Purpose

This study aimed to use conventional ultrasound features, ultrasound radiomics, and machine learning algorithms to establish a predictive model to assess the risk of post-surgical recurrence of gastrointestinal stromal tumors (GISTs).

Methods

This retrospective analysis included 230 patients with pathologically diagnosed GISTs. Radiomic features were extracted from manually annotated images. Radiomic features plus conventional ultrasound features were selected using the SelectKbest analysis of variance and stratified tenfold cross-validation recursive elimination methods. Finally, five different machine learning algorithms (logistic regression [LR], support vector machine [SVM], random forest [RF], extreme gradient boosting [XGBoost], and multilayer perceptron [MLP]) were established to predict risk stratification of GISTs. The predictive performance of the established model was mainly evaluated based on the area under the receiver operating characteristic (ROC) curve (AUC) and accuracy, whereas the predictive performance of the optimal machine learning algorithm and a radiologist's subjective assessment were compared using McNemar’s test.

Results

Seven radiomics features and one conventional ultrasound feature were selected to construct the machine learning models for GIST risk classification. The mentioned five machine learning models were able to predict the malignant potential of GISTs. LR and SVM outperformed other classifiers on the test set, with LR achieving an accuracy of 0.852 (AUC, 0.881; sensitivity, 0.871; specificity, 0.826) and SVM achieving an accuracy of 0.852 (AUC, 0.879; sensitivity, 0.839; specificity, 0.870), and proved significantly better than the radiologist (accuracy, 0.691; sensitivity, 0.645; specificity, 0.813).

Conclusion

Machine learning-based ultrasound radiomics features are able to noninvasively predict the biological risk of GISTs.

中文翻译：

---

使用基于机器学习的超声放射组学预测胃肠道间质瘤的风险分层

目的

本研究旨在利用传统超声特征、超声放射组学和机器学习算法建立预测模型来评估胃肠道间质瘤（GIST）术后复发的风险。

方法

这项回顾性分析包括 230 名经病理诊断为 GIST 的患者。从手动注释的图像中提取放射组学特征。使用 SelectKbest 方差分析和分层十倍交叉验证递归消除方法选择放射组学特征和常规超声特征。最后，建立了五种不同的机器学习算法（逻辑回归[LR]、支持向量机[SVM]、随机森林[RF]、极限梯度提升[XGBoost]和多层感知器[MLP]）来预测GIST的风险分层。所建立模型的预测性能主要根据受试者工作特征（ROC）曲线下面积（AUC）和准确性进行评估，而最佳机器学习算法的预测性能和放射科医生的主观评估则使用麦克尼马尔检验进行比较。

结果

选择七种放射组学特征和一种传统超声特征来构建 GIST 风险分类的机器学习模型。上述五种机器学习模型能够预测 GIST 的恶性潜力。LR 和 SVM 在测试集上优于其他分类器，LR 的准确度为 0.852（AUC，0.881；灵敏度，0.871；特异性，0.826），SVM 的准确度为 0.852（AUC，0.879；灵敏度，0.839；特异性，0.870） ），并被证明明显优于放射科医生（准确度，0.691；灵敏度，0.645；特异性，0.813）。

结论

基于机器学习的超声放射组学特征能够无创地预测 GIST 的生物学风险。