解读偶发卵巢病变:利用纹理分析和机器学习进行特征描述与恶性检测
本研究回顾性纳入427例增强CT检出的偶发卵巢病变患者(良性348例、恶性79例),比较良恶性病变的CT纹理特征并构建结合纹理特征与年龄的三种机器学习模型。总像素、均值、标准差、熵、正像素均值(MPP)和偏度在良恶性组间有显著差异(p<0.05);以10为代价因子优化阈值时,随机森林模型敏感度92%、特异度60%,SVM模型敏感度91%、特异度69%,逻辑回归敏感度92%、特异度61%。作者认为CT纹理分析可提供客观影像分析,机器学习模型检测恶性病变具有高敏感度和中等特异度。
要点
- 偶发卵巢病变良恶性组间总像素、均值、标准差、熵、MPP和偏度差异显著(p<0.05)。
- 随机森林模型敏感度92%、特异度60%;SVM模型敏感度91%、特异度69%;逻辑回归敏感度92%、特异度61%。
- 模型结合CT纹理特征与患者年龄,以10为代价因子选择阈值。
核心结论
- CT纹理分析可对偶发卵巢病变提供客观影像分析,结合纹理特征与年龄的机器学习模型检测恶性病变具有高敏感度和中等特异度。
研究设计与样本
- 经IRB批准、符合HIPAA的回顾性研究,纳入427例增强CT检出的偶发卵巢病变连续患者,其中良性348例(81.5%)、恶性79例(18.5%)。
- 使用商用软件(TexRAD,Feedback Plc,Cambridge,UK)分析CT纹理特征,包括总像素、均值、标准差、熵、正像素均值(MPP)、偏度、峰度和熵。
- 结合纹理特征与患者年龄构建三种机器学习模型,采用十折交叉验证评估性能,构建ROC曲线评估敏感度和特异度,以代价加权法选择截断值。
关键影像表现/征象
- 总像素、均值、标准差、熵、MPP和偏度在良性与恶性组间有显著差异(p<0.05)。
诊断标准与关键数值
- 以10为代价因子优化截断值选择时,随机森林模型敏感度92%、特异度60%。
- 以10为代价因子优化截断值选择时,SVM模型敏感度91%、特异度69%。
- 以10为代价因子优化截断值选择时,逻辑回归模型敏感度92%、特异度61%。
临床应用与阅片要点
- 原文未据此给出推荐。
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
PURPOSE: To compare CT texture features of benign and malignant ovarian lesions and to build a machine learning model to detect malignancy in incidental ovarian lesions.
METHODS: In this IRB-approved, HIPAA-compliant, retrospective study, 427 consecutive patients with incidental ovarian lesions detected on contrast-enhanced CT (348, 81.5% benign and 79, 18.5% malignant) were included. The following CT texture features were analyzed using commercially available software (TexRAD, Feedback Plc, Cambridge, UK): total pixel, mean, standard deviation (SD), entropy, mean value of positive pixels (MPP), skewness, kurtosis and entropy. Three machine learning models were created by combining texture features and patients' age, and performance of these models was assessed using tenfold cross-validation. Receiver operating characteristics (ROC) were constructed to assess sensitivity and specificity. The cutoff value was picked using a cost-weighted method.
RESULTS: Total pixels, mean, SD, entropy, MPP, and skewness were significantly different between benign and malignant groups (p < 0.05). With a selected 10 as a cost factor to optimize cutoff value selection, sensitivity 92%, specificity 60% in the random forest (RF) model, sensitivity 91%, specificity 69% in SVM model, and sensitivity 92%, specificity 61% in the logistic regression, respectively.
CONCLUSION: CT texture analysis could provide objective imaging analysis of incidental ovarian lesions and ML models using CT texture features and age demonstrated high sensitivity and moderate specificity for detection of malignant lesions.
引用
- Park H, Qin L, Guerra P, et al. Decoding incidental ovarian lesions: use of texture analysis and machine learning for characterization and detection of malignancy. Abdom Radiol (NY). 2020. PMID: 32728871. doi:10.1007/s00261-020-02668-3
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。