利用人工智能实现儿科超声心动图的自动化视图分类模型。
An Automated View Classification Model for Pediatric Echocardiography Using Artificial Intelligence.
机构信息
Department of Cardiology, Boston Children's Hospital, and Department of Pediatrics, Harvard Medical School, Boston, Massachusetts.
One Brave Idea, Division of Cardiovascular Medicine, Brigham and Women's Hospital and Harvard Medical School, Boston, Massachusetts.
出版信息
J Am Soc Echocardiogr. 2022 Dec;35(12):1238-1246. doi: 10.1016/j.echo.2022.08.009. Epub 2022 Aug 29.
BACKGROUND
View classification is a key step toward building a fully automated system for interpretation of echocardiograms. However, compared with adult echocardiograms, creating a view classification model for pediatric echocardiograms poses additional challenges, such as greater variation in anatomy, structure size, and views. The aim of this study was to develop a computer vision model to autonomously perform view classification on pediatric echocardiographic images.
METHODS
Using a training set of 12,067 echocardiographic images from patients aged 0 to 19 years, a convolutional neural network model was trained to identify 27 preselected standard pediatric echocardiographic views which included anatomic sweeps, color Doppler, and Doppler tracings. A validation set of 6,197 images was used for parameter tuning and model selection. A test set of 9,684 images from 100 different patients was then used to evaluate model accuracy. The model was also evaluated on a per study basis using a second test set consisting of 524 echocardiograms from children with leukemia to identify six preselected views pertinent to cardiac dysfunction surveillance.
RESULTS
The model identified the 27 preselected views with 90.3% accuracy. Accuracy was similar across age groups (89.3% for 0-4 years, 90.8% for 4-9 years, 90.0% for 9-14 years, and 91.2% for 14-19 years; P = .12). Examining the view subtypes, accuracy was 78.3% for the cine one location, 90.5% for sweeps with color Doppler, 82.2% for sweeps without color Doppler, and 91.1% for Doppler tracings. Among the leukemia cohort, the model identified the six preselected views on a per study basis with a positive predictive value of 98.7% to 99.2% and sensitivity of 76.9% to 94.8%.
CONCLUSIONS
A convolutional neural network model was constructed for view classification of pediatric echocardiograms that was accurate across the spectrum of ages and view types. This work lays the foundation for automated quantitative analysis and diagnostic support to promote efficient, accurate, and scalable analysis of pediatric echocardiograms.
背景
视图分类是构建超声心动图自动解读系统的关键步骤。然而，与成人超声心动图相比，为儿科超声心动图创建视图分类模型还存在其他挑战，例如解剖结构、结构大小和视图的变化更大。本研究旨在开发一种计算机视觉模型，以便自主对儿科超声心动图像进行视图分类。
方法
使用来自 0 至 19 岁患者的 12067 张超声心动图像的训练集，训练卷积神经网络模型以识别 27 个预先选择的标准儿科超声心动图视图，包括解剖扫查、彩色多普勒和多普勒描记。使用 6197 张图像的验证集进行参数调整和模型选择。然后使用来自 100 位不同患者的 9684 张图像的测试集评估模型准确性。还基于每个研究使用第二个测试集评估模型，该测试集由 524 张来自白血病儿童的超声心动图组成，用于识别与心脏功能障碍监测相关的六个预先选择的视图。
结果
该模型以 90.3%的准确率识别了 27 个预先选择的视图。在不同年龄段之间的准确率相似（0-4 岁为 89.3%，4-9 岁为 90.8%，9-14 岁为 90.0%，14-19 岁为 91.2%；P=0.12）。检查视图亚型，电影单部位的准确率为 78.3%，彩色多普勒扫查的准确率为 90.5%，无彩色多普勒扫查的准确率为 82.2%，多普勒描记的准确率为 91.1%。在白血病队列中，该模型基于每个研究识别了六个预先选择的视图，阳性预测值为 98.7%至 99.2%，灵敏度为 76.9%至 94.8%。
结论
为儿科超声心动图构建了视图分类的卷积神经网络模型，该模型在各个年龄段和视图类型中均具有准确性。这项工作为自动定量分析和诊断支持奠定了基础，以促进儿科超声心动图的高效、准确和可扩展分析。
Zotero 插件