中国现代神经疾病杂志 ›› 2021, Vol. 21 ›› Issue (3): 192-196. doi: 10.3969/j.issn.1672-6731.2021.03.011

• 神经外科疾病大数据 • 上一篇    下一篇

2 基于卷积神经网络的硬膜下和硬膜外血肿分割方法的一致性评价

田风选1, 常健博2, 陈亦豪2, 魏俊吉2, 冯铭2, 王任直2, 贺喜武1   

  1. 1. 810007 西宁, 青海省第五人民医院神经外科;
    2. 100730 中国医学科学院 北京协和医学院 北京协和医院神经外科
  • 收稿日期:2021-03-10 出版日期:2021-03-25 发布日期:2021-04-02
  • 通讯作者: 贺喜武,Email:1684793310@qq.com

Agreement evaluation of an automatic segmentation algorithm for quantifying subdural/epidural hemorrhage volume using convolution neural network

TIAN Feng-xuan1, CHANG Jian-bo2, CHEN Yi-hao2, WEI Jun-ji2, FENG Ming2, WANG Ren-zhi2, HE Xi-wu1   

  1. 1 Department of Neurosurgery, The Fifth People's Hospital of Qinghai Province, Xining 810007, Qinghai, China;
    2 Department of Neurosurgery, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100730, China
  • Received:2021-03-10 Online:2021-03-25 Published:2021-04-02

摘要:

目的 探讨基于卷积神经网络的血肿分割算法对硬膜下和硬膜外血肿的测量结果与手动分割结果的一致性。方法 纳入2017年1月至2019年6月中国颅内出血影像数据库129例硬膜下和硬膜外血肿患者计352张CT影像(硬膜下血肿33例计104张影像、硬膜外血肿96例计248张影像),均采用手动分割、算法分割、多田公式3种方法对血肿体积进行测量,以手动分割作为“金标准”,分别与算法分割和多田公式进行一致性检验,并探讨血肿形态和边界对算法的影响。结果 与多田公式相比,算法分割的百分误差最小(23.62%),有94.89%(334/352)的测值在95%一致性界限(95% LoA)内,与“金标准”的一致性良好;但算法分割的波动范围更大,在不对称(P=0.000)和边界清晰(P=0.000)的血肿中表现更佳。结论 基于卷积神经网络构建的算法分割具有一定的临床应用前景,但尚待进一步验证。

关键词: 血肿, 硬膜下, 血肿, 硬膜外, 颅内, 人工智能, 神经网络(计算机), 体层摄影术, X线计算机

Abstract:

Objective To validate the agreement among the convolution neural network segmentation algorithm, Tada formula and manual segmentation for subdural/epidural hemorrhage volume. Methods A total of 129 cases with 352 subdural/epidural hemorrhage CT scans were extracted from Chinese Intracranial Hemorrhage Image Database (CICHID) from January 2017 to June 2019. All CT scans were measured by three methods including manual segmentation, algorithm segmentation and Tada formula. The manual segmentation was regarded as the "golden standard" and the agreement test among three methods was performed. We explored the influence factors in different measurement methods, such as the shape or boundary of hematoma. Results Compared with the Tada formula method, the percentage error of segmentation algorithm was small (23.62%), and the agreement between algorithm and the manual reference was strong, which 94.89% (334/352) of the data was within the 95% limits of agreement (95%LoA), however, the 95%LoA was broad. And the performance of segmentation algorithm showed better in asymmetry (P=0.000) and clear boundary hematoma (P=0.000). Conclusions The segmentation algorithm based on convolution neural network has a certain application prospect, but need to be validated in large sample research.

Key words: Hematoma, subdural, Hematoma, epidural, cranial, Artificial intelligence, Neural networks (computer), Tomography, X-ray computed