摘要
目的 分析基于光谱CT各参数构建的机器学习模型预测甲状腺良恶性结节的可行性.方法 回顾性分析2021年9月～2022年12月经手术病理证实的185例甲状腺结节患者资料.根据病理结果将患者分为恶性结节组(n=106)及良性结节组(n=79),提取10个光谱CT参数构建6种机器学习模型,通过ROC曲线评价各模型预测甲状腺结节良恶性的效能,比较模型曲线下面积的差异.结果 预测甲状腺良恶性结节的极端梯度提升、随机森林、支持向量机、K最近邻学习模型、逻辑回归及决策树的AUC值分别为0.833、0.814、0.813、0.807、0.799、0.776,敏感度分别为0.833、0.833、0.800、0.733、0.767、0.733,特异度分别为0.808、0.769、0.731、0.846、0.808、0.731,准确度分别为0.821、0.804、0.768、0.786、0.786、0.732.结论 基于光谱CT各参数构建预测甲状腺良恶性结节的学习模型效能较好,最优预测模型为XGBoost.
Abstract
Objective To observe the feasibility of machine learning models constructed based on various parameters of spectral CT in predicting the benign and malignant nature of thyroid nodules. Methods A total of 185 patients with thyroid nodules confirmed by surgical pathology from September 2021 to December 2022 were analyzed retrospectively. According to the pathological results, the patients were divided into malignant nodules group (n=106) and benign nodules group (n=79). Ten spectral CT parameters were extracted to establish six machine learning models. The performance of each model in predicting the benign and malignant nature of thyroid nodules was evaluated through ROC curves, and the differences in AUC of the model were compared. Results The AUC values of extreme gradient boosting, random forest, support vector machine, K-nearest neighbors, Logistic regression and decision tree models for predicting thyroid nodule malignancy were 0.833, 0.814, 0.813, 0.807, 0.799, 0.776, respectively. Their sensitivities were 0.833, 0.833, 0.800, 0.733, 0.767, 0.733, their specificities were 0.808, 0.769, 0.731, 0.846, 0.808, 0.731, their accuracies were 0.821, 0.804, 0.768, 0.786, 0.786, 0.732. Conclusion The learning models based on the parameters from spectral CT to predict benign and malignant thyroid nodules had good overall performance, the optimal prediction model was XGBoost.
基金项目
深圳市科技计划资助项目(JCYJ20230807142308018)
国家科技期刊平台
NSTL
万方数据