深度学习多任务注意力网络自动量化颈动脉斑块负荷
本研究提出一种基于深度多任务注意力网络的方法,用于从常规B型超声图像中自动量化颈动脉斑块面积。研究使用1100张专家标注图像,其中200张用于独立验证。结果显示该方法预测斑块面积的平均绝对误差为0.001324,组内相关系数达0.85,85%的测量值在临床可接受误差范围内。
要点
- 深度学习可自动从常规颈动脉超声图像中量化斑块面积,减少操作者依赖性。
- 模型预测斑块面积的组内相关系数为0.85,测量可靠性较高。
- 85%的测量值误差小于4.0 mm²,可检测超过5.0 mm²的斑块进展。
核心结论
- 基于注意力引导的深度多任务学习能够实现颈动脉斑块负荷的准确、可重复量化。
- 该方法可支持可靠检测超过5.0 mm²的临床显著性斑块进展,置信度为95%。
- 与U-Net分割方法相比,平均绝对误差改善了65.2%;与ResNet-50纯回归方法相比改善了37.0%。
研究设计与样本
- 研究类型为方法学验证研究,提出ResNet-50骨干结合卷积块注意力模块的多任务学习框架。
- 共使用1100张专家标注的颈动脉超声图像,其中200张作为独立验证集。
- 参考标准为专家标注的斑块面积,模型联合学习斑块存在性与归一化斑块面积。
关键影像表现/征象
- 模型预测斑块面积与参考值之间的Pearson相关系数为0.712。
- Bland-Altman分析显示平均差异为0.0003,一致性限为-0.0034至0.0039,偏倚极小。
- 组内相关系数达到0.85,表明测量可靠性较高。
诊断标准与关键数值
- 验证集上平均绝对误差为0.001324,相当于临床上的2.44 mm²。
- 验证集上均方根误差为0.001891。
- 85%的测量值落在临床可接受误差阈值内,即误差小于4.0 mm²。
- 以95%置信度可检测超过5.0 mm²最小可检测变化的斑块进展。
临床应用与阅片要点
- 该自动量化方法可用于日常颈动脉超声报告中斑块面积的客观测量,减少人工勾画变异。
- 临床可采用5.0 mm²作为斑块进展的最小可检测变化阈值,以95%置信度判断进展。
- 注意该模型基于专家标注数据训练,在实际使用时需结合超声图像质量控制。
局限性
- 独立验证集仅200张图像,样本量有限。
- 研究为单中心数据,缺乏多中心外部验证。
- 模型预测与参考斑块面积的相关性为0.712,仍存在一定预测误差。
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
Carotid atherosclerotic plaque burden is a well-established biomarker of cerebrovascular and cardiovascular risk, yet its quantitative assessment from ultrasound imaging remains highly operator-dependent and poorly standardized. This study proposes a deep multi-task attention-based learning framework for automated, reproducible quantification of carotid plaque area from routine B-mode ultrasound images. The model integrates a ResNet-50 backbone with convolutional block attention modules (CBAM) to jointly learn plaque presence and normalized plaque area within a unified architecture. A total of 1,100 expert-annotated carotid ultrasound images were used, with 200 images reserved for independent validation. On the validation set, the proposed approach achieved a mean absolute error (MAE) of 0.001324 (corresponding to 2.44 mm² clinically), a root mean square error (RMSE) of 0.001891, and a Pearson correlation coefficient of 0.712 between predicted and reference plaque areas. The model demonstrated 65.2% improvement in MAE over U-Net segmentation pipelines and 37.0% improvement over ResNet-50 regression-only approaches. Bland-Altman analysis revealed minimal bias (mean difference: 0.0003) with narrow limits of agreement (-0.0034 to 0.0039), while intraclass correlation coefficient (ICC) reached 0.85, indicating excellent measurement reliability. Clinical assessment showed that 85% of measurements fell within clinically acceptable error thresholds (<4.0 mm²), enabling detection of plaque progression exceeding the 5.0 mm² minimal detectable change with 95% confidence. These findings demonstrate that attention-guided deep multi-task learning enables accurate, reproducible quantification of carotid plaque burden, directly supporting the reliable detection of clinically significant plaque progression and advancing clinical translation through agreement-centered validation.
引用
- Khan, Bano, Rehman, et al. Deep Multi-Task Attention Network for Automated, Reproducible Quantification of Carotid Plaque Burden in Ultrasound Imaging. Ultrasound Med Biol. 2026. PMID: 42248793. doi:10.1016/j.ultrasmedbio.2026.05.004
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。