人工智能检测和分类食管盘状异物
该研究训练目标检测器识别食管纽扣电池和硬币,并训练图像分类器区分两者。检测器对食管异物检测的敏感度和特异度均为100%;分类器准确分类6/6(100%)纽扣电池和93/95(97.9%)硬币。作者认为人工智能模型在检测和分类食管盘状异物方面有前景,可能用于分流X线片供放射科医生解读。
要点
- 目标检测器对食管异物检测的敏感度和特异度均为100%。
- 图像分类器准确分类6/6(100%)纽扣电池和93/95(97.9%)硬币。
- 2枚硬币被错误分类为纽扣电池,叠放异物均被分类为硬币。
核心结论
- 人工智能模型在检测和分类食管盘状异物方面显示出前景,可能用于分流X线片供放射科医生解读。
研究设计与样本
- 研究训练目标检测器检测纽扣电池和硬币,训练数据为57张X线片(19张食管纽扣电池、19张食管硬币、19张无异物,经内镜确认并按年龄和性别匹配);图像分类器训练数据为从检测器训练集中裁剪的19张纽扣电池和19张硬币X线片;在103张有食管异物和103张无异物X线片上测试。
诊断标准与关键数值
- 目标检测器对食管异物检测的敏感度为100%,特异度为100%。
- 图像分类器对纽扣电池分类准确率为6/6(100%)。
- 图像分类器对硬币分类准确率为93/95(97.9%)。
临床应用与阅片要点
- 作者认为人工智能模型可能用于分流X线片供放射科医生解读。
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
BACKGROUND: Early and accurate radiographic diagnosis is required for the management of children with radio-opaque esophageal foreign bodies. Button batteries are some of the most dangerous esophageal foreign bodies and coins are among the most common. We hypothesized that artificial intelligence could be used to triage radiographs with esophageal button batteries and coins.
OBJECTIVE: Our primary objective was to train an object detector to detect esophageal foreign bodies, whether button battery or coin. Our secondary objective was to train an image classifier to classify the detected foreign body as either a button battery or a coin.
MATERIALS AND METHODS: We trained an object detector to detect button batteries and coins. The training data set for the object detector was 57 radiographs, consisting of 3 groups of 19 images each with either an esophageal button battery, esophageal coin or no foreign body. The foreign bodies were endoscopically confirmed, and the groups were age and gender matched. We then trained an image classifier to classify the detected foreign body as either a button battery or a coin. The training data set for the image classifier consisted of 19 radiographs of button batteries and 19 of coins, cropped from the object detector training data set. The object detector and image classifier were then tested on 103 radiographs with an esophageal foreign body, and 103 radiographs without a foreign body.
RESULTS: The object detector was 100% sensitive and specific for detecting an esophageal foreign body. The image classifier accurately classified all 6/6 (100%) button batteries in the testing data set and 93/95 (97.9%) of the coins. The remaining two coins were incorrectly classified as button batteries. In addition to these images with a single button battery or coin, there were two unique cases in the testing data set: a stacked button battery and coin, and two stacked coins, both of which were classified as coins.
CONCLUSION: Artificial intelligence models show promise in detecting and classifying esophageal discoid foreign bodies and could potentially be used to triage radiographs for radiologist interpretation.
引用
- Rostad, Richer, Riedesel, et al. Esophageal discoid foreign body detection and classification using artificial intelligence. Pediatr Radiol. 2021. PMID: 34850259. doi:10.1007/s00247-021-05240-3
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。