摘要目的:建立甲状腺乳头状癌（PTC）CT影像特征3种机器学习预测模型，运用SHAP值分析最佳模型中各CT特征在模型中的贡献度。方法:回顾性分析2016年1月至2021年1月西湖大学医学院附属杭州市第一人民医院门诊和住院收治，且经病理证实426例440枚PTC的CT影像特征，与467例528枚结节性甲状腺肿（NG）对比，评估咬饼征、增强后范围缩小/模糊、微钙化和形态不规则4个CT特征在2者中的分布。将PTC和NG的CT影像资料以8∶2的比例随机分为训练集和测试集，采用极端梯度提升（XGBoost）、随机森林（RF）和支持向量机（SVM）构建3个机器学习模型。通过受试者工作特征曲线下面积（AUC）、准确率、F1评分等，筛选出最佳模型。使用SHAP值解释最佳模型中各CT特征对模型贡献度。结果:440枚PTC和528枚NG的CT特征中，咬饼征、增强后范围缩小/模糊、微钙化和形态不规则分别为326枚和30枚（ χ2=483.05, P<0.001）、363枚和106枚（ χ2=374.45, P<0.001）、158枚和53枚（ χ2=94.24, P<0.001）、354枚和52枚（ χ2=491.34, P<0.001）。XGBoost、RF和SVM构建的机器学习模型在训练集上的AUC、准确度、F1评分范围分别为0.884~0.925、0.867~0.873、0.844~0.854，在测试集上为0.869~0.923、0.845~0.871、0.803~0.845，其中XGBoost模型在测试集上诊断效能最高。4个CT特征中，形态不规则的绝对SHAP值最高，对诊断PTC为正向贡献。 结论:XGBoost机器学习模型诊断PTC的效能最高；CT特征中，形态不规则对诊断PTC贡献度最高且为正向贡献。

abstractsObjective:To establish three machine learning prediction models based on CT imaging characteristics of papillary thyroid carcinoma (PTC) , and use SHAP (shapley additive explanations) analysis to investigate the contribution of each CT image features in the best model.Methods:CT imaging features in 426 cases of 440 PTCs confirmed pathologically from Jan. 2016 to Jan. 2021 at the affiliated Hangzhou First People’s Hospital of Westlake University Medical School were retrospectively analyzed. compared with 467 cases of 528 nodular goiter (NG) , evaluating the distribution of four CT characteristics: cookie bite sign, enhanced range of narrowing/blur (ERNB) , microcalcifications, and irregular shape. We split the data into 8∶2 ratio for training and testing sets, then constructed three machine learning models using XGBoost, RF, and SVM. Based on AUC, accuracy, F1 score, and other metrics, we selected the best model. Lastly, we used SHAP values to assess each CT feature’s contribution and positive/negative effects on the model.Results:Among 440 PTC and 528 NG nodules, CT features like cookie bite sign, ERNB, microcalcifications, and irregular shape occurred in 326 and 30 ( χ 2=483.05, P<0.001) , 363 and 106 ( χ 2=374.45, P<0.001) , 158 and 53 ( χ 2=94.24, P<0.001) , and 354 and 52 ( χ 2=491.34, P<0.001) nodules, respectively. The machine learning models built using XGBoost, RF, and SVM had AUC, accuracy, and F1 scores ranging from 0.884~0.925, 0.867~0.873, and 0.844~0.854 respectively on the training set. On the test set, the scores ranged from 0.869~0.923, 0.845~0.871, and 0.803~0.845. Among them, the XGBoost model demonstrated the highest diagnostic performance on the test set. Among the four CT features, irregular shape had the highest absolute SHAP value, positively contributing to PTC diagnosis. Conclusion:XGBoost model showed the highest PTC diagnostic performance. Irregular shape had the greatest positive impact on PTC diagnosis.