文献速递儿科影像技术与安全

面向儿科放射科医生的机器学习概念、顾虑与机遇

Machine learning concepts, concerns and opportunities for a pediatric radiologist.
PubMed 文献 · Pediatr Radiol · 2019-03-29

本文介绍机器学习作为人工智能分支在小儿影像中的潜在价值,说明监督、无监督与半监督等学习类型,并阐释数据划分与欠拟合/过拟合两个核心概念。作者还讨论在儿童影像中实施机器学习的挑战,并回顾放射学中检测、分类与分割的机器学习案例。

要点

  • 小儿影像机器学习需极大样本量,并需少量儿科影像医师准确标注图像。
  • 监督、无监督与半监督是文中描述的三类学习方式。
  • 数据划分与欠拟合/过拟合是读者应掌握的两个核心概念。

核心结论

  • 机器学习在小儿影像质量与价值提升方面具有很大潜力,但实施面临数据量、标注、技术与监管及卷积神经网络不透明等挑战。

研究设计与样本

  • 本文为综述性文章,介绍机器学习的学习类型与核心概念,并回顾放射学中的检测、分类和分割案例;未报告样本量与中心数。

临床应用与阅片要点

  • 文中从美国儿科放射学会质量与安全委员会三名儿科放射科医生的角度,分享了对潜在发展领域的看法。

局限性

  • 实施机器学习需要非常大的数据集。
  • 需由数量相对较少的儿科影像医师准确标注这些图像。
  • 存在技术与监管障碍。
  • 卷积神经网络具有不透明性。

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

Machine learning, a subfield of artificial intelligence, is a rapidly evolving technology that offers great potential for expanding the quality and value of pediatric radiology. We describe specific types of learning, including supervised, unsupervised and semisupervised. Subsequently, we illustrate two core concepts for the reader: data partitioning and under/overfitting. We also provide an expanded discussion of the challenges of implementing machine learning in children's imaging. These include the requirement for very large data sets, the need to accurately label these images with a relatively small number of pediatric imagers, technical and regulatory hurdles, as well as the opaque character of convolution neural networks. We review machine learning cases in radiology including detection, classification and segmentation. Last, three pediatric radiologists from the Society for Pediatric Radiology Quality and Safety Committee share perspectives for potential areas of development.

来源:Pediatr Radiol,查看原文。版权归原作者所有。

引用

  1. Moore MM, Slonimsky E, Long AD, et al. Machine learning concepts, concerns and opportunities for a pediatric radiologist. Pediatr Radiol. 2019. PMID: 30923883. doi:10.1007/s00247-018-4277-7

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。