Purpose

The objective of this study was to predict hematoma expansion (HE) by radiomic models based on different machine learning methods and determine the best radiomic model through the comparison.

Method

A total of 108 patients with intracerebral hemorrhage were retrospectively evaluated. Images of baseline non-contrast computed tomography (NCCT) and follow-up NCCT scan within 24 h were retrospectively reviewed. An HE was defined as a volume increase of more than 33% or an increase greater than 12.5 mL from the volume of the baseline NCCT. Texture parameters of the baseline NCCT images were selected by the least absolute shrinkage and selection operator (LASSO) regression. We used support vector machine (SVM), decision tree (DT), conditional inference trees (CIT), random forest (RF), k‑nearest neighbors (KNN), back-propagation neural network (BPNet) and Bayes to build models. Receiver operating characteristic (ROC) analysis and decision curve analysis (DCA) was performed and compared among models.

Results

Every model had a relatively high AUC (all > 0.75), SVM and KNN had the highest AUC of 0.91. There were significant differences between SVM and CIT (Z > 2.266, p = 0.02345), KNN and CIT (Z = 2.4834, p = 0.01301), RF and CIT (Z = 2.6956, p = 0.007027), KNN and BPNet (Z = 2.0122, p = 0.0442), RF and BPNet (Z = 1.9793, p = 0.04778). There was no significant difference among SVM, DT, RF, KNN and Bayes (p > 0.05). The SVM obtained the largest net benefit when the threshold probability was less than 0.33, while KNN obtained the largest net benefit when the threshold probability was greater than 0.33. Combined with ROC and DCA, SVM and KNN performed better in all the models for predicting HE.

Conclusion

Radiomic models based on different machine learning methods can be used to predict HE and the models generated by SVM and KNN performed best.

中文翻译：

---

基于不同机器学习方法预测脑出血扩大的放射组学模型比较

目的

本研究的目的是通过基于不同机器学习方法的放射组学模型预测血肿扩大（HE），并通过比较确定最佳放射组学模型。

方法

回顾性评价了108例脑出血患者。回顾性回顾了基线非对比计算机断层扫描 (NCCT) 的图像和 24 小时内的后续 NCCT 扫描。HE 被定义为体积比基线 NCCT 体积增加超过 33% 或增加超过 12.5 mL。通过最小绝对收缩和选择算子 (LASSO) 回归选择基线 NCCT 图像的纹理参数。我们使用支持向量机 (SVM)、决策树 (DT)、条件推理树 (CIT)、随机森林 (RF)、k 近邻 (KNN)、反向传播神经网络 (BPNet) 和贝叶斯来构建模型。接受者操作特征 (ROC) 分析和决策曲线分析 (DCA) 进行并在模型之间进行比较。

结果

每个模型的 AUC 都相对较高（均 > 0.75），SVM 和 KNN 的 AUC 最高，为 0.91。SVM 和 CIT (Z > 2.266, p  = 0.02345)、KNN 和 CIT (Z = 2.4834, p  = 0.01301)、RF 和 CIT (Z = 2.6956, p  = 0.007027)、KNN 和 BPNet (Z = 2.0122，p  = 0.0442），RF 和 BPNet（Z = 1.9793，p  = 0.04778）。SVM、DT、RF、KNN和Bayes之间没有显着差异（p > 0.05）。SVM 在阈值概率小于 0.33 时获得最大的净收益，而 KNN 在阈值概率大于 0.33 时获得最大的净收益。结合 ROC 和 DCA，SVM 和 KNN 在预测 HE 的所有模型中表现更好。

结论

基于不同机器学习方法的放射组学模型可用于预测 HE，SVM 和 KNN 生成的模型表现最佳。