基于nnU-Net的腹主动脉髂动脉钙化斑块自动检测与分割:用于平扫和增强CT的全躯干动脉粥样硬化负荷评估
本研究用nnU-Net在119例平扫PET-CT和23例增强CT尿路造影上训练腹主动脉髂动脉钙化斑块检测与分割模型,在多个内部和外部数据集上测试。检测精确率88.1%、召回率99.5%、F1分数93.4%,分割Dice分数64.3-83.7%,配对平扫与增强CT的Agatston评分强相关(R2=0.99)。
要点
- 模型检测腹主动脉髂动脉钙化斑块的精确率88.1%、召回率99.5%、F1分数93.4%。
- 分割Dice分数64.3-83.7%,优于基线方法超过10%(p < 0.001)。
- 配对平扫与增强CT的Agatston评分强相关(R2=0.99)。
核心结论
- 自动检测腹主动脉髂动脉斑块能够准确评估全躯干动脉粥样硬化钙化负荷,可能为改善心血管疾病诊断和治疗提供途径。
研究设计与样本
- 研究使用nnU-Net训练腹主动脉髂动脉斑块检测与分割模型,训练数据包括本机构的119例平扫全身PET-CT和23例增强腹部盆腔CT尿路造影。
诊断标准与关键数值
- 检测精确率88.1%、召回率99.5%、F1分数93.4%。
- 分割Dice分数64.3-83.7%,优于基线方法超过10%(p < 0.001)。
- 配对平扫与增强CT的Agatston评分强相关(R2=0.99)。
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
RATIONALE AND OBJECTIVES: Cardiovascular disease (CVD) is closely associated with aortoiliac plaque burden, yet current research on its automated detection and segmentation has largely focused on plaque burden analysis using CT angiography. In this study, we present an automated method for aortoiliac plaque detection and segmentation that enables accurate quantification of calcified plaque burden on both non-contrast and contrast-enhanced CT scans.
MATERIALS AND METHODS: The training data included 119 non-contrast whole-body PET-CT scans and 23 contrast-enhanced abdominopelvic CT urography scans, all obtained from our institution. The testing data comprised 99 contrast-enhanced thoracoabdominopelvic CT scans from the sarcopenia dataset; 93 from the prostate cancer dataset; 1214 paired non-contrast and contrast-enhanced abdominal CT scans from a renal donor cohort; 9199 non-contrast abdominal CT colonography scans from a second institution; and 1446 non-contrast chest CT scans from a third institution. The nnU-Net was used to train a model for aortoiliac plaque detection and segmentation. Detection accuracy was evaluated on non-contrast chest CT scans. Segmentation accuracy was assessed on CT scans with manually labeled plaque regions from the sarcopenia, prostate, renal donor, and CT colonography datasets. The correlation between Agatston scores on paired non-contrast and contrast-enhanced scans was evaluated in the renal donor cohort. Correlations between whole-torso calcified plaque burden (Agatston scores), demographics, and diseases were analyzed using multivariable analysis on the CT colonography dataset.
RESULTS: Aortoiliac plaques were detected with 88.1% precision, 99.5% recall, and a 93.4% F1 score. Segmentation achieved Dice scores of 64.3-83.7% across two internal contrast-enhanced and two external non-contrast CT datasets, outperforming baseline methods by over 10% (p < 0.001). Agatston scores from paired CT scans showed strong correlation (R2 = 0.99). Multivariate analysis showed calcified plaque burden assessment correlated with sex, age, BMI, and smoking (all p < 0.001), as well as alcohol abuse (p = 0.01). The calcified burden assessment was also correlated with CVD, heart failure, myocardial infarction (all p < 0.001), and type 2 diabetes (p = 0.03), but showed no correlation with cancer (p = 0.14) or femoral neck fracture (p = 0.61).
CONCLUSION: Automated aortoiliac plaque detection enables accurate whole-torso atherosclerotic calcified burden assessment, offering a potential pathway for improved CVD diagnosis and treatment.
引用
- Liu J, Batheja V, Mukherjee P, et al. Automated Detection and Segmentation of Aortoiliac Calcified Plaques Using nnU-Net for Whole-Torso Atherosclerotic Burden Assessment on Non-Contrast and Contrast-Enhanced CT Scans. Acad Radiol. 2025. PMID: 41176436. doi:10.1016/j.acra.2025.10.010
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。