机构地区:[1]杭州市萧山区第三人民医院超声科,浙江杭州311251 [2]杭州市第九人民医院放射科,浙江杭州311225 [3]杭州市第九人民医院超声科,浙江杭州311225 [4]西湖大学附属杭州市第一人民医院放射科,浙江杭州310006
出 处:《临床超声医学杂志》2025年第3期248-252,共5页Journal of Clinical Ultrasound in Medicine
摘 要:目的探讨SHAP值在XGBoost超声模型中诊断最大径>1 cm甲状腺乳头状癌(PTC)的临床价值。方法选取我院经手术病理证实的138例PTC患者(145枚结节)和127例结节性甲状腺肿(NG)患者(141枚),将286枚结节以8∶2的比例随机分为训练集(PTC 114枚,NG 115枚)和测试集(PTC 31枚,NG 26枚),比较训练集与测试集,以及训练集中PTC与NG超声检查结果的差异。基于训练集构建XGBoost超声模型;绘制受试者工作特征(ROC)曲线分析XGBoost超声模型对训练集和测试集中PTC的诊断效能,校准曲线分析模型的校准度。通过SHAP值解析XGBoost超声模型,明确各超声因素诊断PTC的权重。结果训练集与测试集结节实性占比比较差异有统计学意义(P<0.05),边缘模糊/不规则/腺外侵犯、纵横比(A/T)>1、超声灰阶比值(UGSR)<0.83、微钙化占比比较差异均无统计学意义。训练集中PTC与NG A/T>1、UGSR<0.83、微钙化、边缘模糊/不规则/腺外侵犯、实性占比比较差异均有统计学意义(均P<0.001)。ROC曲线分析显示,XGBoost超声模型诊断训练集和测试集中PTC的曲线下面积分别为0.941(95%可信区间:0.895~0.987)和0.921(95%可信区间:0.846~0.996),准确率、灵敏度、特异度分别为88.0%、92.9%、83.2%和86.0%、93.5%、77.5%。校准曲线分析显示,XGBoost超声模型诊断训练集和测试集中PTC的校准度均较好。通过SHAP值解析XGBoost超声模型,结果显示各超声因素对诊断PTC均为正向贡献,其权重从高到低依次为:UGSR<0.83、实性、边缘模糊/不规则/腺外侵犯、微钙化和A/T>1。结论利用SHAP值对XGBoost超声模型进行解析可以实现各超声因素诊断效能的量化,在诊断最大径>1 cm PTC中具有较好的临床应用价值。Objective To evaluate the clinical value of SHAP value in XGBoost-based ultrasound model for diagnosing papillary thyroid carcinoma(PTC)with a maximum diameter>1 cm.Method A total of 138 PTC patients(145 nodules)and 127 nodular goiter(NG)patients(141 nodules)confirmed by surgical pathology in out hospital were enrolled.The 286 nodules were randomly divided into a training set(114 PTC,115 NG)and a testing set(31 PTC,26 NG)at a ratio of 8∶2.Ultrasound features were compared between the training and testing sets,as well as between PTC and NG nodules in the training set.A XGBoost ultrasound model was constructed using the training set.Receiver operating characteristic(ROC)curve was drawn to assess the diagnostic performance,and calibration curve evaluated the calibration of the model.SHAP value was used to interpret the model and quantify the contribution of each ultrasound feature for PTC diagnosis.Results There was significant difference in solid composition proportion between the training and testing sets(P<0.05),while there were no significant differences in the proportion of ill-defined/irregular/extrathyroidal extension margins,aspect ratio(A/T)>1,ultrasound grayscale ratio(UGSR)<0.83,and microcalcifications.In the training set,PTC nodules exhibited significantly higher proportion of A/T>1,UGSR<0.83,microcalcifications,ill-defined/irregular/extrathyroidal extension margins,and solid composition compared to NG nodules(all P<0.001).ROC curve analysis showed the area under the curve(AUC)of XGBoost model in the training and testing sets were 0.941(95%CI:0.895~0.987)and 0.921(95%CI:0.846~0.996),respectively,with accuracies of 88.0%and 86.0%,sensitivities of 92.9%and 93.5%,and specificities of 87.2%and 77.5%,respectively.Calibration curve indicated good model fit for both sets.SHAP interpretation revealed that all ultrasound features positively contributed to PTC diagnosis,ranked by weight as follows:UGSR<0.83,solid composition,ill-defined/irregular/extrathyroidal extension margins,microcalcifications,and A/T>1.Con
关 键 词:超声检查 甲状腺乳头状癌 超声灰阶比值 SHAP值 XGBoost
分 类 号:R445.1[医药卫生—影像医学与核医学] R736.1[医药卫生—诊断学]
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...