基于深度学习特征与影像组学特征融合的脑膜瘤质地评估
回顾性纳入2016年12月至2018年12月本院202例术后病理诊断为脑膜瘤的患者,基于T2-FLAIR图像由神经外科医生按Zada质地分级系统分为软或硬。融合深度学习特征与影像组学特征构建分类模型,逻辑回归分类器在测试队列中AUC为0.943(95%CI 0.873-1.000),基于签名构建的列线图C-index在训练队列为0.822(95%CI 0.758-0.885)、测试队列为0.943(95%CI 0.873-1.000)。
要点
- 基于T2-FLAIR图像融合深度学习与影像组学特征可辅助术前评估脑膜瘤质地。
- 逻辑回归分类器使用17个影像组学特征和9个深度学习特征,测试队列AUC为0.943(95%CI 0.873-1.000)。
- 列线图在训练队列C-index为0.822(95%CI 0.758-0.885),测试队列为0.943(95%CI 0.873-1.000)。
核心结论
- 融合深度学习特征与影像组学特征的方法评估脑膜瘤质地具有潜在临床价值,可辅助医师术前判断肿瘤质地。
研究设计与样本
- 回顾性纳入本院2016年12月至2018年12月202例经手术和病理诊断的脑膜瘤患者,随机分为训练队列162例和测试队列40例;基于T2-FLAIR图像由专业神经外科医生按Zada质地分级系统分为软或硬;采用卷积神经网络提取深度学习特征并与影像组学特征融合,经多次特征选择后用四种分类器构建分类模型,以AUC评价性能,并用LASSO构建签名及列线图。
诊断标准与关键数值
- 以Zada质地分级系统将脑膜瘤分为软或硬作为参考标准;逻辑回归分类器使用17个影像组学特征和9个深度学习特征,在测试队列中精确率0.855、召回率0.854、F1分数0.852、AUC 0.943(95%CI 0.873-1.000);列线图C-index在训练队列为0.822(95%CI 0.758-0.885),在测试队列为0.943(95%CI 0.873-1.000),校准良好。
临床应用与阅片要点
- 原文结论认为该方法具有潜在临床价值,可辅助医师术前判断肿瘤质地;决策曲线分析证实列线图预测脑膜瘤质地具有临床 usefulness。
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: This study aims to combine deep learning features with radiomics features for the computer-assisted preoperative assessment of meningioma consistency.
METHODS: 202 patients with surgery and pathological diagnosis of meningiomas at our institution between December 2016 and December 2018 were retrospectively included in the study. The T2-fluid attenuated inversion recovery (T2-Flair) images were evaluated to classify meningioma as soft or hard by professional neurosurgeons based on Zada's consistency grading system. All the patients were split randomly into a training cohort (n = 162) and a testing cohort (n = 40). A convolutional neural network (CNN) model was proposed to extract deep learning features. These deep learning features were combined with radiomics features. After multiple feature selections, selected features were used to construct classification models using four classifiers. AUC was used to evaluate the performance of each classifier. A signature was further constructed by using the least absolute shrinkage and selection operator (LASSO). A nomogram based on the signature was created for predicting meningioma consistency.
RESULTS: The logistic regression classifier constructed using 17 radiomics features and 9 deep learning features provided the best performance with a precision of 0.855, a recall of 0.854, an F1-score of 0.852 and an AUC of 0.943 (95 % CI, 0.873-1.000) in the testing cohort. The C-index of the nomogram was 0.822 (95 % CI, 0.758-0.885) in the training cohort and 0.943 (95 % CI, 0.873-1.000) in the testing cohort with good calibration. Decision curve analysis further confirmed the clinical usefulness of the nomogram for predicting meningioma consistency.
CONCLUSIONS: The proposed method for assessing meningioma consistency based on the fusion of deep learning features and radiomics features is potentially clinically valuable. It can be used to assist physicians in the preoperative determination of tumor consistency.
引用
- Zhang J, Zhao Y, Lu Y, et al. Meningioma consistency assessment based on the fusion of deep learning features and radiomics features. Eur J Radiol. 2023. PMID: 38071910. doi:10.1016/j.ejrad.2023.111250
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。