深度学习R波检测用于心脏MRI心电图门控
本研究回顾性收集120例心脏MRI患者及47例外部队列的心电图,评估门控错误频率,并开发卷积神经网络(CNN)检测R波。结果显示1.5 T和3.0 T门控错误率分别为8.1%和15.2%,CNN在1.5 T的F1分数(99.1%)高于VCG(98.1%,P=0.048),在3.0 T高于Hamilton算法(99.1% vs 92.4%,P=0.049),假阳性率(0.1%)低于Hamilton算法(7.4%,P=0.02),并展示了回顾性图像重建改善图像质量的可行性。
要点
- 1.5 T和3.0 T心脏MRI心电图门控错误率分别为8.1%和15.2%。
- CNN在3.0 T的F1分数(99.1%)高于Hamilton算法(92.4%,P=0.049)。
- CNN在3.0 T的假阳性率(0.1%)低于Hamilton算法(7.4%,P=0.02)。
核心结论
- CNN为心电图R波检测提供了稳健方法,尤其在3.0 T存在MRI诱导伪影时;这些发现支持CNN提升心电图门控和心脏MR图像质量的潜力。
研究设计与样本
- 回顾性收集2022年8月至2023年4月120例在1.5 T和3.0 T下行心脏电影MRI患者的心电图,以及47例外部队列心电图;使用79例MRI和6例外部患者数据开发CNN,其余数据用于测试;CNN性能与向量心电图(VCG)门控和Hamilton算法比较。
诊断标准与关键数值
- 心电图门控错误在1.5 T患者中占8.1%,在3.0 T患者中占15.2%。
- 1.5 T时CNN的F1分数为99.1%,高于VCG的98.1%(P=0.048)。
- 3.0 T时CNN的F1分数为99.1%,高于Hamilton算法的92.4%(P=0.049),与VCG的95.1%无统计学差异(P=0.68)。
- 3.0 T时CNN假阳性率为0.1%,低于Hamilton算法的7.4%(P=0.02),与VCG的4.5%无统计学差异(P=0.28)。
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
Purpose To measure the frequency of electrocardiographic (ECG) gating errors and to assess whether convolutional neural networks (CNNs) can reduce such errors. Materials and Methods ECG tracings from 120 patients who underwent cine cardiac MRI at 1.5 T and 3.0 T, along with an external dataset of ECG tracings from 47 patients, were retrospectively collected (August 2022-April 2023). The frequency of arrhythmias and ECG-gating errors was determined by manual annotation of R waves. A CNN was developed for R-wave detection using data from 79 MRI and six external patients; the remaining data were used for testing. CNN performance was compared using signal processing algorithms, including vectorcardiographic (VCG) gating and the Hamilton algorithm. Results The mean age of the 120 patients undergoing cardiac MRI was 53.7 years ± 19.8 (SD), with 72 male patients. ECG-gating errors were observed in 8.1% of patients at 1.5 T and 15.2% at 3.0 T. At 1.5 T, CNN achieved a higher F1 score (99.1%) than did VCG (98.1%; P = .048). At 3.0 T, CNN achieved a higher F1 score (99.1%) than did the Hamilton algorithm (92.4%; P = .049) and was not statistically different than VCG (95.1%; P = .68). The CNN demonstrated a lower false-positive rate (0.1%) than did the Hamilton algorithm (7.4%; P = .02) at 3.0 T and was not statistically different than VCG (4.5%; P = .28). The feasibility of retrospective image reconstruction to improve cardiac MRI quality was also demonstrated. Conclusion CNNs offer a robust method for ECG R-wave detection, particularly in the presence of MRI-induced artifacts at 3.0 T. These findings support the potential of CNNs to enhance ECG gating and improve cardiac MR image quality. Keywords: Convolutional Neural Network, ECG Gating, Deep Learning R-Wave Detection, Cardiac MRI Reconstruction Supplemental material is available for this article. © RSNA, 2025.
引用
- Mahmoodi A, Hosseiny M, Ermakov V, et al. Deep Learning R-Wave Detection for Electrocardiographic Gating in Cardiac MRI. Radiol Cardiothorac Imaging. 2025. PMID: 41196189. doi:10.1148/ryct.250104
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。