western and Eastern Databases. Acad Radiol. 2026. PMID: 42642320. doi:10.1016/j.acra.2026.08.039

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。