基于深度学习的冠状动脉CT血管造影动脉粥样硬化斑块分割
研究利用308例894次冠状动脉CT血管造影(CCTA)扫描和3035个专家手动分割斑块训练3D U-net深度学习模型,在测试集上量化总斑块、非钙化斑块、钙化斑块和低衰减非钙化斑块体积的组内相关系数(ICC)分别为0.88、0.84、0.99和0.25,优于现有最小成本方法;在外部数据集上迁移学习后性能大幅提升,结论为开放源码深度学习算法与专家分割一致性优秀,但不同斑块特征或设备可能需要迁移学习。
要点
- 深度学习模型量化总斑块、非钙化、钙化和低衰减非钙化斑块体积的ICC分别为0.88、0.84、0.99和0.25。
- 在测试集上深度学习量化上述斑块体积的ICC均高于最小成本方法。
- 外部数据集迁移学习后总斑块ICC从0.62升至0.94,低衰减非钙化斑块从0.48升至0.86。
核心结论
- 开放源码深度学习算法与专家冠状动脉疾病分割达到优秀一致性,但不同斑块特征或设备可能需要迁移学习才能获得准确分割。
研究设计与样本
- 研究为深度学习模型开发与评估,使用308例个体的894次CCTA扫描和3035个由专家读者手动分割的斑块作为参考标准,按60%(186/308)训练、20%(61/308)调参验证、20%(61/308)测试3D U-net模型,并在不同站点采集的50例易损斑块个体外部测试集上评估,另对77例个体数据进行迁移学习后以组内相关系数(ICC)重新评估模型性能。
诊断标准与关键数值
- 测试集上深度学习量化总斑块体积的ICC为0.88(95%CI 0.85-0.91),最小成本方法为0.63(95%CI 0.42-0.76)。
- 测试集上深度学习量化非钙化斑块体积的ICC为0.84(95%CI 0.80-0.88),最小成本方法为0.45(95%CI 0.26-0.59)。
- 测试集上深度学习量化钙化斑块体积的ICC为0.99(95%CI 0.98-0.99),最小成本方法为0.96(95%CI 0.94-0.97)。
- 测试集上深度学习量化低衰减非钙化斑块体积的ICC为0.25(95%CI 0.13-0.37),最小成本方法为-0.01(95%CI -0.13至0.11)。
- 外部数据集迁移学习前深度学习量化总斑块、非钙化、钙化、低衰减非钙化斑块体积的ICC分别为0.62(95%CI 0.01-0.84)、0.54(95%CI -0.04至0.80)、0.91(95%CI 0.85-0.95)、0.48(95%CI 0.18-0.69),迁移学习后分别为0.94(95%CI 0.87-0.97)、0.93(95%CI 0.86-0.96)、0.95(95%CI 0.91-0.97)、0.86(95%CI 0.76-0.92)。
临床应用与阅片要点
- 深度学习3D U-net模型可用于冠状动脉斑块体积量化,与专家读者结果相当。
- 对其它扫描设备或斑块特征,可能需要迁移学习才能达到相似分割结果。
- 所开发的深度学习算法为开放源码,可在任何CT分析软件中实现。
局限性
- 在斑块特征或设备不同的情况下,可能需要迁移学习才能实现准确分割。
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
OBJECTIVES: Volumetric evaluation of coronary artery disease (CAD) allows better prediction of cardiac events. However, CAD segmentation is labor intensive. Our objective was to create an open-source deep learning (DL) model to segment coronary plaques on coronary CT angiography (CCTA).
METHODS: Three hundred eight individuals' 894 CCTA scans with 3035 manually segmented plaques by an expert reader (considered as ground truth) were used to train (186/308, 60%), validate (tune, 61/308, 20%), and test (61/308, 20%) a 3D U-net model. We also evaluated the model on an external test set of 50 individuals with vulnerable plaques acquired at a different site. Furthermore, we applied transfer learning on 77 individuals' data and re-evaluated the model's performance using intra-class correlation coefficient (ICC).
RESULTS: On the test set, DL outperformed the currently used minimum cost approach method to quantify total: ICC: 0.88 [CI: 0.85-0.91] vs. 0.63 [CI: 0.42-0.76], noncalcified: 0.84 [CI: 0.80-0.88] vs. 0.45 [CI: 0.26-0.59], calcified: 0.99 [CI: 0.98-0.99] vs. 0.96 [CI: 0.94-0.97], and low attenuation noncalcified: 0.25 [CI: 0.13-0.37] vs. -0.01 [CI: -0.13 to 0.11] plaque volumes. On the external dataset, substantial improvement was observed in DL model performance after transfer learning, total: 0.62 [CI: 0.01-0.84] vs. 0.94 [CI: 0.87-0.97], noncalcified: 0.54 [CI: -0.04 to 0.80] vs. 0.93 [CI: 0.86-0.96], calcified: 0.91 [CI:0.85-0.95] vs. 0.95 [CI: 0.91-0.97], and low attenuation noncalcified 0.48 [CI: 0.18-0.69] vs. 0.86 [CI: 0.76-0.92].
CONCLUSIONS: Our open-source DL algorithm achieved excellent agreement with expert CAD segmentations. However, transfer learning may be required to achieve accurate segmentations in the case of different plaque characteristics or machinery.
KEY POINTS: • Deep learning 3D U-net model for coronary segmentation achieves comparable results with expert readers' volumetric plaque quantification. • Transfer learning may be needed to achieve similar results for other scanner and plaque characteristics. • The developed deep learning algorithm is open-source and may be implemented in any CT analysis software.
引用
- Jávorszky, Homonnay, Gerstenblith, et al. Deep learning-based atherosclerotic coronary plaque segmentation on coronary CT angiography. Eur Radiol. 2022. PMID: 35524783. doi:10.1007/s00330-022-08801-8
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。