用于超声心动图16节段左室壁厚度评估的深度学习系统的开发与验证
本研究开发并验证了基于左室分割的深度学习算法,在92,984例正常壁厚短轴超声心动图训练后,于肥厚壁厚队列(26,523例)和体外队列(2,238例)中评估,结果显示分割精度与手动分割相似,16节段壁厚测量误差在临床可接受范围内,可能为左室肥厚的准确诊断提供方法。
要点
- 深度学习自动分割左室心肌的平均Dice相似系数为0.87,平均Hausdorff距离正常壁厚队列为4.56mm、肥厚队列为4.81mm。
- 正常壁厚队列16节段壁厚平均绝对误差0.31mm至0.88mm,肥厚队列均在临床可耐受阈值2mm以内。
- 体外队列16节段壁厚平均绝对误差为1.0mm,验证了方法在体内外的准确性。
核心结论
- 深度学习推导的左室心肌轮廓与准确分割可行,可能为评估左室心肌提供方法;该方法在体内外验证了16节段壁厚的准确测量,可能为左室肥厚的准确诊断提供基础与临床应用。
研究设计与样本
- 深度学习算法的开发与验证研究,训练与验证使用92,984例正常壁厚队列的胸骨旁短轴超声心动图,并在肥厚壁厚数据集(26,523例超声心动图)和体外数据集(2,238例超声心动图)中无额外调整地评估性能;研究纳入197例患者,其中155例壁厚正常、42例壁厚增加;主要结局为左室分割性能和16节段左室壁厚度测量准确性。
诊断标准与关键数值
- 正常壁厚队列:左室心肌分割平均Dice相似系数0.87、平均Hausdorff距离4.56mm,16节段壁厚平均绝对误差0.31mm至0.88mm;肥厚壁厚队列:分割平均Dice相似系数0.87、平均Hausdorff距离4.81mm,16节段壁厚平均绝对误差均在临床可耐受阈值2mm以内;体外队列:节段壁厚平均绝对误差1.0mm。
临床应用与阅片要点
- 原文未据此给出推荐。
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
OBJECTIVES: The quantitative analysis of 16-segment left ventricular wall thickness can provide insights into the pathological progression of left ventricular hypertrophy, influencing diagnosis and treatment. However, its applications were time-consuming and had large variability. This study developed an automatic measurement of 16-segment left ventricular wall thickness based on left ventricular segmentation and validated in vivo and in vitro.
MATERIALS AND METHODS: The deep learning algorithm was trained and validated on 92,984 parasternal short-axis echocardiograms in a normal wall thickness cohort. Its performance was evaluated in an increased wall thickness dataset (26,523 echocardiograms) and an in vitro dataset (2238 echocardiograms) without additional adjustments. The primary outcomes were the performance of left ventricular segmentation and the accuracy in measuring 16-segment left ventricular wall thickness.
RESULTS: The study included 197 patients: 155 with normal wall thickness and 42 with increased wall thickness. Hypertension and nephropathy were common in the cohort with increased wall thickness. In the normal wall thickness cohort, the deep learning algorithm achieved similar accuracy to manual left ventricular myocardium segmentation (the mean dice similarity coefficient of 0.87, mean Hausdorff distance of 4.56mm) and accurately measured the 16-segment wall thickness (the mean absolute error from 0.31mm to 0.88mm). In the increased cohort, the mean dice similarity coefficient and Hausdorff distance of segmentation were 0.87 and 4.81mm, and the mean absolute errors of 16-segment wall thickness were within the clinically tolerable threshold of 2mm. In the in vitro cohort, the mean absolute error of segmental wall thickness was 1.0mm.
CONCLUSIONS: Deep learning-derived left ventricular myocardial contours and accurate segmentation were feasible, which may provide a method for assessing left ventricular myocardium. Additionally, the method produced accurate measures of 16-segment wall thickness that were verified in vivo and in vitro, which may provide fundamental and clinical applications for the accurate diagnosis of left ventricular hypertrophy.
引用
- Peng G, Xue W, Sheng Y, et al. Development and Validation of a Deep Learning System for Echocardiographic Assessment of 16-Segment LV Wall Thickness. Ultrasound Med Biol. 2026. PMID: 41832107. doi:10.1016/j.ultrasmedbio.2026.01.014
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。