文献速递乳腺MR

多参数MRI深度学习模型评估浸润性乳腺导管癌脉管侵犯:一项多中心回顾性研究

Multi-parameter MRI deep learning model for lymphovascular invasion assessment in invasive breast ductal carcinoma: A multicenter, retrospective study.
PubMed 文献 · Clin Radiol · 2025-06-25

本研究基于两中心448例浸润性乳腺导管癌患者的回顾性数据,构建多参数MRI深度学习模型预测脉管侵犯状态。T1WI-CE3深度学习模型在验证集和测试集的AUC分别为0.842和0.774,优于T1WI-CE2模型。融合临床-影像特征与多参数MRI深度学习特征的联合模型验证集AUC为0.939(准确率0.892),测试集AUC为0.941(准确率0.872),校准曲线显示预测一致性良好,决策曲线分析证实其临床效用。

要点

  • 联合模型在测试集预测脉管侵犯的AUC为0.941,准确率为0.872。
  • T1WI-CE3深度学习模型验证集AUC 0.842、测试集AUC 0.774,均优于T1WI-CE2模型的0.748和0.619。
  • 校准曲线显示联合模型预测结果与观察结果一致性良好,决策曲线分析支持其临床效用。

核心结论

  • 整合多参数MRI与深度学习的联合模型可用于浸润性乳腺导管癌患者术前脉管侵犯评估,可能辅助临床医生制定更精准的治疗策略。

研究设计与样本

  • 两中心回顾性分析,纳入448例浸润性乳腺导管癌患者;中心1按8:2比例分为训练集和验证集,中心2作为测试集;采用MobileNetV2-3D深度学习模型比较T1WI-CE2与T1WI-CE3性能,选择后者进行融合任务;联合模型整合临床-影像特征与多参数MRI深度学习特征;以AUC评估预测性能,校准曲线评估预测与观察的一致性,决策曲线分析量化不同决策阈值下的净获益。

诊断标准与关键数值

  • T1WI-CE3深度学习模型验证集AUC为0.842,测试集AUC为0.774。
  • T1WI-CE2深度学习模型验证集AUC为0.748,测试集AUC为0.619。
  • 联合模型验证集AUC为0.939,准确率为0.892。
  • 联合模型测试集AUC为0.941,准确率为0.872。

临床应用与阅片要点

  • 联合模型整合临床-影像特征与多参数MRI深度学习特征,测试集AUC为0.941、准确率为0.872。
  • 校准曲线显示联合模型预测结果与观察结果一致性良好,决策曲线分析证实其临床效用。
  • 作者认为该工具可能辅助临床医生制定更精准的治疗策略。

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

AIMS: To investigate the value of multi-parametric magnetic resonance imaging (MRI)-based deep learning (DL) in predicting the Lymphovascular Invasion (LVI) status of invasive breast ductal cancer (IBDC).

MATERIALS AND METHODS: A retrospective analysis of 448 IBDC patients from two centers was conducted, with Center 1 split into training and validation sets (8:2 ratio) and Center 2 used as the test set. A MobileNetV2-3D DL model was used to compare T1WI-CE2 and T1WI-CE3 performance, selecting the latter for fusion tasks. The combined model integrates clinical-radiological features (CRF) and multi-parameter MRI DL features to improve predictive accuracy for LVI. The predictive performance of the model was evaluated by calculating the area under the ROC curve (AUC). Calibration curves were used to evaluate the agreement between predicted outcomes and observations. Furthermore, decision curve analysis (DCA) quantified the net benefit across decision thresholds.

RESULTS: The T1WI-CE3 DL model achieved AUCs of 0.842 and 0.774 on the validation and test sets, respectively, outperforming the T1WI-CE2 DL model, which had AUCs of 0.748 and 0.619. The combined model achieved a validation set AUC of 0.939 with 0.892 accuracy and a test set AUC of 0.941 with 0.872 accuracy. Calibration curves showed good prediction consistency, while DCA confirmed the combined model's clinical utility, supporting its accuracy, reliability, and potential for application.

CONCLUSIONS: This study offers a novel tool for preoperative LVI assessment in IBDC patients by integrating multi-parametric MRI with DL, potentially aiding clinicians in devising more precise treatment strategies.

来源:Clin Radiol,查看原文。版权归原作者所有。

引用

  1. Xu J, Wang G, Wei Y, et al. Multi-parameter MRI deep learning model for lymphovascular invasion assessment in invasive breast ductal carcinoma: A multicenter, retrospective study. Clin Radiol. 2025. PMID: 40695008. doi:10.1016/j.crad.2025.107002

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。