文献速递技术与安全CTMR

手动与深度学习分割提取的影像组学特征:一致性、偏倚与模型性能比较

Radiomic Features Extracted From Manual and Deep Learning-Based Segmentations: Comparison of Agreement, Bias, and Model Performance.
PubMed 文献 · AJR Am J Roentgenol · 2026-10-14

该研究比较了CT和MRI影像组学特征在手动分割与深度学习生成分割(每种模态评估三个模型)之间的一致性和偏倚,发现深度学习分割特征与手动分割特征一致性有限且存在系统偏倚,这些差异影响了下游模型的准确性和泛化能力,对计划大规模影像组学分析的研究者具有启示意义。

要点

  • 深度学习分割提取的影像组学特征与手动分割特征一致性有限。
  • 两种分割来源的特征间存在系统偏倚,影响模型准确性与泛化。
  • 大规模影像组学分析需关注分割方法带来的特征差异。
英文原文(摘要)· 点击展开

CT and MRI radiomic features extracted from deep learning–generated segmentations (three models assessed per modality) commonly exhibited limited agreement with, and presence of bias relative to, features derived from manual segmentations. These systematic differences impacted downstream model accuracy and generalizability. The findings have ramifications for investigators planning large-scale radiomic analyses.

来源:AJR Am J Roentgenol,查看原文。版权归原作者所有。

引用

  1. Dudley JA, Taylor Z, Li H, et al. Radiomic Features Extracted From Manual and Deep Learning-Based Segmentations: Comparison of Agreement, Bias, and Model Performance. AJR Am J Roentgenol. 2026. PMID: 42872852. doi:10.2214/AJR.26.35746

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。