SHAP值在XGBoost MRI模型中鉴别腮腺恶性肿瘤与Warthin瘤的价值  

The value of SHAP value in differentiating parotid malignant tumor from Warthin tumor in XGBoost MRI model

作　　者：张潇涵 朱翰林 韩志江[3] 杨海 ZHANG Xiao-han;ZHU Han-lin;HAN Zhi-jiang;YANG Hai(Department of Radiology,Afiliated Xiaoshan Hospital,Hangzhou Normal University,Hangzhou 311200,China;Department of Radiology,Hangzhou Ninth People's Hospital,Hangzhou 311225,China;Department of Radiology,Hangzhou First People's Hospital,Hangzhou 310006,China;Department of Radiology,Traditional Chinese Medicine Hospital of Wuyi County,Jinhua Zhejiang 321200,China)

机构地区：[1]杭州师范大学附属萧山医院放射科,浙江杭州311200 [2]杭州市第九人民医院放射科,浙江杭州311225 [3]杭州市第一人民医院放射科,浙江杭州310006 [4]浙江省武义县中医院放射科,浙江金华321200

出　　处：《中国临床医学影像杂志》2024年第7期462-466,共5页Journal of China Clinic Medical Imaging

摘　　要：目的:探讨SHAP值在极端梯度提升(Extreme gradient boosting,XGBoost)MRI模型中鉴别腮腺恶性肿瘤(Malignant tumor,MT)与Warthin瘤(Warthin tumor,WT)的价值。方法:回顾分析经手术病理证实的22例22枚MT和38例51枚WT的MRI资料,分析瘤体形态、位置和强化方式,以及T_(1)WI、T_(2)WI、FS-T_(1)WI、FS-T_(2)WI影像征象,经单因素分析筛选有统计学意义的征象,纳入XGBoost模型,使用受试者工作特征曲线下面积(Area under the curve,AUC)、敏感度、特异度评价模型诊断效能。通过可解释机器学习模型(Shapley additive explanations,SHAP)值对模型进行分析,明确各MRI征象在模型中的权重。结果:22枚MT和51枚WT中,持续性强化(P<0.05)、形态不规则(χ^(2)=20.707,P<0.05)、非尾叶(χ^(2)=8.911,P<0.05)、T_(2)WI高信号(χ^(2)=7.581,P<0.05)、FS-T_(1)WI等低信号(P<0.05)、FS-T_(2)WI高信号(χ^(2)=9.016,P<0.05)对鉴别两者有统计学意义,且均更常见于MT中。将6项单因素纳入XGBoost模型分析得出AUC为0.847,敏感度为77.3%,特异度为92.2%;6种MRI征象绝对平均SHAP值为0.21~0.99,其中形态不规则权重最大。结论:形态不规则、T_(2)WI高信号、持续性强化、FS-T_(2)WI高信号、非尾叶、FS-T_(1)WI等低信号对MT和WT鉴别的权重存在差异,对各MRI征象的准确识别,有利于对两者精准诊断。Objective:To explore the value of shapley additional explanations(SHAP)value in differentiating malignant tumor(MT)and Warthin tumor(WT)of parotid gland in extreme gradient boosting(XGBoost)MRI model.Methods:MRI data of22 MT in 22 cases and 51 WT in 38 cases confirmed by surgery and pathology were retrospectively analyzed.The tumors'shape,location and enhancement mode,as well as T_(1)WI,T_(2)WI,FS-T_(1)WI and FS-T_(2)WI imaging signs were analyzed.The statistically significant signs were screened by univariate analysis and included in XGBoost model,using receiver operating characteristic curve area under the curve(AUC),sensitivity and specificity to evaluate the diagnostic efficiency of the model.The model was analyzed by SHAP value,and the weight of MRI signs in the model was made clear.Results:Among 22 MT and51 WT,there were statistical differences in progressive reinforcement(P<0.05),irregular shape(χ^(2)=20.707,P<0.05),non-caudal lobe(χ^(2)=8.911,P<0.05),T_(2)WI high signal(χ^(2)=7.581,P<0.05),FS-T_(1)WI equal or low signal(P<0.05)and FS-T_(2)WI high signal(χ^(2)=9.016,P<0.05).The AUC was 0.847,the sensitivity was 77.3%,and the specificity was 92.2%when six single factors were included in XGBoost model.The absolute average SHAP value of six MRI signs was 0.21~0.99,among which the irregular shape had the largest weight.Conclusion:Irregular shape,T_(2)WI high signal,progressive reinforcement,FS-T_(2)WI high signal,non-caudal lobe and FS-T_(1)WI equal or low signal have different weights in the differential diagnosis of MT and WT.Accurate identification of MRI signs is conducive to accurate diagnosis of MT and WT.

关 键 词：腮腺肿瘤 腺淋巴瘤 磁共振成像 

分 类 号：R739.8[医药卫生—肿瘤] R445.2[医药卫生—临床医学]