文献速递胸部CT

指数化方法结合卷积神经网络按恶性分级分类CT肺结节的效能

Efficacy of exponentiation method with a convolutional neural network for classifying lung nodules on CT images by malignancy level.
PubMed 文献 · Eur Radiol · 2023-07-21

本研究基于LIDC-IDRI数据库的CT图像,构建四个卷积神经网络模型对肺结节按恶性等级分类(1 vs 2、1 vs 3、1 vs 4、1 vs 5),并对像素值施加1.0至10.0(步长0.5)的指数化处理。在恶性1 vs 4中,最大测试准确率和特异度分别提升至多0.012和0.037;在恶性1 vs 5中,最大测试准确率和敏感度分别提升至多0.030和0.0040。结论认为指数化方法提升了CNN分类效能,并可通过选择指数值调整敏感度和特异度。

要点

  • 指数化方法可提升卷积神经网络按恶性等级分类CT肺结节的准确率。
  • 在恶性1 vs 4中,特异度提升至多0.037;在恶性1 vs 5中,敏感度提升至多0.0040。
  • 选择不同指数值可调整模型的敏感度与特异度。

核心结论

  • 指数化方法提升了CNN在CT图像上分类肺结节良恶性的效能,其优势包括提高准确率以及通过选择适当指数值调整敏感度和特异度。

研究设计与样本

  • 研究使用LIDC-IDRI数据库的CT图像,构建四个CNN模型按恶性等级分类(1 vs 2、1 vs 3、1 vs 4、1 vs 5),指数化方法取指数值1.0至10.0、步长0.5;计算准确率、敏感度、特异度和AUC-ROC,并以Mann-Whitney U检验比较指数值1.0与其他各指数值的统计量。

诊断标准与关键数值

  • 在恶性1 vs 4中,最大测试准确率(指数值=2.0、3.0、3.5、4.5、5.0、5.5、6.0、6.5、7.0、7.5、8.0、8.5、9.0、9.5和10.0)和特异度(指数值=6.5、7.0和9.0)分别提升至多0.012和0.037。
  • 在恶性1 vs 5中,最大测试准确率(指数值=6.5和7.0)和敏感度(指数值=1.5)分别提升至多0.030和0.0040。

临床应用与阅片要点

  • 作者指出,通过选择指数值调整敏感度和特异度,可为肺结节患者的筛查、诊断和治疗流程构建合适的CNN模型。

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

OBJECTIVES: The aim of this study was to examine the performance of a convolutional neural network (CNN) combined with exponentiating each pixel value in classifying benign and malignant lung nodules on computed tomography (CT) images.

MATERIALS AND METHODS: Images in the Lung Image Database Consortium-Image Database Resource Initiative (LIDC-IDRI) were analyzed. Four CNN models were then constructed to classify the lung nodules by malignancy level (malignancy level 1 vs. 2, malignancy level 1 vs. 3, malignancy level 1 vs. 4, and malignancy level 1 vs. 5). The exponentiation method was applied for exponent values of 1.0 to 10.0 in increments of 0.5. Accuracy, sensitivity, specificity, and area under the curve of receiver operating characteristics (AUC-ROC) were calculated. These statistics were compared between an exponent value of 1.0 and all other exponent values in each model by the Mann-Whitney U-test.

RESULTS: In malignancy 1 vs. 4, maximum test accuracy (MTA; exponent value = 2.0, 3.0, 3.5, 4.5, 5.0, 5.5, 6.0, 6.5, 7.0, 7.5, 8.0, 8.5, 9.0, 9.5, and 10.0) and specificity (6.5, 7.0, and 9.0) were improved by up to 0.012 and 0.037, respectively. In malignancy 1 vs. 5, MTA (6.5 and 7.0) and sensitivity (1.5) were improved by up to 0.030 and 0.0040, respectively.

CONCLUSIONS: The exponentiation method improved the performance of the CNN in the task of classifying lung nodules on CT images as benign or malignant. The exponentiation method demonstrated two advantages: improved accuracy, and the ability to adjust sensitivity and specificity by selecting an appropriate exponent value.

CLINICAL RELEVANCE STATEMENT: Adjustment of sensitivity and specificity by selecting an exponent value enables the construction of proper CNN models for screening, diagnosis, and treatment processes among patients with lung nodules.

KEY POINTS: • The exponentiation method improved the performance of the convolutional neural network. • Contrast accentuation by the exponentiation method may derive features of lung nodules. • Sensitivity and specificity can be adjusted by selecting an exponent value.

来源:Eur Radiol,查看原文。版权归原作者所有。

引用

  1. Usuzaki T, Takahashi K, Takagi H, et al. Efficacy of exponentiation method with a convolutional neural network for classifying lung nodules on CT images by malignancy level. Eur Radiol. 2023. PMID: 37477673. doi:10.1007/s00330-023-09946-w

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。