文献速递骨肌儿科影像DR

触手可及的人工智能研究:一种用于在儿童腕部X线片上识别佝偻病的目标检测模型

Artificial intelligence research within reach: an object detection model to identify rickets on pediatric wrist radiographs.
PubMed 文献 · Pediatr Radiol · 2021-01-05

该研究利用2007至2018年7岁以下儿童腕部X线片,由两名儿科放射科医师将图像分为佝偻病或正常,并围绕桡骨远端和尺骨远端干骺端绘制带标签的边界框,构建目标检测模型。训练集包括104例佝偻病腕和264例正常腕,测试集包括20例佝偻病腕和37例正常腕,模型对佝偻病腕的敏感度和特异度分别为80%和95%,对正常腕分别为89%和90%,总体准确率为86%。结论认为目标检测可在儿童腕部X线片上识别佝偻病,且小数据集、简单软件工具和适度算力即可开发此类模型。

要点

  • 模型在测试集识别佝偻病腕的敏感度为80%、特异度为95%。
  • 模型在测试集识别正常腕的敏感度为89%、特异度为90%,总体准确率86%。
  • 目标检测模型可用小数据集、简单软件工具和适度算力开发。

核心结论

  • 目标检测可在儿童腕部X线片上识别佝偻病,且小数据集、简单软件工具和适度算力即可开发目标检测模型。

研究设计与样本

  • 该研究经Children's Healthcare of Atlanta机构审查委员会批准,检索放射信息系统2007至2018年7岁以下儿童因评估佝偻病而行的腕部X线检查。
  • 纳入标准为检查类型为「Rickets Survey」或「Joint Survey 1 View」且报告含「rickets」或「rachitic」字样,排除报告含「renal」「kidney」或「transplant」字样者。
  • 由两名儿科放射科医师阅片并将图像分类为佝偻病或正常,围绕桡骨远端和尺骨远端干骺端绘制带标签的边界框进行标注。
  • 训练集来自2007年1月1日至2017年12月31日采集的图像,包括104例佝偻病腕和264例正常腕,用于创建目标检测模型。
  • 测试集由2018日历年采集的图像组成,包括20例佝偻病腕和37例正常腕,测量模型的敏感度、特异度和准确率。

诊断标准与关键数值

  • 测试集中20例佝偻病腕里,16例被正确识别为佝偻病、2例被错误识别为正常、2例无预测,对应佝偻病腕的敏感度80%和特异度95%。
  • 测试集中37例正常腕里,33例被正确识别为正常、2例被错误识别为佝偻病、2例无预测,对应正常腕的敏感度89%和特异度90%。
  • 模型总体准确率为86%。

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: Artificial intelligence models have been successful in analyzing ordinary photographic images. One type of artificial intelligence model is object detection, where a labeled bounding box is drawn around an area of interest. Object detection can be applied to medical imaging tasks.

OBJECTIVE: To demonstrate object detection in identifying rickets and normal wrists on pediatric wrist radiographs using a small dataset, simple software and modest computer hardware.

MATERIALS AND METHODS: The institutional review board at Children's Healthcare of Atlanta approved this study. The radiology information system was searched for radiographic examinations of the wrist for the evaluation of rickets from 2007 to 2018 in children younger than 7 years of age. Inclusion criteria were an exam type of "Rickets Survey" or "Joint Survey 1 View" with reports containing the words "rickets" or "rachitic." Exclusion criteria were reports containing the words "renal," "kidney" or "transplant." Two pediatric radiologists reviewed the images and categorized them as either rickets or normal. Images were annotated by drawing a labeled bounding box around the distal radial and ulnar metaphases. The training dataset was created from images acquired from Jan. 1, 2007, to Dec. 31, 2017. This included 104 wrists with rickets and 264 normal wrists. This training dataset was used to create the object detection model. The testing dataset consisted of images acquired during the 2018 calendar year. This included 20 wrists with rickets and 37 normal wrists. Model sensitivity, specificity and accuracy were measured.

RESULTS: Of the 20 wrists with rickets in the testing set, 16 were correctly identified as rickets, 2 were incorrectly identified as normal and 2 had no prediction. Of the 37 normal wrists, 33 were correctly identified as normal, 2 were incorrectly identified as rickets and 2 had no prediction. This yielded a sensitivity and specificity of 80% and 95% for wrists with rickets and 89% and 90% for normal wrists. Overall model accuracy was 86%.

CONCLUSION: Object detection can identify rickets on pediatric wrist radiographs. Object detection models can be developed with a small dataset, simple software tools and modest computing power.

来源:Pediatr Radiol,查看原文。版权归原作者所有。

引用

  1. Meda KC, Milla SS, Rostad BS, et al. Artificial intelligence research within reach: an object detection model to identify rickets on pediatric wrist radiographs. Pediatr Radiol. 2021. PMID: 33399980. doi:10.1007/s00247-020-04895-8

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。